Recursive launches Flint, an open-weight 40B coding model built by AI
Startup Recursive Superintelligence has released Flint, a free, open-weight coding model designed to run locally on laptops after being developed in under four months largely by its Eureka Machine AI system.
Startup Recursive Superintelligence has launched Flint, an open-weight coding model designed to operate locally on compatible laptops rather than in the cloud. Developed from scratch in under four months with oversight from human researchers, the model was largely generated by Recursive's automated research system, the Eureka Machine, as an early demonstration of artificial intelligence building subsequent AI systems.[1][2][3][4][5]
Released under an Apache 2.0 license and hosted on Hugging Face, Flint is a mixture-of-experts model comprising 39.8 billion total parameters, with roughly 2.5 billion active per token. Recursive claims the model leads its class of sub-50-billion-parameter mixture-of-experts models on agentic coding tasks, pointing to vendor-reported benchmarks of 89.7% on SWE-bench Pro v2, 74.2% on SWE-bench Verified, and 54.9% on Terminal-Bench 2.1 for its BF16 checkpoint.[1][4]
The release is aimed at reducing developer reliance on paid cloud platforms while improving data privacy. Early community performance reports show output speeds of roughly 131 tokens per second on an Apple M4 Pro and nearly 200 tokens per second on M3 and M5 Ultra hardware via the MLX-VLM dev branch. While an 18.5GB 3-bit GGUF quantization fits within 24GB of RAM, testers noted that execution speeds slow considerably past 64,000 tokens of context, and a stricter schema-checked agent test resulted in a 62.5% success rate.[1][3][4]
Key facts
- Recursive Superintelligence released Flint on October 21, 2026, as an open-weight mixture-of-experts coding model under the Apache 2.0 license.
- Flint has 39.8 billion total parameters with approximately 2.5 billion active per token, designed to run locally on laptops.
- The model was built in under four months largely by Recursive's automated AI system, the Eureka Machine, under human oversight.
- Recursive reported vendor-run BF16 benchmark scores of 89.7% on SWE-bench Pro v2, 74.2% on SWE-bench Verified, and 54.9% on Terminal-Bench 2.1.
- Flint's 3-bit GGUF version measures 18.5GB and fits in 24GB of RAM, with reported speeds of roughly 131 tokens per second on an Apple M4 Pro.
- Testing revealed 13 out of 13 long-context recall up to 125,000 tokens with sharp speed declines beyond 64,000 tokens, alongside a 5-out-of-8 score on schema-checked agent tool calls.
Sources · 5 sources
- AG
AGTP@AGTPinsightsPost on X ·
Recursive just released Flint, an open-weight 40B coding model built to run on a laptop. Here's what you need to know. Flint is a mixture-of-experts model from Recursive Superintelligence, released October 21, 2026. It has 39.8B total parameters but only about 2.5B are active per token. A small team of people and agents built it from scratch in under four months, using the company's automated research system, the Eureka Machine. It's licensed Apache 2.0. Recursive says Flint is the best model in its class, meaning MoE models under 50B parameters, on agentic coding benchmarks. The model card reports 89.7% on SWE-bench Pro v2, 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1. Those are vendor-run scores for the BF16 checkpoint. Speed is the pitch. Recursive says Flint is up to 2x faster than comparable models. Users report about 131 tok/s on an M4 Pro, and Prince Canuma saw nearly 200 tok/s on M3 and M5 Ultra using the MLX-VLM dev branch. On a home RTX 4090 with 24GB, one user got 50 tok/s with the nvfp4 version and some layers offloaded to CPU, even after a 200k token prefill. There are caveats. The full 40B model is closer to 48GB systems than common 24GB GPUs. The 3-bit GGUF is 18.5GB, so it fits in 24GB RAM. One tester found 13/13 long-context recall up to 125k tokens, but speed drops sharply past 64k. A stricter agent test with schema-checked tool calls put Flint int4 at 5/8, or 62.5%. Thomas Hatch says Qwen3.8-27B still wins on quality, but Flint has the best quality per compute he's used. Key numbers: - 39.8B total parameters, 2.5B active per token - 125+ tok/s on a MacBook Pro, 150+ on RTX Spark - 89.7% SWE-bench Pro v2, 74.2% SWE-bench Verified - 54.9% Terminal-Bench 2.1 - 3-bit GGUF is 18.5GB - Built in under 4 months, Apache 2.0 Flint is on Hugging Face now as RecursiveSI/flint, with the weights open for local use.
Open source - TE
TechCrunch@TechCrunchPost on X ·
Startup Recursive has released Flint, a free, laptop-sized coding model built largely by its own AI research system, as an early test of AI that improves itself. https://t.co/Q9PYfle0B6
Open source - CB
Crypto BriefingArticle ·
Recursive launches Flint, a fast open-weight coding model built to run locally Flint's local, open-weight model could democratize AI coding tools, reducing reliance on cloud services and enhancing data privacy for developers. The post Recursive launches Flint, a fast open-weight coding model built to run locally appeared first on Crypto Briefing .
Open source - AI
Analytics Insight@analyticsinmePost on X ·
𝗥𝗲𝗰𝘂𝗿𝘀𝗶𝘃𝗲 𝗙𝗹𝗶𝗻𝘁 𝗔𝗜: 𝗙𝗿𝗲𝗲 𝟰𝟬 𝗕𝗶𝗹𝗹𝗶𝗼𝗻 𝗖𝗼𝗱𝗶𝗻𝗴 𝗠𝗼𝗱𝗲𝗹 𝗥𝘂𝗻𝘀 𝗢𝗳𝗳𝗹𝗶𝗻𝗲 Can AI build the next generation of AI? Recursive is putting that idea to the test. Its new Flint coding model runs locally on compatible laptops, keeps code on-device and avoids cloud usage fees. Recursive says its Eureka Machine helped develop the model in under four months, with human researchers overseeing the process. #Recursive #FlintAI #AICoding #EurekaMachine #OpenSourceAI #ArtificialIntelligence #TechNews #AnalyticsInsight #AnalyticsInsightMagazine Read More 👇 https://t.co/SPaYvqF5Br
Open source - TE
TechmemeArticle ·
Recursive releases Flint, an open-weight, 40B-parameter coding model developed with its Eureka Machine AI system, which automated most of the development (Will Knight/Wired) Will Knight / Wired : Recursive releases Flint, an open-weight, 40B-parameter coding model developed with its Eureka Machine AI system, which automated most of the development — Recursive makes the case that the benefits of using AI to build AI outweigh the risks. — For the past few days I've been testing …
Open source