Recursive, the self-improving AI startup led by former Salesforce chief scientist Richard Socher, has shipped its first model. Flint, released on October 21, 2026, is a coding model meant to run on a developer's own laptop or workstation, and its open weights are available on Hugging Face under the Apache 2.0 license. The company frames the release as a practical test of its Eureka Machine, an automated research system that it says contributed to nearly every stage of building the model.
Flint is a mixture-of-experts design with roughly 40 billion total parameters (39.8 billion in the model card) of which about 2.5 billion are active for each token. The card lists 49 layers, mostly expert layers, with 256 routed experts, six chosen per token, plus one shared expert that is always on. Recursive says it trained Flint from scratch through pretraining, mid-training, supervised fine-tuning and reinforcement learning, and that a small team of people and agents finished the job in under four months. The model handles text, reasoning and multi-turn tool calls, and the company recommends keeping prompts within 256,000 tokens.
The performance claims are Recursive's own. The company says Flint leads mixture-of-experts models under 50 billion total parameters on agentic coding tests, reporting 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified, and that it matches dense models of 25 to 40 billion parameters while running more than five times faster. TechCrunch notes that on Recursive's six-benchmark average only Alibaba's Qwen3.8 27B scored higher. The model card also warns that evaluation setups differ between developers, so the comparisons are not perfectly like for like.
Speed is the main pitch. With Recursive's tuned runtime, Flint produces more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 on a 64GB NVIDIA RTX Spark at short context lengths, which the company calls up to twice as fast as similar models. Alongside support for llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang, Recursive is releasing rec, a lightweight inference engine tuned for Apple's M4 Pro and M5 chips.
Recursive describes specific contributions from the Eureka Machine. The system searched expert layouts starting from a 35-billion, 3-billion-active baseline and landed on Flint's sparser shape. After training, it quantized the output layer to 4 bits, against common practice, and recovered accuracy with an offline distillation stage, leaving the NVFP4 build within 0.7 points of the 16-bit model on Recursive's benchmarks. It also wrote custom kernels, including a Metal megakernel that fuses the whole decoding pass on M4 and M5 chips.
Independent reporting adds context. Wired's Will Knight tested Flint on a Mac with 48GB of memory and found it fast and reliable on small programming projects, where he says models of this size often stumble. TechCrunch and Wired both report that Recursive emerged earlier this year with $650 million at a $4.65 billion valuation from backers including GV and Nvidia. Socher told Wired the longer-term goal is full recursive self-improvement under human oversight, an idea that Wired notes several congressional bills propose to ban.
The release still leaves open questions. Recursive acknowledges Flint needs more guidance than large hosted coding assistants and suggests splitting big tasks into smaller steps. Its account of how much the Eureka Machine contributed, and how much researcher time it saved, has not been independently audited, and the model card cautions that Flint can produce faulty code or inappropriate tool calls that users should review.