Issues

• Embedding can only load pure-java, and the pure-java path never receives a batched kernel

  integrallis/models#376 · Filter by author bsbodden opened on Nov 4, 2026
  1

• Per-call scratch allocation in Needle2 and DeBERTa forward passes

  integrallis/models#373 · Filter by author bsbodden opened on Nov 4, 2026

• Output-style switch: ASD-STE100 (Simplified Technical English) conformance

  integrallis/models#367 · Filter by author bsbodden opened on Nov 4, 2026
  1

• Embedding qualification records no throughput or latency: 31 artifacts with no speed figure

  integrallis/models#366 · Filter by author bsbodden opened on Nov 4, 2026

• Measure the SwiGLU shared-pool seam and settle the third compute pool

  integrallis/models#365 · Filter by author bsbodden opened on Nov 4, 2026

• Re-run the CUDA G1 token-parity gate on current code: the shipped evidence is 563 commits old

  integrallis/models#364 · Filter by author bsbodden opened on Nov 4, 2026

• Rerankers silently truncate pairs past their window: 13% of conversational-memory candidates are scored on a prefix

  integrallis/models#362 · Filter by author bsbodden opened on Nov 4, 2026

• Cross-encoder reranking is too slow to default: MiniLM-L6 Q4_K at ~0.8 s per candidate, slower than 70M F16 mxbai

  integrallis/models#360 · Filter by author bsbodden opened on Nov 4, 2026

• Our own reranker and embedder: distill and compress from the strongest permissive checkpoints

  integrallis/models#356 · Filter by author bsbodden opened on Nov 3, 2026

• Target models for local agent-memory defaults: extraction, admission gate (incl. System-1), reranking, embedding, judge — and the runtime work each needs

  integrallis/models#355 · Filter by author bsbodden opened on Nov 3, 2026
  1

• Reranking qualification measures logit parity and latency, not ranking quality

  enhancementimportant
  integrallis/models#353 · Filter by author bsbodden opened on Nov 3, 2026

• Embedding qualification measures oracle parity, not retrieval quality

  enhancementimportant
  integrallis/models#352 · Filter by author bsbodden opened on Nov 3, 2026

• Thread scaling: we gain 0.4% from 8 to 16 threads where the comparator gains 10.7%

  enhancementimportant
  integrallis/models#351 · Filter by author bsbodden opened on Nov 3, 2026

• mellum: prefill at 0.53x of ollama leaves the published entry no end-to-end latency headroom

  integrallis/models#338 · Filter by author bsbodden opened on Nov 2, 2026
  2

• MoE decode on the Llama path: mellum now passes (0.896); maple 0.21x and lfm2moe 0.47x still short

  enhancement
  integrallis/models#322 · Filter by author bsbodden opened on Oct 31, 2026
  7

• Port parakeet: needs a transcription model class we do not have

  enhancement
  integrallis/models#284 · Filter by author bsbodden opened on Oct 24, 2026
  2

• Port lfm2moe: best value on the list after Tier A

  enhancement
  integrallis/models#283 · Filter by author bsbodden opened on Oct 24, 2026
  3

• qxotic architecture parity: 10 ports, what each needs, and where it stops

  enhancement
  integrallis/models#279 · Filter by author bsbodden opened on Oct 24, 2026
  8