Recursive launches Flint, an open-weight 40B MoE coding model built by its automated research system

Recursive (Recursive SI), founded by Richard Socher, released its first open-weights model Flint on October 22 — a 40B-total-parameter MoE with 2.5B active parameters, purpose-built for local agentic coding. The weights are on Hugging Face under Apache 2.0 (an Unsloth GGUF version is already available), with day-0 support for Apple Silicon (in partnership with Nativ AI and mlx-vlm). CEO Josh Tobin said the company is betting that a large share of everyday programming will shift from the cloud to small models running on laptops or LAN servers.

Confirmed

• Architecture: a 40B-A2.5B coding MoE trained from scratch; per Robert Lange's architecture overview, it uses 256 experts (512 width), top-6 routing, and a 3:1 mix of sliding-window attention (SWA) and full attention; 262K context with support for reasoning and tool calling
• Speed: 125+ tokens/s decoding on a MacBook Pro M4 (the architecture overview cites 131 tok/s on an M4 Pro), and 150+ tokens/s on an NVIDIA DGX Spark; officially claimed to be up to 2x faster than comparable local models
• Benchmarks: 89.7 on SWE-Bench Pro v2; 54.9 on Terminal-Bench 2.1 (vs. 38.7 for Qwen3.5); officially claimed to lead same-class sparse MoEs across six benchmarks, beating Qwen3.5 and Muse Glimmer
• Quantization: 4-bit lmhead quantization loses only 0.7 points
• Development: Socher revealed Flint was built by his automated research system, the "Eureka Machine," which runs full research loops of ideation → implementation → validation and decides next steps based on results; it went from zero to release in under 4 months

Why it matters

• Flint is widely seen as a landmark case of AI accelerating AI research (the RSI path): a small team built a model in months with a fully automated research system, and Socher says the Eureka Machine was involved in nearly every step of Flint's development
• Reddit user returnity, after testing it, called it "the Qwen3.5 35B replacement everyone has been waiting for," suggesting Flint could challenge the Qwen small-powerhouse lineup for coding agents
• Fast local inference (125+ tokens/s on a laptop) plus open weights offers a new proof point for shifting cloud-based coding to local machines

2026-10-22 ~ 2026-10-22 · 73 related posts

• Flint lands on Hugging Face: 40B MoE with 2.5B active params for on-device coding — pcuenq · 2026-10-22
• [source] 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 Ships Flint: a 39.8B MoE Coding Model With 2.5B Active Params and Its Own Inference Engine — BerryBlush_Rogers691 · 2026-10-22

22 near-duplicate retellings: cong_ml · ChengleiSi · ajratner · dr_cintas · _rockt · _rockt · HeyAmit_ · ChengleiSi · ying11231 · jeffclune · CodeByPoonam · ChengleiSi · ChengleiSi · _rockt · yifengzhu_ut · noahdgoodman · ChengleiSi · damianplayer · ChengleiSi · _rockt · HeyAmit_ · rohanpaul_ai