64GB Macs are no longer a dead zone for agentic coding: 27B model at 256k context, measured
M5_Maxxx · reddit · 2026-11-01
After ten days running Qwen3 27B (16GB MLX quant, LM Studio) for agentic coding on a 64GB M4 Max, the author argues the old "64GB is a dead zone" advice has flipped.
- 256k context, agent sessions at 100-200k tokens; at 164k tokens the engine uses 48GB (46GB GPU): 16GB weights + 11GB KV cache (64KB/token, only 16/64 layers full attention) + a second context copy for prompt cache and linear attention checkpoints + MLX prefill scratch
- Usability hinges on the prompt cache: 93% hit rate means 1k new tokens per turn; a full prefill miss takes 20 minutes
- LM Studio's MLX vision path refuses KV quantization; the text-only path allows it but drops cache checkpoints, causing more full reprocessing on hybrid models
- 48GB doesn't fit unless dropping to 128k context or moving to llama.cpp with q8 KV (25-30GB at 256k, slower prefill); on 64GB it fits with only 150MB free
- Core argument: with agents, context costs as much as the model, and 27B is now good enough for real work — making 64GB the sweet spot; bigger MoE or dual models still want 128GB
More from coding & agent
- Claude API raises request limit from 32MB to 128MB, image size cap to 3000px — FinanceYF5 · 2026-11-01
- Indie hacker's viral video made with a Skill that auto-builds social-proof walls — yihui_indie · 2026-11-01
- Vibe Kanban's Shutdown Teaches a Lesson: Keep Agent Coordination MCP-Portable — onitsoga · 2026-11-01
- Kimi autonomously used a 4x3090 rig and YuE2 to score its own 45-second launch film — minchoi · 2026-11-01
- Kimi models a walkable 3D forest in Blender, rendered in-browser with Three.js — minchoi · 2026-11-01
- Kimi K3.1 generates a 3D turbo engine visual in about 40 minutes from one prompt — minchoi · 2026-11-01