Recursive Superintelligence released Flint on October 21, 2026: a mixture-of-experts coding model meant to run on local hardware. Official materials put total size near 40 billion parameters with about 2.5 billion active per token. The company says its Eureka Machine automated research system helped propose, implement, and test much of the architecture, quantization, and kernels in under four months.
Key Facts
| Details | |
|---|---|
| Model | Recursive Flint (Hugging Face: RecursiveSI/flint) |
| Released | October 21, 2026 |
| Size | About 39.8B total parameters; ~2.5B active per token (including output head), per the model card |
| License | Apache 2.0 |
| Context | Model card documents 262,144 tokens; company blog says up to 512k and recommends staying within 256k for best results |
| Architecture | 49-layer MoE: 2 dense layers then 47 MoE layers; 256 routed experts, top-6, plus a shared expert; mostly 513-token sliding-window attention |
| Local speed (company claims) | More than 125 tokens/s on M4 Pro MacBook Pro; more than 150 tokens/s on 64GB NVIDIA RTX Spark at short contexts |
| Company-reported coding scores (BF16) | SWE-bench Verified 74.2%; SWE-bench Pro v2 89.7%; Terminal-Bench 2.1 54.9%; LiveCodeBench v6 80.3% |
| Hardware notes | 4-bit weights about 23 GB; MarkTechPost and company materials cite 48 GB Apple Silicon or NVIDIA GPU for comfortable local use |
| Serving | llama.cpp, Ollama, Unsloth, LM Studio, vLLM, SGLang; company also released rec for Apple Silicon. vLLM published day-0 Docker/serve instructions |
Background
Recursive emerged from stealth in May 2026 with $650 million in funding at a $4.65 billion valuation, TechCrunch reported, with backers including GV, Nvidia, and AMD Ventures. CEO Richard Socher previously served as Salesforce's chief scientist.
Flint is framed as an early public test of AI-assisted model development rather than fully unsupervised self-improvement. Humans still oversee the Eureka Machine. TechCrunch and The Decoder note that Recursive acknowledges Flint may need more step-by-step guidance than large closed coding assistants on hard tasks.
vLLM's October 21 blog focuses on serving: NVFP4 and BF16 checkpoints on a single GPU, day-0 image vllm/vllm-openai:flint, and kernel work that brought batch-1 decode at 100K context to roughly 485 tokens/s on one GB200 in their end-to-end measurement. Those server figures are separate from Recursive's laptop speed claims.
Reactions and Impact
Wired's Will Knight, cited by TechCrunch, built small programs with Flint on a Mac with 48GB of memory and said it completed his coding requests. Independent third-party benchmark replications were still limited at writing time; MarkTechPost and The Decoder stress that several chart numbers are company-reported.
What We Don't Know Yet
- How much wall-clock time Eureka Machine saved versus a fully human research schedule (Recursive did not publish a figure).
- Independent reproductions of SWE-bench Pro v2 and related agentic scores outside Recursive's harness notes.
- Whether later Eureka-built models will measurably improve the Eureka Machine itself, which is the company's longer bet.
- Production API pricing, if any; the release emphasizes downloadable weights.