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

Model datasheet · Ornith 1.5

Ornith-1.5 9B

70.6 on SWE-bench Verified from a 9B dense model that fits a 16GB machine at 4-bit — a number that would have belonged to a 30B a year ago. Text only, MIT, and only its maker has measured it so far.

Vendor
DeepReinforce
Architecture
Dense · 9B
Context
262,144 tokens
License
MIT
Released
2026-08-19
Vision
No

→ GGUF (llama.cpp / LM Studio / Ollama)

§1 Characteristics by quantization band

Q4–Q5 — 4–5 bit

5.8GB weights · e.g. Q4_K_M , mlx-4bit + KV: 256MB at 8K · 1GB at 32K · 4GB at 128K
Coding 6/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q4_K_M · aggregated · 2026-08-30

    Editorial derivation: one point down. A 9B has less redundancy to spare than a 30B, so 4-bit costs it more.

Tool calling / agentic 5/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q4_K_M · aggregated · 2026-08-30

    Editorial derivation: one point down; structured tool output degrades before prose does.

Q6–Q8 — 6–8 bit

7.6GB weights · e.g. Q6_K + KV: 256MB at 8K · 1GB at 32K · 4GB at 128K
Coding 7/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q6_K · aggregated · 2026-08-30

    Editorial derivation: carried from the full band. At 7.6GB this fits a 16GB machine comfortably.

Tool calling / agentic 6/10
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)

18.4GB weights · e.g. bf16 + KV: 256MB at 8K · 1GB at 32K · 4GB at 128K
Coding 7/10
data checked Aug 2026
  • vendor-model-card — SWE-bench Verified = 70.6 @ bf16 · vendor-reported · 2026-08-30

    Extraordinary for 9B — above Granite 4.2 30B (57.0) and close to models three times its size. Exactly the kind of claim that needs independent replication, and has none yet.

Tool calling / agentic 6/10
data checked Aug 2026
  • vendor-model-card — Terminal-Bench 2.1 = 47 @ bf16 · vendor-reported · 2026-08-30

    Well behind the 35B sibling's 68.5 — multi-step agentic work is where the size difference shows.

§2 Known issues & what fixes them

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

70.6 on SWE-bench Verified from a 9B model would be remarkable if independently confirmed. It is not: the number is DeepReinforce's own five-run average, published eleven days ago, on a model trained against tasks it generated for itself.

No fix
No known fix — choose a different model

No fix — a caveat. If you need a coding model you can trust at this size today, the scores on this site with independent backing are lower and older.

Evidence · 2 sources · community-consensus
ISSUE-02 mild Q4–Q5 · coding

No Q2–Q3 build is published for this size, so the smallest practical download is the 5.8GB Q4_K_M. That is fine on 16GB and awkward on 8GB, where the 4-bit weights plus a working context leave very little room.

No fix
No known fix — choose a different model

On an 8GB machine, pick a smaller model rather than a smaller quant of this one — the quant does not exist.

Evidence · 1 source · anecdotal