Hugging Face: Rust specialist (scale400) · Rust/prose mixture (mixed400). Both require the pinned Qwen3-8B target; each page includes direct weight downloads and serving instructions. All six experimental checkpoints remain available in the GitHub release.
An existing EAGLE-3 draft adapted for frozen Qwen3-8B achieved 15.63% lower paired latency than the upstream draft on 32 fresh Rust prompts in the matched-control study (exploratory 95% interval: 14.26–17.11%). H100 80GB, BF16, greedy generation, one request at a time.
Read the full write-up → · Reproduce the analysis · Browse artifacts · Model cards · Download weights
- Replication: the initial 512-example / 100-update gain repeated at about 10.7% with two training seeds and new serving sessions.
- Data versus training: a 2,048-example pool added little at 100 updates. At 400 updates, it was 1.76% faster than the 512-example control (0.40–2.84%). Longer training produced the larger gain.
- Trade-offs: Rust-only adaptation slowed an eight-prompt JSON diagnostic. Mixed data improved prose relative to Rust-only training, while slowing Rust slightly; JSON recovery was inconclusive.
- Output checks: 3,120 follow-up requests, 120 unique prompts, identical text and completion-token counts across all tested drafts/sessions. This is finite greedy-output agreement, not proof of sampling equivalence or functional correctness.
This is draft adaptation and evaluation, not a new architecture, target fine-tuning, quantization, or a better Rust coding model. EAGLE-3, the pretrained draft, SGLang, and SpecForge are upstream work. Concurrency-dependent output variation in the original pilot remains unresolved.
Python 3.11 or later, standard library only:
python scripts/verify_release.pyThis recomputes all 90 follow-up comparisons and native counters from the distributed per-request evidence, checks output parity and artifact hashes, and reproduces the pilot comparisons. It makes no network requests and starts no training.
Optional local checks: python -m pip install pytest==8.4.2 then python -m pytest -q. One optional integration check requires the pinned Qwen tokenizer. Charts use the separately listed plotting dependency.
Source code, pinned prompts, generated conversations, raw measurements, training audits, plots, and machine-readable results are included. Weights belong in separate release assets, not Git history; see model cards and hashes. The source snapshot is approximately tens of megabytes. Each draft is about 799 MB and needs the pinned target model.
The headline is a within-study paired comparison, not a subtraction of medians across different sessions. Confidence intervals are exploratory and conditional on the tested models/prompts. Full methods, the failed concurrency check, data provenance, and limitations are in the write-up.
Completed provider usage: $17.27 before credits; conservative internal estimate: $41.88; cap: $60. These are different accounting measures. No active GPU containers remained at closeout; retained storage is separate. Cloud reproduction needs a new explicit budget and namespace. A clean GPU image rebuild has not been independently validated.