Known-issues sheets for local language models Rev 0.1.0 · reviewed every 7 days

Model datasheet · Ornith 1.5

Ornith-1.5 35B-A3B

The highest SWE-bench Verified score in this catalog — 79.0, from 3B active parameters, under MIT. Trained by having the model invent its own tasks and scaffolds rather than learning from a fixed human-written set. Eleven days old, and every number is DeepReinforce's own.

Vendor
DeepReinforce
Architecture
MoE · 35B total / ~3B active
Context
262,144 tokens
License
MIT
Released
2026-08-19
Vision
Yes

→ GGUF (llama.cpp / LM Studio / Ollama) → MLX (Apple Silicon)

§1 Characteristics by quantization band

Q2–Q3 — 2–3 bit

16.7GB weights · e.g. Q3_K_M + KV: 128MB at 8K · 512MB at 32K · 2GB at 128K
Coding 6/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q3_K_M · aggregated · 2026-08-30

    Editorial derivation: two points down. 3-bit expert routing degrades faster than 3-bit dense weights.

Tool calling / agentic 5/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q3_K_M · aggregated · 2026-08-30

    Editorial derivation: two points down from the full band.

Q4–Q5 — 4–5 bit

21.9GB weights · e.g. Q4_K_M , mlx-4bit + KV: 128MB at 8K · 512MB at 32K · 2GB at 128K
Coding 7/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q4_K_M · aggregated · 2026-08-30

    Editorial derivation: one point down, the standard MoE deduction used across this site — sparse experts are individually more quant-sensitive than dense weights. This is the band that fits a 32GB machine.

Tool calling / agentic 6/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q4_K_M · aggregated · 2026-08-30

    Editorial derivation: one point down for expert quantization; long agentic runs compound small errors.

Q6–Q8 — 6–8 bit

30.5GB weights · e.g. Q6_K + KV: 128MB at 8K · 512MB at 32K · 2GB at 128K
Coding 8/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q6_K · aggregated · 2026-08-30

    Editorial derivation: carried from the full band — 6-bit holds MoE expert weights without meaningful loss. At 30.5GB this band needs a 64GB machine.

Tool calling / agentic 7/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q6_K · aggregated · 2026-08-30

    Editorial derivation: carried from the full band.

FP16 — full precision (fp16/bf16)

71.1GB weights · e.g. bf16 + KV: 128MB at 8K · 512MB at 32K · 2GB at 128K
Coding 8/10
data checked Aug 2026
  • vendor-model-cardSWE-bench Verified = 79 @ bf16 · vendor-reported · 2026-08-30

    The highest figure in this catalog, ahead of Qwen3.6 27B's 77.2 — from 3B active parameters. Held at 8 rather than 9 because it is DeepReinforce's own run, averaged over five attempts with their harness; nobody else has reproduced it in the eleven days since release.

Tool calling / agentic 7/10
data checked Aug 2026
  • vendor-model-cardTerminal-Bench 2.1 = 68.5 @ bf16 · vendor-reported · 2026-08-30

    Agentic terminal work over many steps. The model is trained by building its own scaffolds, so agentic behaviour is the design target rather than a side effect. Vendor-measured.

§2 Known issues & what fixes them

ISSUE-01 moderate Q2–Q3 / Q4–Q5 / Q6–Q8 / FP16 · coding, tool-calling

Every figure is DeepReinforce's own, averaged over five attempts with their harness and sampling settings. The headline — 79.0 on SWE-bench Verified, the highest in this catalog — is also the one most in need of a second opinion, and the model is eleven days old. The training method compounds the concern: the model generates its own tasks and scaffolds, so the possibility of the evaluation resembling the training distribution is real rather than theoretical.

No fix
No known fix — choose a different model

Nothing to fix — a caveat to hold. Treat the ranking as provisional until an independent harness reproduces it.

Evidence · 2 sources · community-consensus
ISSUE-02 mild Q2–Q3 / Q4–Q5 / Q6–Q8 / FP16 · coding

3B active parameters buys speed, not memory. All 35B of weights stay resident: 21.9GB at Q4_K_M, 30.5GB at Q6_K. On a 32GB machine only the 4-bit band is realistic, and the 6-bit band needs 64GB.

Fix
Move up a quant band [strong]

If you have the memory, Q6_K is worth it on a sparse model — expert weights suffer more from aggressive quantization than dense ones do.

Evidence · 1 source · anecdotal