Flint: 40B MoE with 2.5B active params hits 125+ tok/s on M4 Pro

cong_ml · x · 2026-10-22

Recursive releases Flint, a mixture-of-experts model with 40B total parameters and just 2.5B active per token, leading their <50B MoE agentic-coding comparisons. Measured at short context: 125+ tok/s on M4 Pro, 150+ on a 64GB RTX Spark, with recommended context up to 256k.

During development, their Eureka Machine searched expert layouts and widths for the target hardware, refining 35B/3B-active into 40B/2.5B — more total capacity, less work per token, faster local decoding.

Related event: Recursive releases Flint, an open-weights 40B MoE coding model built in under 4 months by its automated Eureka research system(50 posts)→

Original post →

More from Models

• Recursive AI ships Flint, a 40B-A2.5B MoE claimed to match Qwen3.6-27B, weights on Hugging Face — ChengleiSi · 2026-10-22
• DeepSeek ships V4.1-Flash-1017, now the #1 open model on RSI-Inference Bench — scaling01 · 2026-10-22
• OpenAI's Cybersecurity Guardrails Block a Dev From Auditing His Own Patched Server — Next_Technology6361 · 2026-10-22
• TPN-006 hits Hugging Face: 7.7GB compressed model beats its 17GB Mistral teacher — markjeffrey · 2026-10-22
• Kimi K3.1's latest checkpoint draws 49k-token SVG portrait zero-shot in 22 minutes — lyraxana · 2026-10-22
• MiniMax launches M3.1 Flash at 1/3 the cost and 3x the speed of M3, weights out Oct 23 — MiniMax_AI · 2026-10-22

Models channel →