Recursive Releases Flint, 40B MoE Open-Weight Coding Model
Recursive, founded by Richard Socher, has released Flint, its first open-weights model, designed for fast agentic coding running locally on laptops. The entire model was built from scratch in under 4 months by Eureka Machine, an automated AI research system—widely seen as a real-world test of "AI doing research on its own."
Confirmed
- Flint was released by Recursive (founded by Richard Socher, who also founded You.com) as its first open-weights model, positioned for fast local coding (agentic coding).
- Architecture and specs: MoE design with 40B total parameters and only 2.5B activated per token, developed by a small team.
- Speed: officially claims over 125 tokens/s generation on a MacBook Pro; also noted to run locally on devices like the RTX Spark.
- How it was built: Socher revealed Flint is the real-world output of the "Eureka Machine"—a fully automated AI research system that runs complete research loops of "ideation → implementation → validation" and decides next steps based on results. The system was used throughout nearly every part of Flint's development, enabling a very small team to build the model in under 4 months.
Why it matters
- Flint demonstrates the feasibility of automated research systems producing usable models end to end: with AI automating the research loop, a small team dramatically cut the time and manpower needed for model development.
- The 40B-total / 2.5B-active MoE design plus open weights lets a high-quality coding model run fast locally on consumer laptops, lowering the barrier to local agentic coding.
2026-10-22 ~ 2026-10-22 · 44 related posts
Primary sources
- Recursive launches Flint: 40B MoE local coding model built in 4 months by its automated research system — RichardSocher ·
- Socher: Flint was built in <4 months by an automated AI research loop, shrunk 71% via low-bit training — RichardSocher ·
- Richard Socher's Recursive debuts Flint, an AI-built local coding model hitting 125 tokens/sec on a MacBook — dr_cintas ·
- Flint benchmark details: 89.7 on SWE-Bench Pro v2, ahead of Qwen3.5 and Muse Glimmer — RichardSocher · 2026-10-22
- [source] Socher: Flint was built in <4 months by an automated AI research loop, shrunk 71% via low-bit training — RichardSocher · 2026-10-22
- [source] Recursive launches Flint: 40B MoE local coding model built in 4 months by its automated research system — RichardSocher · 2026-10-22
- Recursive launches Flint: an open 40B MoE model built for local agentic coding — josh_tobin_ · 2026-10-22
- Flint hits 131 tokens/s on M4 Pro: open weights, no API fees, local-first coding — josh_tobin_ · 2026-10-22
- Ex-OpenAI's Josh Tobin ships Flint: sub-50B MoE leads agentic coding, 125 tok/s on M4 Pro — josh_tobin_ · 2026-10-22
- Flint was built from scratch in under 4 months, leads sub-50B MoE on coding benchmarks — josh_tobin_ · 2026-10-22
- Josh Tobin launches Flint, a local agentic coding model running on Ollama and llama.cpp — josh_tobin_ · 2026-10-22
- Recursive's Flint hits 131 tok/s on M4 Pro: 40B MoE with 2B active tuned for local coding — _rockt · 2026-10-22
- Recursive ships Flint: 40B MoE with 2.5B active params for laptop coding at 125 tok/s — _rockt · 2026-10-22
- Recursive SI launches Flint: open-weights local coding model at 125+ tokens/sec on MacBook Pro — JafarNajafov · 2026-10-22
- You.com launches Flint, a local coding model built by its Eureka Machine in under 4 months — omarsar0 · 2026-10-22
- vLLM cuts Flint's kernels from 22 to 7 per MoE layer with custom fusion — vllm_project · 2026-10-22
- vLLM Cuts Flint Kernel Launches from 22 to 7 per Layer with Recursive SI — vllm_project · 2026-10-22
- Richard Socher unveils Recursive's first model Flint, built in under 4 months — daniel_mac8 · 2026-10-22
- Flint: 40B MoE with 2.5B active params hits 125+ tok/s on M4 Pro — cong_ml · 2026-10-22
- Recursive open-sources Flint under Apache 2.0, runnable locally via Ollama — cong_ml · 2026-10-22
- Recursive.SI's Eureka Machine builds Flint-40B-A2B from scratch in 4 months, fast local coding — ChengleiSi · 2026-10-22
- Recursive's Eureka Machine builds Flint, a 40B open-weight local coding model, in under 4 months — CodeByPoonam · 2026-10-22
- Recursive (Caiming Xiong) releases Flint, an open-weights local coding agent model — ChengleiSi · 2026-10-22
- Recursive team on shipping Flint: full pre-training to RL pipeline built from scratch — ChengleiSi · 2026-10-22
- Recursive open-sources Flint, a 39.8B MoE coding-agent model with 2.5B active params — ollama · 2026-10-22
- Recursive's Flint: 40B MoE local coding model, 125+ tok/s on M4 Pro, built in 4 months by automation — josh_tobin_ · 2026-10-22
- Recursive's Flint: 39.8B MoE local coding model hits 125+ tok/s on a MacBook Pro — ivan_bezdomny · 2026-10-22
- Recursive launches Flint, a local agentic coding model, built in 4 months from scratch — josh_tobin_ · 2026-10-22
- Recursive bets automated AI research labs will outrun headcount-driven labs — ChengleiSi · 2026-10-22
- Recursive launches Flint: local coding model built in 4 months, 125+ tok/s on MacBook — ChengleiSi · 2026-10-22
- Recursive launches Flint, a local coding model hitting 125+ tokens/sec on MacBook, built in under 4 months — damianplayer · 2026-10-22
16 near-duplicate retellings: cong_ml · ChengleiSi · ajratner · dr_cintas · _rockt · _rockt · HeyAmit_ · ChengleiSi · ying11231 · jeffclune · CodeByPoonam · ChengleiSi · ChengleiSi · _rockt · yifengzhu_ut · noahdgoodman