news
19459002> LookAtMyQuant 2: Qwen3.5-4B in 1.2 GB 19459404 LookAtMyQuant 2 19459402 takes Qwen3.5-4B to 1.23 GB, 2.3 bits per weight. The file resolves 330 of 540 exact-answer mathematics issues. Prism ML’s Bonsai 2 Ternary 4B resolves 328 at 1.20 GB, and the initial resolves 359. Its KL from the initial
on held-out text is 0.45 nats, versus 0.59 for Bonsai. 19459400 The dish is post-training quantization with trellis codes, adjusted on the design’s own worked mathematics services with digit tokens weighted 10 ×, and with bits designated per row by Fisher info. 4,000 actions of distillation follow, in which every layer is re-quantized as soon as from the initial weights. The page has the findings, the complete dish and its expense, about $34 of GPU time. The dish’s code is here, and the next area runs it. The 1.23 GB file itself is the 19459155 release, loaded for the Apple Silicon runtime.
Qwen3.8-27B in 7.2 GB The exact same dish takes Qwen3.8-27B from 53.7 GB to 7.2 GB, 2.1 bits per weight, adjusted and distilled on English prose and chat, code and mathematics. The file fixes 401 of the 540 mathematics issues, as numerous as the initial, and Unsloth’s 19459156 GGUF of the very same size resolves 348. It likewise ratings above that GGUF on HumanEval( 73.8% versus 67.1%), MMLU (78.8% versus 75.0%) and GSM8K( 94.5% versus 92.7%).
19459400 Next to it sits a 4-bit DSpine drafter 19459402 (1.02 GB, transformed from with and loaded with 19459159, which approximately doubles decode speed on mathematics and code on Apple silicon. 19459400 19459400 Inkling-2-Mini in 5.4 GB and 8.9 GB 19459404 The exact same compiler and proximal fine-tuning ( 19459160, encompassed a 26.4 B physical/ 5.4 B active 3 ×- looped MoE ( looped 3 ×=48 rational sees, 128 professionals top-8), compresses from 18.89 GB of text weights to 19459411 5.38 GB ( bits/weight) and 19459411 8.87 GB ( 19459164 bits/weight) utilizing 4 KB SRAM hashed block-vector trellis codes (/ , row trellis codes ( 19459167/ , and a 19459169 flooring on the tail pail ( 19459171. 19459400 Examined on the exact same BF16 instructor, 19459172 (, greedy, reasoning_effort=19659004, all 19459175 executable issues, Vision VQA tests (with the unblemished 637 MB BF16 vision tower), and 16 +13- language TextMix windows: Design Text GB 19659007 ( 19459179 19659008 GSM8K (200) 19659009 HumanEval (164) Pooled vs Teacher 19659011 Initial BF16 Teacher 18.89 19659013 0.0000 19659014 0.0000 147/ 200 (73.5%) 66/ 164 (40.2%) 100.0% 19659018 LookAtMyQuant 5.4 GB ( 5.38 0.0895 19659021 0.2871 134/ 200 (67.0%) 60/ 164 (36.6%) 19659024 91.2% Unsloth 19659026 7.76 0.0808 0.1236 19659029 127/ 200 (63.5%) 46/ 164 (28.0%) 19659031 81.6% 19659032 LookAtMyQuant 8.9 GB ( 8.87 0.0208 19659035 0.0571 19659036 148/ 200 (74.0%) 58/ 164 (35.4%) 96.8% 19659039 Unsloth 19659040 9.14 19659041 0.0486 0.0811 19659043 132/ 200 (66.0%) 49/ 164 (29.9%) 85.3% Unsloth 11.66 19659048 0.0189 19659049 0.0278 19659050 148/ 200 (74.0%) 68/ 164 (41.5%) 101.4% At 19459411 5.38 GB 19459413, 19459186 outscores Unsloth’s 9.14 GB 19459184 on both GSM8K and HumanEval at 41% smaller sized size (up from 76/200 and 24/164 untuned). At 19459411 8.87 GB 19459413, matches the BF16 initial on GSM8K and cuts KL on English/chat/code and 19459190 throughout all 13 languages listed below Both run straight through 19459192 and 19459193 with custom-made Apple Silicon Metal kernels, additionally coupled with the 178.9 MB 4-bit DSpark drafter : on M5 Max, 5.4 GB 19459413 deciphers at 19659054 serial and 19659055 mean ( 19659056 mathematics, thinking, 19659058 code, 19659059 prose; 19659060 prefill at , 19459202 active RAM), and 8.9 GB 19459413 deciphers at 19659061 serial and 19659062 mean ( 19659063 thinking, 19659064 mathematics; prefill at , 19459209 active RAM). 19459400
Install and utilize the Apple Silicon runtime 19459404 On Apple silicon, set up the runtime plan straight( this sets up , 19459211, 19459212, and 19459213 without the PyTorch compiler reliances):
19659066 From a regional clone, 19459214 sets up the 19459192 command, or run 19459216 without setting up. 19459400 19459411 Command line. Pass either a regional design folder or a Hugging Face repo ID: lamq generate shinmegamiboson/Qwen3.8-27B-LookAtMyQuant-7.2GB -pWhat is 17 times 23? Show the steps.19459454lamq generate shinmegamiboson/Qwen3.8-27B-LookAtMyQuant-7.2GB –drafter shinmegamiboson/Qwen3.8-27B-LookAtMyQuant-7.2GB-drafter -pWrite a Python function that merges overlapping intervals.–stats lamq chat shinmegamiboson/Qwen3.8-27B-LookAtMyQuant-7.2GB –drafter shinmegamiboson/Qwen3.8-27B-LookAtMyQuant-7.2GB-drafter lamq generate shinmegamiboson/Inkling-2-Mini-LookAtMyQuant-5.4GB –drafter shinmegamiboson/Inkling-2-Mini-LookAtMyQuant-drafter -p19459454What is 17 times 23? Show the steps.–stats”> lamq create shinmegamiboson/Qwen3.8 -27 B-LookAtMyQuant-7.2 GB – p 19659067 lamq create shinmegamiboson/Qwen3.8 -27 B-LookAtMyQuant-7.2 GB — drafter shinmegamiboson/Qwen3.8 -27 B-LookAtMyQuant-7.2 GB-drafter – p — statistics lamq chat shinmegamiboson/Qwen3.8 -27 B-LookAtMyQuant-7.2 GB — drafter shinmegamiboson/Qwen3.8 -27 B-LookAtMyQuant-7.2 GB-drafter lamq produce shinmegamiboson/Inkling -2- Mini-LookAtMyQuant-5.4 GB — drafter shinmegamiboson/Inkling -2- Mini-LookAtMyQuant-drafter – p 19659069– statistics checks out stdin when 19459218 is left out, supports , , 19459221, 19459222, 19459223, from mlx_lm import generate from runtime_mlx import load model, tok=load(19459454shinmegamiboson/Qwen3.8-27B-LookAtMyQuant-7.2GB) prompt=tok.apply_chat_template( L_SQUARE_B.L_SQUARE_B. 19459454role19459454:user,19459454content19459454:What is 17 times 23? Show the steps.19459454 R_SQUARE_B.R_SQUARE_B., add_generation_prompt=True, enable_thinking=False, ) print(generate(model, tok, prompt, max_tokens=200))”>, and 19459225 checks a jam-packed design’s format and measurements without packing weights. Python combination. returns an 19459228 set suitable with and 19459230 (auto-detecting designs such as Qwen3.5-4B/ Qwen3.8-27B and exports such as Gemma 4 E2B): 19459400
from 19659073 mlx_lm 19659074 import 19659075 produce 19659076 from 19659077 runtime_mlx 19659078 import load design 19459434, tok 19659081=load 19459434( 19659083 timely 19659084=Qwen3.8-27B-LookAtMyQuant-7.2GB, drafter=shinmegamiboson tok 19459433 apply_chat_template 19459434( , 19459433 add_generation_prompt=Real , enable_thinking 19659092=19659093 Incorrect ,) 19459433 print 19459434 ( 19459433 create ( 19459433 design 19459434, 19459433 tok 19459434, 19459433 timely , max_tokens 19659094=19659095 200 ) For streaming, multi-turn chat, or speculative translating through one user interface, usage :
from runtime_mlx SmolLM2-360M’)19459454python -c19459454from datasets import load_dataset as l; l(‘Salesforce import Runtime 19659102 rt=Runtime 19459434 load 19459434( , drafter=19659107, ) for 19659108 portion in 19659110 rt 19459434 19459433 stream 19459434( 19659111, 19459433 max_tokens 19659112=256 : print ( piece 19459434 text , end 19659114=19659115, flush 19659116=19659117 Real ) Run the compiler on SmolLM2-360M in 3 commands Set up the reliances and download the design and WikiText-2. The code checks out SmolLM2-360M’s config when it begins, so do
this.
Step. 19459413 This compresses one piece at a time, leaves the rest exact, and records the KL to the initial design. The complete set for this design, 2,312 measurements, took about 50 minutes on an L40S. 19459400 19659121 Assign. 19459413 This selects one code per piece, with the least summed KL inside the budget plan. It requires no GPU. This command utilizes the SmolLM2-360M measurements in this repository, with a spending plan of 3.044 bits per weight.
19659122 The strategy it composes has the very same widths as 19459400 19459411 Develop and score. 19459413 This quantizes the design to the strategy, one matrix at a time. It then reports perplexity and KL to the initial on WikiText-2 test. 19459400 19659123 Our run of that strategy remains in 19459235 It reached perplexity 16.64 and KL 0.0926, versus an initial perplexity of 15.31. Include for a fast check that whatever works. 19459400 19459400
Run the LookAtMyQuant 2 dish on Qwen3.5-4B Each action is a called task; prints its command. The tasks require Qwen3.5-4B in 19459238 on the volume and the text mix( 19459239. Every file they make goes to on the volume.
Task Script What it does Our run 19659128 The initial responses science, code, direction and tool triggers, and MATH and GSM8K training issues. This likewise downloads MATH 19659129 The calibration set: the original’s greedy services to 384 MATH training issues, in 2,048-token windows 15 minutes, H100 19659131 The mathematics text the fine-tuning distils on: services to MATH issues and to 19459247 issues from another seed, with incorrect estimations fixed and just ideal responses kept 33 minutes, H100 Each output row’s Fisher info at digit-predicting positions 1 minute, H100 Post-training quantization: trellis codes along rows, digit tokens weighted 10 × in the Hessians, QEP at α=0.5, bits per row from the row strategy, e-means codebooks 37 minutes, H100 4,000 actions of distillation on hidden weights pulled towards their codes, with every layer re-quantized when from the initial weights 2.7 hours, B200 19659139, The 540-problem standard and a 2nd set of 216, for the initial, Bonsai, the develop and the fine-tuned develop 20 minutes a design, H100 19659142 KL from the initial on WikiText-2 and on the text mix’s
13 sources 19659143 1 minute a design, H100 19659144 , and run the very same code on 2 layers, 60 actions and 36 issues, in a couple of minutes each. 19459400 19459411 The row strategy. 19659145 checks out 19459263, which provides 1 to 4 bits to each of 1,520 row groups. made it. Its 19459265 command constructs and loads a control with a set split of rows, 19459266 designates rows inside the control’s jam-packed bytes with 19459267, and builds under the strategy and checks the jam-packed size versus it. 19459400 19459411 The codebooks. The 4 e-means codebooks the construct translates with remain in offers the earlier ones, and a strategy made before the modification keeps them unless the flag states otherwise. 19459400 Current equal-byte dish enhancements. 19459413 2 zero-byte additions are readily available in the compiler and tweak scripts: 19459400 19659146 Softmax-Fisher token table ranking ( in 19459272 and. Ranks connected embedding/output containers by a 50/50 mix of output Softmax-Fisher 19459274 and corpus frequency (and untied output heads by Softmax-Fisher alone). Mid-training equal-byte row reallocation ( 19459275 in. 19459413 Paired with , steps each layer’s continuous-vs-stored proximal KL space 19459278 to move bits towards deep unabsorbed layers while keeping specific file size (. Integrated with on Qwen3.5-4B at similar 1.23 GB (, WikiText-2 KL fell from 0.425 to 0.387 (− 9.0 %) and 13-source KL from 0.447 to 0.378 (− 15.5 %, enhancing all 13 sources). 19459411 Where these tasks vary from our run. 19459413 Our GSM8K options were composed previously, with believing constantly off. Our text KL on 13 sources utilized 52 windows selected from the text mix by a different tool; 19459282 takes the mix’s own assessment windows, on which our file ratings
0.44 where the page reports 0.45. 19459400
Large designs operate on Modal, one GPU per task names every task, and 19459284 runs them. 19459285 downloads designs and the text mix utilized for calibration and assessment. 19459400
The runs behind the website utilized these devices, each with 8 CPU cores: 19659150 Design 19659151 GPU 19659152 Time SmolLM2-360M 19659154 L40S About 50 minutes to determine, in 9 parallel tasks Gemma 4 E2B, determining 19659157 H100 7– 26 minutes per task, in 4 tasks 19659159 Gemma 4 E2B, constructing a strategy H100 28– 58 minutes per strategy 19459286 makes plan of llama.cpp’s own tensor types, so the outcome runs in llama.cpp as it is. ratings public quantizations the exact same method as our strategies.
Both run inside the image in modal_gguf.pywhich develops llama.cpp with CUDA.
The Apple Silicon runtime checks out Lloyd– Max and trellis strategies
quantize.py --export composes a strategy in the format runtime_mlx/ checks out. A strategy with trellis codes likewise requires --trellis-rows 16,8which runs each trellis course along a row of its matrix, so that translating an entry checks out just that row. On Apple silicon, pip install . sets up mlxand mlx-lm; requirements.txt holds the PyTorch compiler reliances.
On Gemma 4 E2B at 2.5 GB, the trellis strategy has 26% lower KL than the Lloyd– Max strategy and translates about 22%slower. runtime_mlx/test_trellis_gen.py and runtime_mlx/test_trellis.py inspect the trellis kernels versus the matrices the quantizer developed.
python runtime_mlx/bench.py lamq path/to/export
A drafter makes the 1.2 GB file create 1.5 to 2 times quicker on Apple silicon
runtime_mlx/fast_q35.py deciphers the jam-packed 1.2 GB Qwen3.5-4B file one token at a time, directly from its trellis codes. runtime_mlx/drafter_q35.py includes speculative decoding. A little drafter proposes the 7 tokens after the design’s newest one in a single pass. The design then inspects all 8 positions in one pass over its coded weights(runtime_mlx/block_q35.pyand keeps the propositions that match its own option, so the text is precisely its own greedy text.
On an M5 Max, over 30 held-out triggers of as much as 256 tokens, generation with the drafter was 2.0 times as quick: 176 tokens a 2nd versus 88, attended alternation in one procedure on 2026-11-11. A check kept 4.3 tokens typically. Over 8 daily triggers it was 1.5 times as quick. The gain follows how foreseeable the text is: 2.5 times on algebra, 2.0 on code, 1.2 on chat and 1.05 on a story, and among the 30 held-out triggers ran slower with the drafter. The text matched plain greedy decoding on every timely. These are speeds after the very first token; checking out the timely takes about 0.1 s in any case.
2 later modifications remain in this code. Each was determined versus the variation before it on 40 paired generations: proposing from the 96,033 most typical tokens was 6.0% faster, and selecting the winning token inside the jam-packed item was 5.3% faster. Direct mode likewise allows self-calibrating adaptive alternative by default (adaptive=True; pass --no-adaptiveto disable): when rolling approval drops listed below the determined break-even ratio of drafted-check latency to serial-step latency (accounting for cache flush and drafter context replay), it stops briefly preparing for brief specific serial bursts and resumes when approval recuperates, maintaining bit-exact greedy parity while preventing downturns on low-acceptance prose.The drafter is a DSpine design, trained with SpecForge from the released DFlash drafter for Qwen3.5-4B on the 1.2 GB file’s own greedy responses. modal_drafter.py runs the training as one Modal task, and it cost $42. On 300 held-out triggers the skilled drafter gets 4.65 tokens per check, versus 4.30 for the stock DFlash drafter.
pack_drafter.pyshops the drafter at 4 bits a weight in the construct’s own code, without any calibration information: 464 MB versus 1.83 GB at 16 bits. On 60 held-out triggers the 4-bit drafter kept 4.35 tokens per check versus 4.38, and the text was the exact same. drafter_q35.py loads either kind and composes the jam-packed one out thick as it loads. The 4-bit drafter is the lm2-drafter-4bit release. It runs next to the jam-packed 1.2 GB file, the lm2-qwen35-4b-1.2gb release.
gh release download lm2-qwen35-4b-1.2gb –repo ShinMegamiBoson/lookatmyquant -D q35-1.2gb gh release download lm2-drafter-4bit –repo ShinMegamiBoson/lookatmyquant -D drafter-4bit python runtime_mlx/bench_drafted.py q35-1.2gb drafter-4bit”>
gh release download lm2-qwen35-4b-1.2gb --repo ShinMegamiBoson/lookatmyquant -D q35-1.2gb
gh release download lm2-drafter-4bit --repo ShinMegamiBoson/lookatmyquant -D drafter-4bit
python runtime_mlx/bench_drafted.py q35-1.2gb drafter-4bit
The criteria defaults to direct confirmation. Usage --mode chat-fork to attempt a 2nd prospect at an uncertain early position, or --mode copy-wide for recognized adjoining copies. Copy mode utilizes the source cursor and tunes the gate/up and down forecasts for 16-slot checks. Both modes validate all discharged tokens with the jam-packed target. To criteria the chat fork, run:
python runtime_mlx/bench_drafted.py q35-1.2gb drafter-4bit --mode chat-fork --prompts chat
For a copy standard, pass --mode copy-wide --heldout path/to/copy-prompts.jsonThe JSON entries bring source prompt_lenand input_ids from the jam-packed design’s
tokenizer.
The checks, and the packer for a drafter of your own:[pythonruntime_mlx/test_fast_q35pyLAMQ_Q35_PACKED=path/to/packedpythonruntime_mlx/test_block_q35pypythonruntime_mlx/check_drafterpypath/to/packedpath/to/drafterpythonruntime_mlx/bench_draftedpypath/to/packedpath/to/drafter–heldoutpath/to/drafter/heldout/heldoutjson–take30pythonpack_drafterpypath/to/drafterpath/to/drafter-4bit”>
python runtime_mlx/test_fast_q35.py
LAMQ_Q35_PACKED=path/to/packed python runtime_mlx/test_block_q35.py
python runtime_mlx/check_drafter.py path/to/packed path/to/drafter
python runtime_mlx/bench_drafted.py path/to/packed path/to/drafter --heldout path/to/drafter/heldout/heldout.json --take 30
python pack_drafter.py path/to/drafter path/to/drafter-4bit
runtime_mlx/pack_dense.py composes the jam-packed file from the fine-tuned construct’s thick weights, and runtime_mlx/modal_pack.py runs it on Modal. check_drafter.py compares the drafter in MLX with SpecForge’s own design on referral obstructs that drafter/reference.py composes.
Where whatever is
Course
What it holds
compiler.py
Procedure and designate
compiler_arch.py
How the compiler discovers a design’s layers, tables and phases, and the text mix
quantize.py
Develops a program or a strategy, then ratings it
row_importance.py rowgroup_alloc.py row_allocation.py compiler_artifact.py
Bits per row: each row’s significance, the allowance, the constructs it compares, and the jam-packed format that rates them
math_calibration.py checked_solutions.py other_text.py calc_check.py token_kinds.py
The text the dish adjusts and tweaks on, the checker for mentioned computations, and the type of token each position anticipates
finetune.py finetune_lib.py fast_delta.py
Fine-tuning that ends on the codes, and a quicker linear-attention guideline for Qwen3.5
math_bench.py math_eval.py text_kl.py
The exact-answer mathematics criteria, its runner for the Qwen3.5 household and Bonsai, and KL from the initial on text
gguf_compiler.py public_eval.py eval_tasks.py
The llama.cpp path, public quantizations and job standards
pyproject.toml runtime_mlx/runtime.py runtime_mlx/cli.py
Installable Apple Silicon runtime plan, merged load/ RuntimePython API, and lamq CLI
lamq_export.py runtime_mlx/
The export format and the Apple Silicon runtime
runtime_mlx/fast_q35.py block_q35.py block_argmax_q35.py pack_dense.py modal_pack.py
The runtime for jam-packed Qwen3.5-4B and Qwen3.8-27B files: one token at a time, a block of 8 or 16 tokens in one pass, and the packer
runtime_mlx/inkling/ runtime_mlx/test_inkling.py
Apple Silicon MLX runtime, 4 KB SRAMht2/ht3Metal kernels, and 4-bit DSpark block-8 speculative drafter for jam-packed Inkling-2-Mini designs
runtime_mlx/drafter_q35.py draft_vocab_q35.py convert_dspine.py check_drafter.py bench_drafted.py drafter/ modal_drafter.py pack_drafter.py
Speculative deciphering with a DSpine drafter, Tencent DSpine conversion, drafter training, and the drafter kept at 4 bits
opgraph/
The grammar: a design as a tensor network, and the rewrites and codes that act upon it
run_completeness.py make_report.py
The check that 9 released techniques are replicated precisely
analyze_compiler.py
The tables of strategy results
run_batch.py modal_app.py modal_gguf.py modal_fetch.py
Called tasks and the Modal runners
results/compiler/plans/ results/compiler/rows/
Every strategy, and the row prepare for Qwen3.5-4B at 1.2 GB
results/protocols/
The 2nd mathematics set of 216 issues, and the hashes of every concern stayed out of training text
cache/codes/
The e-means trellis codebooks at 1 to 4 bits
results/batch/ results/gguf/
The measurements and strategy runs behind the website
results/COMPILER.md results/COMPLETENESS.md
Reviews of the compiler on SmolLM2-360M and Qwen3.5-4B, and of the recreation check
site/
The site. Run python3 -m http.server 8000 inside it, then check out http://localhost:8000
site2/
The LookAtMyQuant 2 page. It runs the exact same method
tests/
Tests of the grammar and the codes
The tests require SmolLM2-360M and WikiText-2
Each test file operates on its own from the repository root, and all 7 surface in under a minute.[pythontests/test_framepypythontests/test_grammarpypythontests/test_codespypythontests/test_elementspypythontests/test_nestedpypythontests/test_rowgroup_allocpypythontests/test_reallocationpy”>
python tests/test_frame.py
python tests/test_grammar.py
python tests/test_codes.py
python tests/test_elements.py
python tests/test_nested.py
python tests/test_rowgroup_alloc.py
python tests/test_reallocation.py
runtime_mlx/test_core.py checks the runtime versus the quantizer’s mathematics. Run python runtime_mlx/make_ref.py Which composes 1.6 GB of
referral files.
runtime_mlx/test_fast_q35.py and runtime_mlx/test_block_q35.py examine the Qwen3.5 kernels versus NumPy items of individually unpacked codes, without any design. When LAMQ_Q35_PACKED names a jam-packed file, the 2nd likewise checks a block of tokens versus translating them one at a time. When LAMQ_Q38_PACKED names the jam-packed Qwen3.8-27B file, it runs the exact same examine a design whose output head is saved apart from its embedding.
What this repository neglects
Design weights.Whatever downloads from Hugging Face. The Qwen3.5-4B tasks anticipate that design in models/Qwen3.5-4B on the volume.
Criteria copies. eval_tasks.py downloads its criteria. On an offline maker it checks out exported copies from data/bench/which are not here. The Modal images run offline, so cache the criteria on the volume.
The script that put together site/data.jsonIt was a set of one-off actions. The source numbers are all in results/
LookAtMyQuant 2’s thick weights and result files. The 1.2 GB file is a release in the jam-packed format. The thick weights the standards operated on are not released, and neither are the per-problem outcomes behind site2/data.json or the script that composes the page’s numbers from them.
The drafter at 16 bits. The drafter as trained is 1.8 GB and is not released. The 4-bit one is a release file, and modal_drafter.py trains it.
this.
Run the LookAtMyQuant 2 dish on Qwen3.5-4B Each action is a called task; prints its command. The tasks require Qwen3.5-4B in 19459238 on the volume and the text mix( 19459239. Every file they make goes to on the volume.
Task Script What it does Our run 19659128 The initial responses science, code, direction and tool triggers, and MATH and GSM8K training issues. This likewise downloads MATH 19659129 The calibration set: the original’s greedy services to 384 MATH training issues, in 2,048-token windows 15 minutes, H100 19659131 The mathematics text the fine-tuning distils on: services to MATH issues and to 19459247 issues from another seed, with incorrect estimations fixed and just ideal responses kept 33 minutes, H100 Each output row’s Fisher info at digit-predicting positions 1 minute, H100 Post-training quantization: trellis codes along rows, digit tokens weighted 10 × in the Hessians, QEP at α=0.5, bits per row from the row strategy, e-means codebooks 37 minutes, H100 4,000 actions of distillation on hidden weights pulled towards their codes, with every layer re-quantized when from the initial weights 2.7 hours, B200 19659139, The 540-problem standard and a 2nd set of 216, for the initial, Bonsai, the develop and the fine-tuned develop 20 minutes a design, H100 19659142 KL from the initial on WikiText-2 and on the text mix’s
13 sources 19659143 1 minute a design, H100 19659144 , and run the very same code on 2 layers, 60 actions and 36 issues, in a couple of minutes each. 19459400 19459411 The row strategy. 19659145 checks out 19459263, which provides 1 to 4 bits to each of 1,520 row groups. made it. Its 19459265 command constructs and loads a control with a set split of rows, 19459266 designates rows inside the control’s jam-packed bytes with 19459267, and builds under the strategy and checks the jam-packed size versus it. 19459400 19459411 The codebooks. The 4 e-means codebooks the construct translates with remain in offers the earlier ones, and a strategy made before the modification keeps them unless the flag states otherwise. 19459400 Current equal-byte dish enhancements. 19459413 2 zero-byte additions are readily available in the compiler and tweak scripts: 19459400 19659146 Softmax-Fisher token table ranking ( in 19459272 and. Ranks connected embedding/output containers by a 50/50 mix of output Softmax-Fisher 19459274 and corpus frequency (and untied output heads by Softmax-Fisher alone). Mid-training equal-byte row reallocation ( 19459275 in. 19459413 Paired with , steps each layer’s continuous-vs-stored proximal KL space 19459278 to move bits towards deep unabsorbed layers while keeping specific file size (. Integrated with on Qwen3.5-4B at similar 1.23 GB (, WikiText-2 KL fell from 0.425 to 0.387 (− 9.0 %) and 13-source KL from 0.447 to 0.378 (− 15.5 %, enhancing all 13 sources). 19459411 Where these tasks vary from our run. 19459413 Our GSM8K options were composed previously, with believing constantly off. Our text KL on 13 sources utilized 52 windows selected from the text mix by a different tool; 19459282 takes the mix’s own assessment windows, on which our file ratings
0.44 where the page reports 0.45. 19459400
Large designs operate on Modal, one GPU per task names every task, and 19459284 runs them. 19459285 downloads designs and the text mix utilized for calibration and assessment. 19459400
The runs behind the website utilized these devices, each with 8 CPU cores: 19659150 Design 19659151 GPU 19659152 Time SmolLM2-360M 19659154 L40S About 50 minutes to determine, in 9 parallel tasks Gemma 4 E2B, determining 19659157 H100 7– 26 minutes per task, in 4 tasks 19659159 Gemma 4 E2B, constructing a strategy H100 28– 58 minutes per strategy 19459286 makes plan of llama.cpp’s own tensor types, so the outcome runs in llama.cpp as it is. ratings public quantizations the exact same method as our strategies.
Both run inside the image in modal_gguf.pywhich develops llama.cpp with CUDA.
The Apple Silicon runtime checks out Lloyd– Max and trellis strategies
quantize.py --export composes a strategy in the format runtime_mlx/ checks out. A strategy with trellis codes likewise requires --trellis-rows 16,8which runs each trellis course along a row of its matrix, so that translating an entry checks out just that row. On Apple silicon, pip install . sets up mlxand mlx-lm; requirements.txt holds the PyTorch compiler reliances.
On Gemma 4 E2B at 2.5 GB, the trellis strategy has 26% lower KL than the Lloyd– Max strategy and translates about 22%slower. runtime_mlx/test_trellis_gen.py and runtime_mlx/test_trellis.py inspect the trellis kernels versus the matrices the quantizer developed.
python runtime_mlx/bench.py lamq path/to/export
A drafter makes the 1.2 GB file create 1.5 to 2 times quicker on Apple silicon
runtime_mlx/fast_q35.py deciphers the jam-packed 1.2 GB Qwen3.5-4B file one token at a time, directly from its trellis codes. runtime_mlx/drafter_q35.py includes speculative decoding. A little drafter proposes the 7 tokens after the design’s newest one in a single pass. The design then inspects all 8 positions in one pass over its coded weights(runtime_mlx/block_q35.pyand keeps the propositions that match its own option, so the text is precisely its own greedy text.
On an M5 Max, over 30 held-out triggers of as much as 256 tokens, generation with the drafter was 2.0 times as quick: 176 tokens a 2nd versus 88, attended alternation in one procedure on 2026-11-11. A check kept 4.3 tokens typically. Over 8 daily triggers it was 1.5 times as quick. The gain follows how foreseeable the text is: 2.5 times on algebra, 2.0 on code, 1.2 on chat and 1.05 on a story, and among the 30 held-out triggers ran slower with the drafter. The text matched plain greedy decoding on every timely. These are speeds after the very first token; checking out the timely takes about 0.1 s in any case.
2 later modifications remain in this code. Each was determined versus the variation before it on 40 paired generations: proposing from the 96,033 most typical tokens was 6.0% faster, and selecting the winning token inside the jam-packed item was 5.3% faster. Direct mode likewise allows self-calibrating adaptive alternative by default (adaptive=True; pass --no-adaptiveto disable): when rolling approval drops listed below the determined break-even ratio of drafted-check latency to serial-step latency (accounting for cache flush and drafter context replay), it stops briefly preparing for brief specific serial bursts and resumes when approval recuperates, maintaining bit-exact greedy parity while preventing downturns on low-acceptance prose.The drafter is a DSpine design, trained with SpecForge from the released DFlash drafter for Qwen3.5-4B on the 1.2 GB file’s own greedy responses. modal_drafter.py runs the training as one Modal task, and it cost $42. On 300 held-out triggers the skilled drafter gets 4.65 tokens per check, versus 4.30 for the stock DFlash drafter.
pack_drafter.pyshops the drafter at 4 bits a weight in the construct’s own code, without any calibration information: 464 MB versus 1.83 GB at 16 bits. On 60 held-out triggers the 4-bit drafter kept 4.35 tokens per check versus 4.38, and the text was the exact same. drafter_q35.py loads either kind and composes the jam-packed one out thick as it loads. The 4-bit drafter is the lm2-drafter-4bit release. It runs next to the jam-packed 1.2 GB file, the lm2-qwen35-4b-1.2gb release.
gh release download lm2-qwen35-4b-1.2gb –repo ShinMegamiBoson/lookatmyquant -D q35-1.2gb gh release download lm2-drafter-4bit –repo ShinMegamiBoson/lookatmyquant -D drafter-4bit python runtime_mlx/bench_drafted.py q35-1.2gb drafter-4bit”>
gh release download lm2-qwen35-4b-1.2gb --repo ShinMegamiBoson/lookatmyquant -D q35-1.2gb
gh release download lm2-drafter-4bit --repo ShinMegamiBoson/lookatmyquant -D drafter-4bit
python runtime_mlx/bench_drafted.py q35-1.2gb drafter-4bit
The criteria defaults to direct confirmation. Usage --mode chat-fork to attempt a 2nd prospect at an uncertain early position, or --mode copy-wide for recognized adjoining copies. Copy mode utilizes the source cursor and tunes the gate/up and down forecasts for 16-slot checks. Both modes validate all discharged tokens with the jam-packed target. To criteria the chat fork, run:
python runtime_mlx/bench_drafted.py q35-1.2gb drafter-4bit --mode chat-fork --prompts chat
For a copy standard, pass --mode copy-wide --heldout path/to/copy-prompts.jsonThe JSON entries bring source prompt_lenand input_ids from the jam-packed design’s
tokenizer.
The checks, and the packer for a drafter of your own:[pythonruntime_mlx/test_fast_q35pyLAMQ_Q35_PACKED=path/to/packedpythonruntime_mlx/test_block_q35pypythonruntime_mlx/check_drafterpypath/to/packedpath/to/drafterpythonruntime_mlx/bench_draftedpypath/to/packedpath/to/drafter–heldoutpath/to/drafter/heldout/heldoutjson–take30pythonpack_drafterpypath/to/drafterpath/to/drafter-4bit”>
python runtime_mlx/test_fast_q35.py
LAMQ_Q35_PACKED=path/to/packed python runtime_mlx/test_block_q35.py
python runtime_mlx/check_drafter.py path/to/packed path/to/drafter
python runtime_mlx/bench_drafted.py path/to/packed path/to/drafter --heldout path/to/drafter/heldout/heldout.json --take 30
python pack_drafter.py path/to/drafter path/to/drafter-4bit
- 13 sources 19659143 1 minute a design, H100 19659144 , and run the very same code on 2 layers, 60 actions and 36 issues, in a couple of minutes each. 19459400 19459411 The row strategy. 19659145 checks out 19459263, which provides 1 to 4 bits to each of 1,520 row groups. made it. Its 19459265 command constructs and loads a control with a set split of rows, 19459266 designates rows inside the control’s jam-packed bytes with 19459267, and builds under the strategy and checks the jam-packed size versus it. 19459400 19459411 The codebooks. The 4 e-means codebooks the construct translates with remain in offers the earlier ones, and a strategy made before the modification keeps them unless the flag states otherwise. 19459400 Current equal-byte dish enhancements. 19459413 2 zero-byte additions are readily available in the compiler and tweak scripts: 19459400 19659146 Softmax-Fisher token table ranking ( in 19459272 and. Ranks connected embedding/output containers by a 50/50 mix of output Softmax-Fisher 19459274 and corpus frequency (and untied output heads by Softmax-Fisher alone). Mid-training equal-byte row reallocation ( 19459275 in. 19459413 Paired with , steps each layer’s continuous-vs-stored proximal KL space 19459278 to move bits towards deep unabsorbed layers while keeping specific file size (. Integrated with on Qwen3.5-4B at similar 1.23 GB (, WikiText-2 KL fell from 0.425 to 0.387 (− 9.0 %) and 13-source KL from 0.447 to 0.378 (− 15.5 %, enhancing all 13 sources). 19459411 Where these tasks vary from our run. 19459413 Our GSM8K options were composed previously, with believing constantly off. Our text KL on 13 sources utilized 52 windows selected from the text mix by a different tool; 19459282 takes the mix’s own assessment windows, on which our file ratings
Large designs operate on Modal, one GPU per task names every task, and 19459284 runs them. 19459285 downloads designs and the text mix utilized for calibration and assessment. 19459400
The runs behind the website utilized these devices, each with 8 CPU cores: 19659150 Design 19659151 GPU 19659152 Time SmolLM2-360M 19659154 L40S About 50 minutes to determine, in 9 parallel tasks Gemma 4 E2B, determining 19659157 H100 7– 26 minutes per task, in 4 tasks 19659159 Gemma 4 E2B, constructing a strategy H100 28– 58 minutes per strategy 19459286 makes plan of llama.cpp’s own tensor types, so the outcome runs in llama.cpp as it is. ratings public quantizations the exact same method as our strategies.
Both run inside the image in modal_gguf.pywhich develops llama.cpp with CUDA.
The Apple Silicon runtime checks out Lloyd– Max and trellis strategies
quantize.py --export composes a strategy in the format runtime_mlx/ checks out. A strategy with trellis codes likewise requires --trellis-rows 16,8which runs each trellis course along a row of its matrix, so that translating an entry checks out just that row. On Apple silicon, pip install . sets up mlxand mlx-lm; requirements.txt holds the PyTorch compiler reliances.
On Gemma 4 E2B at 2.5 GB, the trellis strategy has 26% lower KL than the Lloyd– Max strategy and translates about 22%slower. runtime_mlx/test_trellis_gen.py and runtime_mlx/test_trellis.py inspect the trellis kernels versus the matrices the quantizer developed.
python runtime_mlx/bench.py lamq path/to/export
A drafter makes the 1.2 GB file create 1.5 to 2 times quicker on Apple silicon
runtime_mlx/fast_q35.py deciphers the jam-packed 1.2 GB Qwen3.5-4B file one token at a time, directly from its trellis codes. runtime_mlx/drafter_q35.py includes speculative decoding. A little drafter proposes the 7 tokens after the design’s newest one in a single pass. The design then inspects all 8 positions in one pass over its coded weights(runtime_mlx/block_q35.pyand keeps the propositions that match its own option, so the text is precisely its own greedy text.
On an M5 Max, over 30 held-out triggers of as much as 256 tokens, generation with the drafter was 2.0 times as quick: 176 tokens a 2nd versus 88, attended alternation in one procedure on 2026-11-11. A check kept 4.3 tokens typically. Over 8 daily triggers it was 1.5 times as quick. The gain follows how foreseeable the text is: 2.5 times on algebra, 2.0 on code, 1.2 on chat and 1.05 on a story, and among the 30 held-out triggers ran slower with the drafter. The text matched plain greedy decoding on every timely. These are speeds after the very first token; checking out the timely takes about 0.1 s in any case.
2 later modifications remain in this code. Each was determined versus the variation before it on 40 paired generations: proposing from the 96,033 most typical tokens was 6.0% faster, and selecting the winning token inside the jam-packed item was 5.3% faster. Direct mode likewise allows self-calibrating adaptive alternative by default (adaptive=True; pass --no-adaptiveto disable): when rolling approval drops listed below the determined break-even ratio of drafted-check latency to serial-step latency (accounting for cache flush and drafter context replay), it stops briefly preparing for brief specific serial bursts and resumes when approval recuperates, maintaining bit-exact greedy parity while preventing downturns on low-acceptance prose.The drafter is a DSpine design, trained with SpecForge from the released DFlash drafter for Qwen3.5-4B on the 1.2 GB file’s own greedy responses. modal_drafter.py runs the training as one Modal task, and it cost $42. On 300 held-out triggers the skilled drafter gets 4.65 tokens per check, versus 4.30 for the stock DFlash drafter.
pack_drafter.pyshops the drafter at 4 bits a weight in the construct’s own code, without any calibration information: 464 MB versus 1.83 GB at 16 bits. On 60 held-out triggers the 4-bit drafter kept 4.35 tokens per check versus 4.38, and the text was the exact same. drafter_q35.py loads either kind and composes the jam-packed one out thick as it loads. The 4-bit drafter is the lm2-drafter-4bit release. It runs next to the jam-packed 1.2 GB file, the lm2-qwen35-4b-1.2gb release.
gh release download lm2-qwen35-4b-1.2gb –repo ShinMegamiBoson/lookatmyquant -D q35-1.2gb gh release download lm2-drafter-4bit –repo ShinMegamiBoson/lookatmyquant -D drafter-4bit python runtime_mlx/bench_drafted.py q35-1.2gb drafter-4bit”>
gh release download lm2-qwen35-4b-1.2gb --repo ShinMegamiBoson/lookatmyquant -D q35-1.2gb
gh release download lm2-drafter-4bit --repo ShinMegamiBoson/lookatmyquant -D drafter-4bit
python runtime_mlx/bench_drafted.py q35-1.2gb drafter-4bit
Both run inside the image in
modal_gguf.pywhich develops llama.cpp with CUDA.
The Apple Silicon runtime checks out Lloyd– Max and trellis strategies
quantize.py --export composes a strategy in the format runtime_mlx/ checks out. A strategy with trellis codes likewise requires --trellis-rows 16,8which runs each trellis course along a row of its matrix, so that translating an entry checks out just that row. On Apple silicon, pip install . sets up mlxand mlx-lm; requirements.txt holds the PyTorch compiler reliances.
On Gemma 4 E2B at 2.5 GB, the trellis strategy has 26% lower KL than the Lloyd– Max strategy and translates about 22%slower. runtime_mlx/test_trellis_gen.py and runtime_mlx/test_trellis.py inspect the trellis kernels versus the matrices the quantizer developed.
python runtime_mlx/bench.py lamq path/to/export
A drafter makes the 1.2 GB file create 1.5 to 2 times quicker on Apple silicon
runtime_mlx/fast_q35.py deciphers the jam-packed 1.2 GB Qwen3.5-4B file one token at a time, directly from its trellis codes. runtime_mlx/drafter_q35.py includes speculative decoding. A little drafter proposes the 7 tokens after the design’s newest one in a single pass. The design then inspects all 8 positions in one pass over its coded weights(runtime_mlx/block_q35.pyand keeps the propositions that match its own option, so the text is precisely its own greedy text.
On an M5 Max, over 30 held-out triggers of as much as 256 tokens, generation with the drafter was 2.0 times as quick: 176 tokens a 2nd versus 88, attended alternation in one procedure on 2026-11-11. A check kept 4.3 tokens typically. Over 8 daily triggers it was 1.5 times as quick. The gain follows how foreseeable the text is: 2.5 times on algebra, 2.0 on code, 1.2 on chat and 1.05 on a story, and among the 30 held-out triggers ran slower with the drafter. The text matched plain greedy decoding on every timely. These are speeds after the very first token; checking out the timely takes about 0.1 s in any case.
2 later modifications remain in this code. Each was determined versus the variation before it on 40 paired generations: proposing from the 96,033 most typical tokens was 6.0% faster, and selecting the winning token inside the jam-packed item was 5.3% faster. Direct mode likewise allows self-calibrating adaptive alternative by default (adaptive=True; pass --no-adaptiveto disable): when rolling approval drops listed below the determined break-even ratio of drafted-check latency to serial-step latency (accounting for cache flush and drafter context replay), it stops briefly preparing for brief specific serial bursts and resumes when approval recuperates, maintaining bit-exact greedy parity while preventing downturns on low-acceptance prose.The drafter is a DSpine design, trained with SpecForge from the released DFlash drafter for Qwen3.5-4B on the 1.2 GB file’s own greedy responses. modal_drafter.py runs the training as one Modal task, and it cost $42. On 300 held-out triggers the skilled drafter gets 4.65 tokens per check, versus 4.30 for the stock DFlash drafter.
pack_drafter.pyshops the drafter at 4 bits a weight in the construct’s own code, without any calibration information: 464 MB versus 1.83 GB at 16 bits. On 60 held-out triggers the 4-bit drafter kept 4.35 tokens per check versus 4.38, and the text was the exact same. drafter_q35.py loads either kind and composes the jam-packed one out thick as it loads. The 4-bit drafter is the lm2-drafter-4bit release. It runs next to the jam-packed 1.2 GB file, the lm2-qwen35-4b-1.2gb release.
gh release download lm2-qwen35-4b-1.2gb –repo ShinMegamiBoson/lookatmyquant -D q35-1.2gb gh release download lm2-drafter-4bit –repo ShinMegamiBoson/lookatmyquant -D drafter-4bit python runtime_mlx/bench_drafted.py q35-1.2gb drafter-4bit”>
gh release download lm2-qwen35-4b-1.2gb --repo ShinMegamiBoson/lookatmyquant -D q35-1.2gb gh release download lm2-drafter-4bit --repo ShinMegamiBoson/lookatmyquant -D drafter-4bit python runtime_mlx/bench_drafted.py q35-1.2gb drafter-4bit
The criteria defaults to direct confirmation. Usage --mode chat-fork to attempt a 2nd prospect at an uncertain early position, or --mode copy-wide for recognized adjoining copies. Copy mode utilizes the source cursor and tunes the gate/up and down forecasts for 16-slot checks. Both modes validate all discharged tokens with the jam-packed target. To criteria the chat fork, run:
python runtime_mlx/bench_drafted.py q35-1.2gb drafter-4bit --mode chat-fork --prompts chat
For a copy standard, pass --mode copy-wide --heldout path/to/copy-prompts.jsonThe JSON entries bring source prompt_lenand input_ids from the jam-packed design’s
tokenizer.
The checks, and the packer for a drafter of your own:[pythonruntime_mlx/test_fast_q35pyLAMQ_Q35_PACKED=path/to/packedpythonruntime_mlx/test_block_q35pypythonruntime_mlx/check_drafterpypath/to/packedpath/to/drafterpythonruntime_mlx/bench_draftedpypath/to/packedpath/to/drafter–heldoutpath/to/drafter/heldout/heldoutjson–take30pythonpack_drafterpypath/to/drafterpath/to/drafter-4bit”>
python runtime_mlx/test_fast_q35.py LAMQ_Q35_PACKED=path/to/packed python runtime_mlx/test_block_q35.py python runtime_mlx/check_drafter.py path/to/packed path/to/drafter python runtime_mlx/bench_drafted.py path/to/packed path/to/drafter --heldout path/to/drafter/heldout/heldout.json --take 30 python pack_drafter.py path/to/drafter path/to/drafter-4bit
runtime_mlx/pack_dense.py composes the jam-packed file from the fine-tuned construct’s thick weights, and runtime_mlx/modal_pack.py runs it on Modal. check_drafter.py compares the drafter in MLX with SpecForge’s own design on referral obstructs that drafter/reference.py composes.
Where whatever is
| Course | What it holds |
|---|---|
compiler.py |
Procedure and designate |
compiler_arch.py |
How the compiler discovers a design’s layers, tables and phases, and the text mix |
quantize.py |
Develops a program or a strategy, then ratings it |
row_importance.py rowgroup_alloc.py row_allocation.py compiler_artifact.py |
Bits per row: each row’s significance, the allowance, the constructs it compares, and the jam-packed format that rates them |
math_calibration.py checked_solutions.py other_text.py calc_check.py token_kinds.py |
The text the dish adjusts and tweaks on, the checker for mentioned computations, and the type of token each position anticipates |
finetune.py finetune_lib.py fast_delta.py |
Fine-tuning that ends on the codes, and a quicker linear-attention guideline for Qwen3.5 |
math_bench.py math_eval.py text_kl.py |
The exact-answer mathematics criteria, its runner for the Qwen3.5 household and Bonsai, and KL from the initial on text |
gguf_compiler.py public_eval.py eval_tasks.py |
The llama.cpp path, public quantizations and job standards |
pyproject.toml runtime_mlx/runtime.py runtime_mlx/cli.py |
Installable Apple Silicon runtime plan, merged load/ RuntimePython API, and lamq CLI |
lamq_export.py runtime_mlx/ |
The export format and the Apple Silicon runtime |
runtime_mlx/fast_q35.py block_q35.py block_argmax_q35.py pack_dense.py modal_pack.py |
The runtime for jam-packed Qwen3.5-4B and Qwen3.8-27B files: one token at a time, a block of 8 or 16 tokens in one pass, and the packer |
runtime_mlx/inkling/ runtime_mlx/test_inkling.py |
Apple Silicon MLX runtime, 4 KB SRAMht2/ht3Metal kernels, and 4-bit DSpark block-8 speculative drafter for jam-packed Inkling-2-Mini designs |
runtime_mlx/drafter_q35.py draft_vocab_q35.py convert_dspine.py check_drafter.py bench_drafted.py drafter/ modal_drafter.py pack_drafter.py |
Speculative deciphering with a DSpine drafter, Tencent DSpine conversion, drafter training, and the drafter kept at 4 bits |
opgraph/ |
The grammar: a design as a tensor network, and the rewrites and codes that act upon it |
run_completeness.py make_report.py |
The check that 9 released techniques are replicated precisely |
analyze_compiler.py |
The tables of strategy results |
run_batch.py modal_app.py modal_gguf.py modal_fetch.py |
Called tasks and the Modal runners |
results/compiler/plans/ results/compiler/rows/ |
Every strategy, and the row prepare for Qwen3.5-4B at 1.2 GB |
results/protocols/ |
The 2nd mathematics set of 216 issues, and the hashes of every concern stayed out of training text |
cache/codes/ |
The e-means trellis codebooks at 1 to 4 bits |
results/batch/ results/gguf/ |
The measurements and strategy runs behind the website |
results/COMPILER.md results/COMPLETENESS.md |
Reviews of the compiler on SmolLM2-360M and Qwen3.5-4B, and of the recreation check |
site/ |
The site. Run python3 -m http.server 8000 inside it, then check out http://localhost:8000 |
site2/ |
The LookAtMyQuant 2 page. It runs the exact same method |
tests/ |
Tests of the grammar and the codes |
The tests require SmolLM2-360M and WikiText-2
Each test file operates on its own from the repository root, and all 7 surface in under a minute.[pythontests/test_framepypythontests/test_grammarpypythontests/test_codespypythontests/test_elementspypythontests/test_nestedpypythontests/test_rowgroup_allocpypythontests/test_reallocationpy”>
python tests/test_frame.py python tests/test_grammar.py python tests/test_codes.py python tests/test_elements.py python tests/test_nested.py python tests/test_rowgroup_alloc.py python tests/test_reallocation.py
runtime_mlx/test_core.py checks the runtime versus the quantizer’s mathematics. Run python runtime_mlx/make_ref.py Which composes 1.6 GB of
referral files.
runtime_mlx/test_fast_q35.py and runtime_mlx/test_block_q35.py examine the Qwen3.5 kernels versus NumPy items of individually unpacked codes, without any design. When LAMQ_Q35_PACKED names a jam-packed file, the 2nd likewise checks a block of tokens versus translating them one at a time. When LAMQ_Q38_PACKED names the jam-packed Qwen3.8-27B file, it runs the exact same examine a design whose output head is saved apart from its embedding.
What this repository neglects
models/Qwen3.5-4B on the volume.eval_tasks.py downloads its criteria. On an offline maker it checks out exported copies from data/bench/which are not here. The Modal images run offline, so cache the criteria on the volume.site/data.jsonIt was a set of one-off actions. The source numbers are all in results/site2/data.json or the script that composes the page’s numbers from them.modal_drafter.py trains it.Discover more from PMN S.P.O.R.T.S - A PRIME MEDIA NETWORK BRAND
Subscribe to get the latest posts sent to your email.





