REDDIT · REDDIT// 32d agoBENCHMARK RESULT

Qwen3-Coder-Next tops Qwen3.5 in Claude Code

A LocalLLaMA user reports that Qwen3-Coder-Next 80A3B and Qwen3.5 35B both ran at roughly 132K context inside 36GB of combined VRAM, but Coder-Next was far more dependable inside Claude Code. In this side-by-side local test, Qwen3.5 repeatedly stalled mid-job and needed workarounds, while Qwen3-Coder-Next handled tool calls cleanly and felt much closer to Sonnet-level reliability.

// ANALYSIS

Community evals like this are messy, but they matter because agentic coding lives or dies on tool-call stability, not just raw model size or benchmark bragging rights.

–The key result is reliability, not speed: the poster says Qwen3-Coder-Next stayed stable through Claude Code jobs while Qwen3.5 35B often stopped in the middle.
–Both models reportedly fit long context on a dual-GPU 36GB setup, which makes the stability gap more important than the raw 80B-versus-35B comparison.
–That fits Qwen’s broader positioning around coding and agentic workflows, where tool use and long-horizon execution matter more than one-shot code generation.
–It is still anecdotal and hardware-specific, but it is exactly the kind of field report local-first Claude Code users want before burning time on quant and template experiments.

// TAGS

qwen3-coder-nextclaude-codellmai-codingbenchmarkopen-weights

DISCOVERED

32d ago

2026-03-10

PUBLISHED

34d ago

2026-03-08

RELEVANCE

8/ 10

AUTHOR

ikaganacar