Recursive's First Model

Recursive — an AI startup founded last year by former Salesforce chief scientist Richard Socher — on October 21 introduced its first model, Flint. This is the company's first open model. Flint is a language model with a mixture-of-experts (MoE) architecture for coding agents: it has 39.8 billion total parameters, with about 2.5 billion active parameters per token. The model was released under the Apache 2.0 license, and its weights are openly available for free download on Hugging Face.

According to the company, Flint is designed for tasks such as exploring codebases, fixing bugs, writing new features, and creating tests. The model supports text in/out, reasoning, and multi-step tool calls. The most notable point: Flint is built to run on a laptop — in the company's tests it ran nearly twice as fast as comparable local models.

The Model Was Built by AI

Recursive emphasizes that Flint was largely created with the help of the company's automated research system, Eureka Machine. A small team built the model from scratch — in under four months. This is the second layer of the story: it's not just about the model itself, but how it was built — the automated research system took over much of the architecture-selection, training, and evaluation cycles.

"Flint is a mixture-of-experts model with 39.8 billion total parameters, using about 2.5 billion active parameters per token, released under the Apache 2.0 license," the official model card on Hugging Face says.

Per the model card, Flint has 49 layers — 2 dense layers and 47 MoE layers; each MoE layer contains 256 experts, of which 6 are selected per token. The context window is 262,144 tokens, and the knowledge cutoff is stated as June 2026. The model was trained from scratch: pre-training, mid-training, supervised fine-tuning, and reinforcement learning stages. The repository includes a BF16 checkpoint, model configuration, tokenizer, chat template, and an SGLang launch kit; modeling code for Transformers hasn't been added yet. The company published the model in the RecursiveSI/flint repository on Hugging Face; alongside the weights, the chat template and sample client code for SGLang serving are open as well.

Speed and Scores

In Recursive's tests, Flint generated about 131 tokens per second on a MacBook Pro with an Apple M4 Pro chip. For comparison: similar models like Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning delivered 58–69 tokens per second under the same conditions — Flint is nearly twice as fast. On the Nvidia RTX Spark device, speed exceeded 150 tokens per second. The distributed weights are in BF16 format and total about 80 gigabytes; separate quantized variants are prepared for Nvidia and Apple Silicon, and an SGLang-based launch kit is included — designed to run a single instance on a single B300 or H200 GPU. The model supports tool calls through an OpenAI-compatible interface; for complex tasks an output budget of at least 32,768 tokens is set, with generation settings of temperature 1.0, top-p 0.95, and top-k 20.

Benchmark results (from the official model card): SWE-bench Pro v2 at 89.7%, SWE-bench Verified at 74.2%, Terminal-Bench 2.1 at 54.9%. The company claims that with these scores, Flint leads among MoE models under 50 billion parameters in agentic coding. Evaluation conditions may differ across sources: in its measurements, Recursive counted solutions found via web search in Terminal-Bench as failures, while other developers may not have applied such a correction.

Company and Funding

Recursive was founded last year by Richard Socher; before that, Socher served as chief scientist at Salesforce. This May the company emerged from stealth and raised $650 million — a deal that valued the company at $4.65 billion. Investors include GV, Nvidia, and AMD Ventures. The first detailed report on the announcement was published in TechCrunch on October 21.

The model card also notes limitations: although safety-related examples were included in the supervised fine-tuning stage, Flint may still generate incorrect or harmful content, buggy code, or inappropriate tool calls.

More details in the TechCrunch story and the official model card on Hugging Face.