You can deploy this model and start earning money today!

morgan/
qwen38-27b-mtp-dev20-k4b06-lr1e5-q4-g64-islands

Name: qwen38-27b-mtp-dev20-k4b06-lr1e5-q4-g64-islands
Availability: InStock

multi-token-prediction

Views

No views yet

Qwen 3.8 27B MTP Dev20 K4 β0.6 LR 1e-5 — Q4/G64 + BF16 QKV islands

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 target backbone and tokenizer are intentionally not included.

The head began with the K=4, acceptance-loss weight β=0.6, learning-rate 1e-5 Dev20 fine-tuning arm. Its eight matrix weights were converted with MLX 0.32.0 affine 4-bit/group-64 quantization; seven one-dimensional normalization weights remain BF16. The resulting 31-tensor packed head is extended with six proposal-only precision-island tensors:

precision_islands.q.weight: BF16 [1024, 5120]
precision_islands.q.indices: I32 [1024]
precision_islands.k.weight: BF16 [1024, 5120]
precision_islands.k.indices: I32 [1024]
precision_islands.v.weight: BF16 [1024, 5120]
precision_islands.v.indices: I32 [1024]

For this causal composition, the island index arrays are copied exactly from the promoted public QKV-islands mechanism at amal-david/qwen38-mtp-head-q4-qkv-islands-v1@8081fee431e304076b6f6296d6eb5dc7a3fc91af. All K and V rows are retained, along with its 1,024 selected Q rows. The BF16 island weights are exact row gathers from this fine-tuned head. No selector was recomputed.

The compatible challenge runtime installs these rows only on the MTP proposal attention projection. The fixed target model still verifies every emitted token. The artifact does not contain executable remote code.

Exact artifact identity

Field	Value
`model.safetensors` bytes	`270404736`
Raw file SHA-256	`a6f2640ba96f99ef18b7e4fadd80bca432d93a2dfc3b331c6dc3fe98277eced8`
Model-only tree SHA-256	`5cbc55373ad5bc3339d9143e2cc0216c57ba29206d9e93a10d6e464feac84765`
Tensor count	`37`
Fine-tuned BF16 source SHA-256	`0a94bbb07f134c6347a788775087f47e19dbc6cc6bc35b5b8cd5a82adc4b05ca`
Packed Q4/G64 base SHA-256	`a6df85360b805b747a60c1a51c6784cf624d374f1fed63b2c272a7a0bdc2cf9f`

The model-only tree digest is SHA-256 over this exact line:

a6f2640ba96f99ef18b7e4fadd80bca432d93a2dfc3b331c6dc3fe98277eced8 model.safetensors

Validation

Exact 37-tensor ABI validated with MLX 0.32.0.
All 31 packed tensors are byte-identical to the fine-tuned Q4/G64 base.
All three index arrays are byte-identical to the promoted public mechanism.
All island rows are byte-identical gathers from the fine-tuned BF16 source.
Reapplying the constructor to the incumbent sources reproduced the promoted 270,404,624-byte artifact byte-for-byte, including raw SHA-256 517bb133d7ca6e228a5129710b3cb2c25aa9944753b9f9a225fa1e8135df5e65.

See provenance.json for the machine-readable record.

morgan/qwen38-27b-mtp-dev20-k4b06-lr1e5-q4-g64-islands

Qwen 3.8 27B MTP Dev20 K4 β0.6 LR 1e-5 — Q4/G64 + BF16 QKV islands

Exact artifact identity

Validation

morgan/
qwen38-27b-mtp-dev20-k4b06-lr1e5-q4-g64-islands