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

Model datasheet · KAT-Coder

KAT-Coder-V2.5-Dev

Qwen3.6-35B-A3B with a coding-specific post-train bolted on — same architecture, same footprint, tuned hard for agentic software work and reported to use far fewer tokens getting there. Vision was stripped from the open release. A useful demonstration of what post-training alone is worth.

Vendor
Kwaipilot
Architecture
MoE · 35B total / ~3B active
Context
262,144 tokens
License
Apache-2.0
Released
2026-07-23
Vision
No

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

§1 Characteristics by quantization band

Q2–Q3 — 2–3 bit

16.2GB weights · e.g. Q3_K_M + KV: 160MB at 8K · 640MB at 32K · 2.5GB 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 dense weights at the same bit width.

Q4–Q5 — 4–5 bit

21.4GB weights · e.g. Q4_K_M , mlx-4bit + KV: 160MB at 8K · 640MB at 32K · 2.5GB 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 on this site. 21.4GB — the band that fits a 32GB machine.

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 from an already modest full-band score.

Q6–Q8 — 6–8 bit

30.1GB weights · e.g. Q6_K + KV: 160MB at 8K · 640MB at 32K · 2.5GB 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 is near-lossless even on sparse experts. 30.1GB needs a 64GB machine.

FP16 — full precision (fp16/bf16)

69.4GB weights · e.g. bf16 + KV: 160MB at 8K · 640MB at 32K · 2.5GB at 128K
Coding 8/10
data checked Aug 2026
  • vendor-model-card — SWE-bench Verified = 69.4 @ bf16 · vendor-reported · 2026-08-30

    SWE-bench Multilingual 63.0, SWE-bench Pro 45.96. Kwaipilot report their own base model, Qwen3.6-35B-A3B, at 64.40 on the same harness — where Alibaba report 73.4 for it. The two vendors' harnesses disagree by nine points on an identical model, so compare the uplift within this row, not across vendors.

Tool calling / agentic 6/10
data checked Aug 2026
  • vendor-model-card — Terminal-Bench 2.1 = 41.02 @ bf16 · vendor-reported · 2026-08-30

    Well behind Ornith-1.5 35B-A3B (68.5) at the same size and active-parameter count — this model is tuned for patch generation more than for long terminal sessions.

§2 Known issues & what fixes them

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

Kwaipilot measure their own base model — Qwen3.6-35B-A3B — at 64.40 on SWE-bench Verified, while Alibaba publish 73.4 for the same weights. Nine points of disagreement on an identical model tells you the harness, scaffold and sampling settings move the score as much as the model does. KAT-Coder's own 69.4 is measured on Kwaipilot's harness, so it can be compared to their 64.40 base figure but not to numbers from anyone else's.

No fix
No known fix — choose a different model

No fix — read within-vendor deltas, not cross-vendor rankings. This is why the site records the exact metric string and source for every number.

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

The base model, Qwen3.6-35B-A3B, is multimodal. The open-weight KAT-Coder release ships language weights only — the vision components were removed. If you picked it expecting the base model's screenshot handling, it is not there.

No fix
No known fix — choose a different model

Use the Qwen3.6 35B-A3B base if you need vision at this size; it is Apache-2.0 too.

Evidence · 1 source · anecdotal