Flint hits 131 tokens/s on M4 Pro: open weights, no API fees, local-first coding
josh_tobin_ · x · 2026-10-22
Continuing the Flint launch thread, Josh Tobin frames the model as a step toward an everyday coding experience that is fast, customizable, private and free: open weights running on your own hardware with no API fees, decoding at over 125 tokens/s on an M4 Pro MacBook Pro at short contexts.
Detailed charts: Flint MLX hits 131.22 tokens/s short-context on M4 Pro (48GiB), llama.cpp 87.98, and NVFP4 on RTX Spark 155.82 — ahead of sparse MoE peers like Qwen3.6 35B (65.55) and Gemma 4 (61.89) on the quality-speed tradeoff.
Related event: Flint Benchmarks: Tops Sub-50B MoE, Runs at 131 tokens/s on M4 Pro(2 posts)→
More from Infra
- Agentic era needs its own infrastructure stack: security, privacy and governance — atShruti · 2026-10-22
- IBM Brings Spyre Accelerator to PyTorch Compiler and vLLM, Outlined at PyTorchCon 2026 — PyTorch · 2026-10-22
- Nscale listed 10,000MW of 'potential' capacity in its S-1 — built under 1% — SumitGup · 2026-10-22
- Six months later, the best AI costs 20× less — Tgoc · 2026-10-22
- Single RTX 5080 runs Qwen3.8-Flash-Next at 150 tok/s decode, 4.4K tok/s prompt — yehyakar · 2026-10-22
- SpaceX posts first operating profit as AI revenue nearly triples to $7.6B, passing Starlink — fortune · 2026-10-22