Qwen3.6-35B-A3B Stumbles on Coding Loops
This Reddit post is basically a real-world stress test of Qwen3.6-35B-A3B inside an agentic coding workflow. The user says the model is not “entirely stupid,” but it becomes hard to use because it loops on file creation and repeatedly hallucinates simple formatting mistakes like broken YAML syntax. That lines up with the broader pattern for open coding models: benchmark results can be solid, but actual usefulness depends heavily on quantization, prompting, tool-call handling, and how well the surrounding agent runtime constrains retries.
Hot take: the model is probably not “bad at coding” so much as “good enough to expose weak agent scaffolding.” When a model starts looping on empty file writes or self-correcting the same syntax error forever, that is often a tool-use/control failure amplified by quantization rather than a pure coding-capability verdict.
- –Official materials position Qwen3.6-35B-A3B as an open MoE coding model with 35B total parameters and 3B active per token, aimed at agentic coding and multimodal reasoning.
- –The benchmark picture is decent: Alibaba’s announcement says it beats Qwen3.5-27B on several coding benchmarks and scores 73.4 on SWE-bench Verified and 51.5 on Terminal-Bench 2.0.
- –The Reddit report suggests the weak point is execution stability, not basic competence: it can attempt the right task, then get stuck in repetitive tool calls or low-level syntax churn.
- –A Q5 quantization can easily worsen exactness, especially for tool-call formatting, YAML/JSON punctuation, and multi-step file-writing loops.
- –If you want it to feel usable, the surrounding agent matters a lot: stricter tool-call validation, lower retry counts, better stop conditions, and possibly a less aggressive quantization.
- –Bottom line: good for coding in the “can solve real tasks” sense, but not reliable enough to trust blindly in an unconstrained agent loop.
DISCOVERED
4h ago
2026-04-27
PUBLISHED
4h ago
2026-04-27
RELEVANCE
AUTHOR
laughingfingers