You can deploy this model and start earning money today!

morgan/
qwen38-27b-mtp-dev40-cut12-q4-g64-q2-rerank

Name: qwen38-27b-mtp-dev40-cut12-q4-g64-q2-rerank
Availability: InStock

multi-token-prediction

Views

No views yet

Qwen 3.8 27B MTP Dev40 cut-12k — Q4/G64 + QKV islands and Q2 rerank

This repository contains a proposal-head artifact for native-MTP speculative decoding with Qwen 3.8 27B. It is not a standalone language model. The fixed target backbone and tokenizer are intentionally not included.

The artifact composes the selected Dev40 cut-12k MTP head with the public promoted Qwen 3.8 challenge head ABI:

The selected 15-tensor BF16 export supplies the trained core values.
MLX 0.32.0 packs eight trained matrices with affine 4-bit, group-size-64 quantization. Seven normalization vectors remain BF16. Together these form the 31-tensor packed core.
The promoted Q, K, and V precision-island index arrays are copied byte-for-byte. The six island tensors are exact BF16 gathers from the selected Dev40 Q, K, and V matrices under those preserved indices.
The promoted affine-2 draft_lm_head weight, scales, and biases are copied byte-for-byte. The compatible runtime uses this triple for its 32-token shortlist and performs the promoted exact affine-4 rerank.

The head only proposes future tokens. The fixed target model verifies every emitted token, and the trusted parent owns exact target top-two evidence, row accounting, rollback, replay, and scoring. This artifact contains no executable remote code.

Exact artifact identity

Field	Value
`model.safetensors` bytes	`427742696`
Raw file SHA-256	`5040ec7eb47e4d852990824b87205220d6ee12517dcfdd15206cb01a91758389`
Model-only tree SHA-256	`1f207eb91faff87e8dcdc0175ebfff0b5b15b2fb0477e8e8ea143ac70e8b9335`
Tensor count	`40`
Tensor payload bytes	`427738112`
`config.json` bytes	`3570`
`config.json` SHA-256	`fb2a5bd03bc4a2ac6a161f2f2eb5bf9fc3e4e23bf50a55c2ba41fe6b692d2b95`
Fine-tuned BF16 source bytes	`849400336`
Fine-tuned BF16 source SHA-256	`fdd4a8dbaad207e9d9a296a43e2e9495dedee7a24f367a58b2de04ff4b15ae77`
Fine-tuned BF16 model-only tree	`90a87a1be1ef0f38a6817305e94d3c14fb0d7158f70a5009b6d98933add74c31`
Packed Q4/G64 core SHA-256	`fee57c5947f025ead8f8989655375fce15c5e2fb73a8832b4c8d0a55e6de2931`

The model-only tree digest is SHA-256 over this exact line, including its final newline:

5040ec7eb47e4d852990824b87205220d6ee12517dcfdd15206cb01a91758389 model.safetensors

Preserved frontier payloads

The source frontier head is amal-david/qwen38-mtp-head-q2-q4-rerank-v1@ae6282749a52e052496dd5300b4aa441df7301e8. Its raw model SHA-256 is d038fd41e2d5dab1b3905c115d859fdc98dfbfde9862c14ebb82c2b3247ec2f1, and its model-only tree SHA-256 is 559b24ebca354018e4402fdb1f5af1afe5a0721bd2ebf04133500d846f7d5f71.

Composition validation proved that the q2 readout triple, all three island index arrays, and tensor wire order are byte-exact copies of that frontier artifact. It also proved that all 31 packed core tensors match the Dev40 Q4/G64 build and that every island row exactly matches its BF16 source row.

Training evidence

The selected checkpoint came from job q38-d40v2-cut12000-16cf28f-20260818t163000-a16c01 and W&B run q38mtp-20260818-163155-e97e8d11. Training used source commit 16cf28f37509900a03ce0c022f1cf051bf27bc4f, Speculators commit c80a242a735519cf9781cacb0894e6bfac821c1a, the Dev40v2 minimum-step-12000 dataset selection, K=4, acceptance-loss beta 0.6, seed 42, sequence length 11264, a 12,000-step cosine horizon with 600 warmup steps, AdamW learning rate 5e-6, and Muon learning rate 5e-5.

The common 4,049-row validation objective was 0.545973635. A same-protocol 256-prompt acceptance evaluation measured mean committed length 3.5363566990079582, accepted-draft rate 0.6340891747519896, and per-depth acceptance 0.8397470838329881 / 0.6843998691812929 / 0.5579417856753516 / 0.4542679603183255. The best and final exports had the same acceptance metrics. CUDA serial-stream agreement was 113/256 prompts; that exactness diagnostic is non-authoritative because Qwen GDN is numerically non-invariant on the CUDA path. Apple MLX supplied deployment exactness.

Apple MLX validation

The final local screen ran on an Apple M4 Pro Mac mini with 48 GiB of memory, macOS 26.5.2 (25F84), Xcode 26.6, Swift 6.3.3, and Apple Metal 32023.883. It used the live promoted source 0c90733d383f6b987a29682bf9eb9458a6172bfa, a 512-token seed, a 128-token decode window, MTP depth 8, fan auto, and three consecutive real GPU samples at or below 40C before each measured leg.

Arm	Gate GPU	Seconds/token	Exact tokens	Ledger
Current frontier control	`38.2417641 C`	`0.0557334144`	`128/128`	`134/134`
Dev40 cut-12k	`39.3879738 C`	`0.0560098598`	`128/128`	`136/136`

Both legs passed parity, reported zero residual divergence, closed their reference-row ledgers, and loaded the exact declared head provenance. The Dev40/control ratio was 0.9950643438, so Dev40 was 0.4935656% slower on this one-prompt M4 screen. This result is fully disclosed and treated as directional/noise; only the official M5 runner can determine ranked promotion.

The exact local evidence SHA-256 values are:

control report: d2db2513063d8f39510346e9cff0662e8b3fc7b88ae2f00bbc54d83da4e8733f
Dev40 report: d77906ad32cd060b340930cb6fadd0dc019677f9ac61d0cc15e2c3fc59fa1672
comparison: 43ba389f146d3f8d34b4ac6ed1847302a9e02d605e0e79faee63ad346fa10eb8
thermal log: 988078ab1ac6b5023a42b7bea578eb183b9945798687a68a55c4d9387c924b59
gate log: e41bfa8bfe63aa9c4949a793a430b0135fde01d61abb198b84cf8b6fc7bd7e10

See provenance.json for the machine-readable record. Any consuming manifest must pin an immutable repository revision and independently reproduce the declared byte count and tree digest.

morgan/qwen38-27b-mtp-dev40-cut12-q4-g64-q2-rerank