5 August 2026 · Tel Aviv
What is actually on the bench.
A hand-curated record of the current focus and public open loops. This is intentionally dated; it is a working state, not a claim of permanent priority.
Make the MTP drafter smaller than the overhead it is meant to remove.
The primary study is StudentSV: a half-width MTP block with its own 32,768-token output head, trained for the real speculative-decoding path instead of inheriting the target model's full 248,320-row projection.
The study is concluded, and the answer is no—for a functional reason, not a size one. The verifier-as-oracle audit sorted every compression into two families: fidelity cuts (precision) cost at most 1.9 acceptance-retention points even under severe shift, while function cuts—pruning, low-rank, and the distilled StudentSV itself—keep 97% on calibration-register text and pay 10–19 points where the text actually shifts. Parameters serving rare content accumulate no mass on any finite calibration sample. The recipe that survives is zero-training: trim the vocabulary to the model's own 32k hottest rows, quantize to NVFP4, verify full-vocabulary—1.82× at 9B, 2.88× composed at 27B, measured end-to-end in memra, distribution-exact by construction. Harness at avifenesh/hqmtp; the paper draft is written and in internal review rounds.
- Subject
- Qwen3.5-9B · co-trained MTP → StudentSV
- Methods
- Small-vocabulary head · half-width block · chain distillation
- Standard
- Separate training signal, acceptance, FLOPs, and measured wall time.
- Working paper
- Study concluded · harness public ↗
The studies and systems that keep touching each other.
The model experiments need a runtime, and the runtime keeps generating better research questions. These are the nearest active loops.
The Rust + CUDA inference engine that exposes the real costs of MTP drafting, verification, quantization, expert spill, and Blackwell kernel choices. Renamed from bw24; v0.69 ships prebuilt binaries and serves official FP8 checkpoints with out-of-the-box MTP speculation at 2.6–2.8× plain decode.
public · v0.69 B / training recipe-labLayer-loop weight sharing plus ε=λ/(N√L) residual scaling, tested from zero in eleven pre-registered rounds. Data-constrained at d=768 the looped model beat FLOPs-matched vanilla in all three mixer families—attention 3/3 seeds, pure SSM 2/2, hybrid 2/2; seven paired runs, zero sign flips. The hybrid rounds isolated the rule "loop the state-mixer, never the retriever". The d≥1024 rung is next.
public · scaling rung next C / retention mem-retrofitProduct-key memory grafted post-hoc onto dense Qwen3-4B, five sequential update rounds against LoRA. The retrofit is free at lr/10; the published forgetting advantage never cleared a confidence interval in 12 of 12 tries. A powered ~30-seed run is what can change the verdict.
public · concluded at n=3 D / compression REAP × quantFirst physical artifact public on Hugging Face: Hy3-REAP-Layer103p5, a receipt-bound expert overlay selected on routing displacement. The public REAP50 mask was rejected at the screening gate (0.24 domain-macro vs the unpruned 0.73); layer-balanced pruning with private calibration holds 83.5% retention at the 100GB cap. Measured per-projection damage across seven formats now feeds one global allocator that chooses precision and pruning together under the byte ceiling.
working paper 02 E / review revutoA local PR reviewer that learns from maintainer feedback, with the model supplier and repository kept replaceable.
updated 04 Aug F / maintain Valkey GLIDEReview, API and reliability work across a large multi-language client, plus the support and release details that never appear in feature lists.
maintainerFinished work, waiting on maintainers.
Authored changes that are complete and sitting in review at the projects they target—not an activity feed.
Issues that are really research prompts.
The issue tracker is often where a vague irritation becomes a falsifiable question.