Recursive releases Flint, an open-weight 40B-A2.5B local coding model largely developed by its automated research system, the Eureka Machine
Confirmed
Importance: major (4 of 5)The takeaway
Recursive, Richard Socher’s lab, released Flint on Oct 21, 2026: an Apache-2.0 mixture-of-experts coding model (39.8B total, ~2.5B active) that runs on a laptop at 125–150 tokens/s.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- Major (4 of 5)
- Last verified
- 22 October 2026
Your AI and this story
- GPT-6 Astra174 days after its cutoff
- Claude Opus 5.5113 days after its cutoff
- Gemini 4 Argon204 days after its cutoff
- Grok 4.7143 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 113 days before it.
Key facts
- Architecture (model card): decoder-only MoE, 39.8B total parameters, ~2.5B active per token; 49 layers (2 dense + 47 MoE), 256 routed experts with 6 selected plus a shared expert; 35 sliding-window (513-token) and 14 full-attention layers; 262,144-token context; knowledge cutoff June 2026
- Benchmarks (Recursive): 54.9% Terminal-Bench 2.1, 89.7% SWE-bench Pro v2, 74.2% SWE-bench Verified; best SWE-bench Verified among models under 50B parameters per TechCrunch; only Qwen3.8 27B scored higher on Recursive’s six-test average
- Speed: ~131 tokens/s on an M4 Pro MacBook Pro (about 2x Gemma 4 and Nemotron 3.5 Lightning on the same machine), 150+ on RTX Spark
- Recursive blog: ‘Our small team of people and agents built it from scratch in under four months’; the Eureka Machine ‘helped develop nearly every part of Flint’s model-building stack, from data and training recipes to infrastructure and inference optimization’
- Eureka Machine contributions (per Recursive via TechCrunch): sparser model shape, low-bit training (71% smaller with almost no quality loss), custom kernels; Recursive did not say how much researcher time it saved
- License and access: Apache 2.0; Hugging Face RecursiveSI/flint (BF16, ~80 GB) plus int4 and NVFP4 GGUF variants; free on OpenCode and pi.dev
- Company: founded 2025, CEO Richard Socher (ex-Salesforce chief scientist); $650M at $4.65B valuation (May 2026; GV, Nvidia, AMD Ventures); $410M AWS compute deal (June)
- Reach: Socher’s launch post 2.8M views
What happened
Recursive, which emerged from stealth in May with the stated long-term goal of supervised recursive self-improvement, shipped its first model. Flint targets fast, private coding on local hardware and comes with open weights. Recursive credits the Eureka Machine, a system that proposes ideas, runs experiments and picks next steps, with most of the development work, under human supervision.
TechCrunch notes Recursive recommends breaking big tasks into smaller steps, and that the current system is a supervised research assistant, far from the open-ended self-improvement some researchers fear. A hands-on WIRED review is cited only through TechCrunch’s summary (not opened here).
Why it matters
AI-automated AI R&D has been a central concern in lab risk reports. Flint is a concrete, downloadable product whose makers say an AI system did most of the research, which makes the claim testable by outsiders.
Sources
4 sources from 4 sites. Numbers match the chips in the text.
4 sources: 3 primary, 1 press
Primary
- Recursive: Introducing Recursive Flintrecursive.com, official
- Hugging Face: RecursiveSI/flint model cardhuggingface.co, docs
- Richard Socher on X (2.8M views)x.com, official
Press
Changes
- Filed (X)