Qwen3.8 27B

Roughly 88% at 2-bit against roughly 90% at full precision — a real but small loss, and far better than 2-bit usually manages. One point down.

The study reports a noticeable fall at 2-bit on agentic coding while remaining competitive. Two points down from the measured 4-bit result.

Editorial derivation: one point down, in line with the measured GPQA drop at the same width.

Editorial derivation: two points down. Structured tool output is the first thing to go at 2-bit, and the study's agentic scores fall there.

No discount. This site normally docks a point at 4-bit, but here it was measured: the 17GB Q4_K_M matched the 55GB full model on an agentic coding benchmark. This is the band to run — it fits a 32GB machine.

Measured at roughly 90% against the full model's roughly 90%. No discount applied because none was observed.

Held at 4-bit in the same study. Carried unchanged.

Carried unchanged — the agentic benchmark is the one that held exactly.

Measured, not derived: an independent study found Q4_K_M matching the full model on Terminal-Bench 2.1, so 8-bit sits between two points that are equal. Carried unchanged.

Same study: ~90% at both bf16 and Q4_K_M. Carried unchanged.

IFBench was one of the three benchmarks run across the quant ladder with no meaningful degradation at 4-bit or above.

Carried unchanged; agentic behaviour held at 4-bit in the same study.

Terminal-Bench 2.1: 73.0, SWE-bench Pro: 61.7, NL2Repo-Bench: 42.3. The highest figures in this catalogue on all three — Qwen3.6 27B manages 83.9 on LiveCodeBench and Ornith-1.5 35B-A3B 68.5 on Terminal-Bench.

Independent, and the reason this cell moved from 8 to 9: #9 overall on Code Arena's WebDev board and the only model in its size class in the top 10, plus the #1 open-weight slot on Arena.ai's Image-to-WebDev board. Third-party evaluation of the coding claim now exists; it did not on 2026-08-30.

Independent aggregator, cited by link under their terms. Places it immediately behind a 2.8T-parameter model on agentic work.

Humanity's Last Exam: 30.8. The GPQA figure was independently reproduced at ~90% in a third-party quantization study, which is the first outside confirmation of any Qwen3.8 number.

Above Nemotron 3 Nano's 71.5 on the same benchmark. Also reproduced independently in the quantization study below. IFBench is program-checkable, so there is little interpretive room.

AndroidWorld 81.9, WebArena-Verified 64.8 — none of them Qwen-built, unlike the QwenSWEBench and CoWorkBench numbers on the same card, which are excluded here for that reason.