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

Model datasheet · Qwen3.6

Qwen3.6 35B-A3B

Near-flagship agentic coding at 3B-active speed: 73.4 on SWE-bench Verified while activating a twelfth of its weights. The trade is memory, not quality — you hold all 35B in RAM to run 3B worth of compute, and it carries the 3.6 family's tool-calling defect.

Vendor
Alibaba
Architecture
MoE · 35B total / ~3B active
Context
262,144 tokens
License
Apache-2.0
Released
2026-04-22
Vision
Yes

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

§1 Characteristics by quantization band

Q2–Q3 — 2–3 bit

17GB weights · e.g. Q3_K_M + KV: 160MB at 8K · 640MB at 32K · 2.5GB at 128K
Math & reasoning 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 MoE routing degrades faster than 3-bit dense — do not assume the dense 27B's unusually clean Q3_K_M result transfers here.

Coding 6/10
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.

Tool calling / agentic 4/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q3_K_M · aggregated · 2026-08-30

    Editorial derivation: below usable. The family's empty-tool-call defect plus 3-bit expert drift — see the weakness entry covering this band.

Q4–Q5 — 4–5 bit

22GB weights · e.g. Q4_K_M , mlx-4bit + KV: 160MB at 8K · 640MB at 32K · 2.5GB at 128K
Math & reasoning 7/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q4_K_M · aggregated · 2026-08-30

    Editorial derivation: one point down. Unlike the dense 27B — where a measured study found 4-bit indistinguishable from bf16 — sparse experts are individually more quant-sensitive, the same reasoning applied to gemma-4-26b-a4b. This is the band that fits a 32GB machine (22.29GB Q4_K_M).

Coding 7/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q4_K_M · aggregated · 2026-08-30

    Editorial derivation: one point down for MoE expert quantization.

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

    Editorial derivation: a shaky tool layer plus expert quantization.

Q6–Q8 — 6–8 bit

31GB weights · e.g. Q6_K + KV: 160MB at 8K · 640MB at 32K · 2.5GB at 128K
Math & reasoning 8/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q6_K · aggregated · 2026-08-30

    Editorial derivation: 6-bit carries MoE expert weights without meaningful loss. Note the file is 30.95GB — this band needs a 64GB machine in practice.

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.

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, defect included.

FP16 — full precision (fp16/bf16)

70GB weights · e.g. bf16 + KV: 160MB at 8K · 640MB at 32K · 2.5GB at 128K
Math & reasoning 8/10
data checked Aug 2026
  • vendor-model-cardAIME26 = 92.7 @ bf16 · vendor-reported · 2026-08-30

    GPQA: 86.0, MMLU-Pro: 85.2 — within a couple of points of the dense 27B sibling (94.1 / 87.8) at a twelfth of the active compute. Vendor-reported.

Coding 8/10
data checked Aug 2026
  • vendor-model-cardSWE-bench Verified = 73.4 @ bf16 · vendor-reported · 2026-08-30

    SWE-bench Pro: 49.5. Trails the dense Qwen3.6 27B (77.2 / 53.5) by roughly four points — same editorial band, and the gap buys a large speed advantage at 3B active.

Tool calling / agentic 6/10
data checked Aug 2026

§2 Known issues & what fixes them

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

Shares the 3.6 family defect: the model announces a tool call and then emits an empty one, ending the agent loop with the task unfinished. Reporters state explicitly that it affects the 35B as well as the 27B, and it reproduces on Alibaba's own FP8 build — so it is neither a size-specific nor a quantization artifact.

Fix
Update your runtime [partial]

Run a current llama.cpp/vLLM with the official 3.6 chat template rather than a 3.5 template carried over; this clears most of the empty calls.

Fix
Prompt technique [partial]

Retry on an empty tool call instead of terminating, and prefer few coarse tools over many fine-grained ones.

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

The A3B speed comes with a memory bill that a parameter count hides: all 35B of weights must be resident to activate 3B of them. Q4_K_M is 22.29GB against the dense 27B sibling's 17.1GB — so the faster model is the one that does NOT fit the smaller machine, and Q6_K (30.95GB) needs 64GB.

Fix
Move up a quant band [strong]

Inverted advice: on a 32GB machine the dense Qwen3.6 27B at Q4_K_M gives you higher benchmark scores in less memory. Choose this model for throughput on a large machine, not to save space.

Evidence · 1 source · anecdotal