Recursive's Flint hits 131 tok/s on M4 Pro: 40B MoE with 2B active tuned for local coding
_rockt · x · 2026-10-22
Recursive is showing off Flint, a 40B-parameter (2B active) model built for local coding. In official local-hardware decode benchmarks, Flint hits 131.22 tokens/s on an Apple M4 Pro 48GB (MLX) and 155.82 tokens/s on RTX Spark (NVFP4), with the best normalized AUC through 128K context among sparse MoE and dense peers like Llama, Laguna, Qwen3.5, and Gemma. The author touts it as "great at local coding, and fast."
Related event: Recursive Releases Flint, 40B MoE Open-Weight Coding Model(44 posts)→
More from Infra
- Micron AI memory roadmap: bandwidth moves onto the die, HBM supply tight from 2029 — BenBajarin · 2026-10-22
- Qualcomm's non-phone revenue seen growing 60% next year, driven by auto and embedded — pdamodaran · 2026-10-22
- After PHP advisories, a WebAssembly runtime vendor lays out its security validation plan — MickeySteamboat · 2026-10-22
- Hugging Face: Specializing an SLM's vocabulary can cut inference cost by 77% — Hugging Face Blog · 2026-10-22
- Ollama Runs Open Models on NVIDIA Blackwell via Crusoe, Live in Two Weeks — ollama · 2026-10-22
- CUDA's moat quietly eroding as Codex can rewrite kernels for AMD, investor notes — paigeinsf · 2026-10-22