Birgitta Böckeler tests local models for agentic coding on Apple Silicon

A four-week, hands-on memo finds local coding models fast and runnable on 48–64GB Macs, but still unreliable at tool calling and far below big models.

Birgitta Böckeler spent roughly four weeks running local models for agentic coding and wrote up the factors that decide whether they're viable (martinfowler.com, July 7, 2026; resurfaced by @bibryam on October 4). If you're weighing a local setup against cloud APIs, her experience points to a usable speed story and a weak reliability story.

Her setup was an Apple M3 Max with 48GB RAM and an M5 Pro with 64GB. She used LM Studio, 4-bit quantized Qwen3.6 and Gemma 4 models, and the OpenCode and Pi harnesses with no Skills or MCP servers active. Her findings:

• RAM is the core constraint. Models of 15–25GB were comfortable. A 48GB model on the 64GB machine crashed partway into a conversation.
• Context needs 32K–64K. Default runtime settings are too small for agentic work, and the KV cache competes with the model weights for memory.
• Speed was good compared with a year ago, though it slows as conversations grow.
• Tool calling often failed, for example with wrong parameter names, though models could usually recover. Harnesses use different edit-tool schemas, which she suspects doesn't help.
• Reasoning can hurt. Smaller models looped in their reasoning, and turning reasoning off was faster and performed the same to slightly better in her automated runs.
• Outcomes were "hit and miss" and "nowhere near" big hosted models. She judged only functional correctness, not code quality.

She found Qwen3.6 35B-A3B MoE the best balance of size and quality. This is one developer's informal evaluation, not a benchmark. A follow-up memo with detailed task results is promised. The part of the memo we read doesn't compare costs with cloud usage.

Sources

• Viability of local models for coding · martinfowler.com
• @bibryam: Viability of local models for coding · @bibryam