Model datasheet · Qwen3.6
Qwen3.6 27B
The 27B dense agentic-coding model that beat a 397B MoE on SWE-bench Verified — and, unlike its Qwen3.8 successor, has had four months for independent evaluators to check the claim. Q4_K_M is measurably indistinguishable from bf16 here; the real caveat is tool-calling reliability, not quantization.
- Vendor
- Alibaba
- Architecture
- Dense · 27B
- 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
data checked Aug 2026
- quant-degradation-community — AIME-120 (temperature 0.6) = 73.3 @ Q3_K_M · aggregated · 2026-08-30
Q3_K_M actually scored 73.3 vs bf16's 70.8 — within run-to-run noise, not a real gain. Discounted one point because the band is unreliable as a whole: Q3_K_S collapsed to 54.2 on the same test. See the quantOverride.
data checked Aug 2026
- quant-degradation-community — SVG-generation duels + perplexity vs bf16 @ Q3_K_M · aggregated · 2026-08-30
Editorial derivation: 3-bit still produced coherent structured output in the study's generation tests; one-point discount for band-internal variance.
data checked Aug 2026
- quant-degradation-community — top-1 token agreement vs bf16 @ Q3_K_M · aggregated · 2026-08-30
Editorial derivation: an already-shaky tool layer plus 3-bit token drift. Two points below the full band.
Q4–Q5 — 4–5 bit
17GB weights · e.g. Q4_K_M , Q5_K_M , mlx-4bit + KV: 512MB at 8K · 2GB at 32K · 8GB at 128Kdata checked Aug 2026
- quant-degradation-community — AIME-120 (temperature 0.6) = 70.8 @ Q4_K_M · aggregated · 2026-08-30
rawValue is the bf16 reference the study measured; Q4_K_M landed within noise of it. Editorial derivation: no discount for 4-bit on this model.
data checked Aug 2026
- quant-degradation-community — Terminal-Bench 2.1 @ Q4_K_M · aggregated · 2026-08-30
Editorial derivation: the 17.1GB Q4_K_M matched the full model on an agentic coding benchmark. This is the band to run — it fits a 32GB machine with context to spare.
data checked Aug 2026
- quant-degradation-community — top-1 token agreement vs bf16 @ Q4_K_M · aggregated · 2026-08-30
Editorial derivation: carried from the full band, defect included.
data checked Aug 2026
- quant-degradation-community — KL divergence + top-1 agreement vs bf16 @ Q6_K · aggregated · 2026-08-30
Editorial derivation: measured study of this exact model found every quant at 4-bit and above indistinguishable from full precision up to noise. Carried from the full band unchanged.
data checked Aug 2026
- quant-degradation-community — KL divergence + top-1 agreement vs bf16 @ Q6_K · aggregated · 2026-08-30
Editorial derivation: carried from the full band — no measurable loss at 6–8 bit.
data checked Aug 2026
- quant-degradation-community — KL divergence + top-1 agreement vs bf16 @ Q6_K · aggregated · 2026-08-30
Editorial derivation: carried from the full band. The empty-tool-call defect is architectural, not a quantization artifact — quantizing up does not fix it.
FP16 — full precision (fp16/bf16)
55GB weights · e.g. bf16 + KV: 512MB at 8K · 2GB at 32K · 8GB at 128Kdata checked Aug 2026
- vendor-model-card — GPQA Diamond = 87.8 @ bf16 · vendor-reported · 2026-08-30
AIME26: 94.1, HMMT Feb 25: 93.8, MMLU-Pro: 86.2. Vendor-reported, but the AIME line is the one number an independent party has re-run locally (see the mid-band cell).
data checked Aug 2026
- vendor-model-card — SWE-bench Verified = 77.2 @ bf16 · vendor-reported · 2026-08-30
LiveCodeBench v6: 83.9, SWE-bench Pro: 53.5, Terminal-Bench 2.0: 59.3. Beats the 397B-A17B Qwen3.5 MoE on most coding lines.
- vals-ai — Vals Index placement (link-only — no numbers republished) @ bf16 · aggregated · 2026-08-30
Independent evaluator: #22 on the Vals Index overall, #8 of 18 among open-weight models, four months after release. Corroborates the tier, not the exact vendor figures — held at 8 rather than 9 because the flagship agentic numbers are still vendor-measured.
data checked Aug 2026
- vendor-model-card — AndroidWorld = 70.3 @ bf16 · vendor-reported · 2026-08-30
Benchmarks are strong on paper.
- community-consensus — reported agent-loop failures (empty tool calls) @ bf16 · aggregated · 2026-08-30
Scored two points below the benchmark tier: the 3.6 family emits empty tool calls that terminate agent loops, reported across 27B and 35B and at official FP8. See weaknesses/qwen3.6-27b.json.
§2 Per-quant deviations
Q3_K_S · Math & reasoning drops to 4/10
Q3_K_S is the trap in this band — it loses ~16 points of AIME accuracy while the barely-larger Q3_K_M loses nothing. If you are squeezing this model into ≤3 bits, take Q3_K_M.
§3 Known issues & what fixes them
The model regularly announces that it is about to call a tool and then emits an empty tool call, which terminates the agent loop with the task unfinished. Reported across the whole 3.6 family (27B and 35B) and reproduced on Alibaba's own FP8 build, so it is not a quantization artifact — users upgrading from 3.5, which had no such problem, notice it immediately.
Largely a chat-template and harness problem: run a current llama.cpp/vLLM build with the official 3.6 template rather than a 3.5 template carried over. This clears most, not all, of the empty calls.
Have the harness retry on an empty tool call instead of ending the run, and prefer fewer, coarser tools over many fine-grained ones.
Evidence · 2 sources · community-consensus
- Qwen3.6-27B frequently stopped with empty tool call (github, 2026-08-30)
- Anyone is having issues with tool calling with the 3.6 family? (vendor, 2026-08-30)
The Q2–Q3 band is not uniform for this model. In a systematic quantization study, Q3_K_M held full-precision math accuracy (73.3 on AIME-120 vs bf16's 70.8) while Q3_K_S fell to 54.2 — a ~16-point drop for a file barely a gigabyte smaller. 2-bit quants degrade visibly in generation quality on top of that.
Within the same band, choose Q3_K_M over Q3_K_S — it costs almost nothing extra and recovers the whole gap. If you have 19GB free, Q4_K_M is indistinguishable from bf16 on this model.