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 LR3 trained MTP head with the public promoted Qwen
3.8 challenge head ABI. The composition changes proposal values while
preserving the promoted runtime mechanisms:
The 15-tensor LR3 BF16 export supplies the trained core values.
Eight trained matrices are packed with MLX 0.32.0 affine 4-bit,
group-size-64 quantization. Seven normalization vectors remain BF16. These
values form the 31-tensor packed core.
The promoted Q, K, and V precision-island index arrays are copied
byte-for-byte. The six BF16 island tensors are exact gathers from the LR3
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 only to build a
32-token shortlist, then performs the promoted exact affine-4 rerank.
The Q-row selector is intentionally not recomputed. Preserving the public
frontier layout holds the runtime and shortlist mechanism fixed so the Apple
comparison isolates the trained core and gathered island values.
The head only proposes future tokens. The fixed target model still verifies
every emitted token, and the trusted parent still owns exact target top-two
evidence, row accounting, rollback, replay, and scoring. The artifact contains
no executable remote code.
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.
The LR3 compositor verified that the q2 readout triple, all three precision-
island index arrays, and tensor wire order are byte-exact copies of that
frontier artifact. It also verified that all 31 packed core tensors match the
LR3 Q4/G64 build and every island row matches its LR3 BF16 source row.
Training evidence
LR3 was trained from source commit
79cce6cb594aefc53294cdb0aabb277997ac863d. It used a K=4 objective with
acceptance-loss beta 0.6, a 4,000-step cosine horizon, 200 warmup steps,
AdamW learning rate 3e-6, and Muon learning rate 3e-5. The selected
checkpoint was step 2,469 and was also the final checkpoint. Its validation
objective was 0.5906565.
The source is the 94,562-row safe Dev20 tree with SHA-256
85a444fe8476085ac92f1f110e0b01aca33cc98b74758ea2019910d121942130.
Training uses the 20,000 rows with the smallest
BLAKE2b-128(seed:source_index) keys. The selected-row tree SHA-256 is
d17ed48d18d0c4b260bec9c478fcab98c54c9104df0cbad02284cc68a8703211,
and the exact membership-manifest SHA-256 is
57c62cb68d4ede120698acdc1ab53d6dcf0e9ac781e2b716b6f4ffca76c8fff1.
The prepared cache contains those 20,000 rows plus 4,049 held-out validation
rows, for 24,049 rows total.
The CUDA directional diagnostic measured mean committed length
3.4625480154 and accepted-draft rate 0.6156370. It matched the CUDA serial
stream on 96 of 256 samples. That exactness result is diagnostic-only because
the CUDA Qwen-GDN path is numerically non-invariant; Apple MLX remains the
deployment correctness authority.
The matched control measured committed length 3.4360958033064857, accepted-
draft rate 0.6090239508266214, and per-depth acceptance
0.8256676557863502 / 0.6611911827045358 / 0.5277660025434506 / 0.4214709622721492. LR3 improved committed length by
0.0264522120584313, or about 0.7698%, in this directional diagnostic.
LR3 per-depth acceptance was 0.832266325 / 0.668053777 / 0.533504055 / 0.428723858.
Raw file size, raw SHA-256, model-only tree digest, config identity, tensor
count, and tensor payload bytes were recomputed locally.
All 40 tensor names and wire order match the promoted q2/q4 rerank ABI.
All 31 packed core tensors are byte-identical to the LR3 MLX Q4/G64 build.
The q2 readout triple and all three island index arrays are byte-identical to
the promoted frontier head.
Every Q, K, and V BF16 island row is an exact gather from the LR3 BF16 source.
The exact current campaign source builds in release mode with the frozen
dependency graph when this head manifest is the only source-tree change.
Two counterbalanced, real-40C-gated Apple sequences completed with exact
tokens, zero residual divergence, closed ledgers, fan auto, and exact
per-arm head provenance.
Apple MLX selection
The terminal deployment screen ran on an Apple M4 Pro Mac mini with 48 GiB of
memory, macOS 26.5.2 (25F84), Xcode 26.6 (17F113), Swift 6.3.3, and Metal
toolchain com.apple.dt.toolchain.Metal.32023.883. It used exact campaign base
a187ec6c479bfec7e62fbb55a68c39aa128ceca7, a 128-token window, MTP depth 8,
natural cooling, fan auto, and three real GPU cool gates at or below 40C per
arm. These are directional local results, not official ranked scores.
The first sequence used order control/LR3/LR3/control:
Arm
Head
MTP seconds/token
01
Current frontier
0.05670436006039381
02
LR3
0.055505343712866306
03
LR3
0.055216390639543533
04
Current frontier
0.056804000400006771
The control mean was 0.0567541802302003 seconds/token and the LR3 mean was
0.0553608671762049. LR3 was 2.454996% faster by the matched absolute-time
criterion.
The second sequence used order LR3/B05/B05/LR3:
Arm
Head
MTP seconds/token
01
LR3
0.055070945061743259
02
B05
0.055201351642608643
03
B05
0.055350874550640583
04
LR3
0.055238242261111736
The LR3 mean was 0.0551545936614275 seconds/token and the B05 mean was
0.0552761130966246. LR3 was 0.219841% faster, which is inside the
predeclared ±0.5% noise band. LR3 was selected because it decisively beat the
current frontier and did not lose the B05 comparison.
Every arm emitted all 128 expected tokens, reported zero residual divergence,
and closed its reference-row ledger. LR3 and B05 each had 17 rounds, effective
mean draft length 6.5294117647058822, and accepted-draft rate 1.0 on this
public prompt. The current-head control had 20 rounds, effective mean draft
length 5.7, and accepted-draft rate 0.956140350877193.
See provenance.json for the machine-readable record. Any consuming manifest
must pin an immutable Hugging Face revision and independently reproduce the
declared artifact identity before use.
