Practitioner demonstrates a fully offline coding agent using open-weight Qwen3.5 9B
2026-10-27
A Latent Space piece published October 27 described running a coding agent (OpenCode, connected to a local Ollama instance) entirely offline using the open-weight Qwen3.5 9B model, for tasks like explaining code, investigating failing tests, refactoring, and finishing features without network access. Per Qwen's own model card as cited in the piece, Qwen3.5 9B scores 65.6 on LiveCodeBench v6 and 66.1 on BFCL-V4 (function calling), versus 55.8/50.3 for the smaller Qwen3.5 4B, while gpt-oss-20b scores 74.6 on LiveCodeBench v6 (BFCL-V4 not reported for it). The Qwen3.5 9B download is reported at roughly 6.6-7.6GB.
Significance 3: A credible practitioner demonstration of real offline agentic-coding capability with a newer open-weight model generation not yet in the tracker's current entities is relevant to the local-model verdicts, though the cited benchmarks are lab-reported rather than independently verified.
Implications · machine-drafted, not owner judgment
Qwen3.5 9B's published scores suggest Alibaba has iterated past the Qwen3 generation currently backing the local-fast verdict; if independently confirmed, this model is a candidate to re-examine for local-fast, particularly since it fits well within a 64GB hardware budget at under 8GB on disk. It is also a concrete demonstration that useful agentic coding workflows now run fully offline on consumer hardware.
- Independent (non-Qwen-card) benchmark confirmation of Qwen3.5 9B's coding and tool-use scores
- Whether larger Qwen3.5 variants also fit a 64GB local-hardware budget
Sources
- curator Coding From the Himalayas With an Offline AI Agent retrieved 2026-10-27
Affected axes
Coding · Agentic / tool use · Runs well locally on 64GB Apple Silicon