Recursive releases Flint, an Apache 2.0 open-weight coding model

Recursive released Flint, an open-weight mixture-of-experts coding model, on Hugging Face on October 21, 2026. The model card lists 39.8 billion total parameters and approximately 2.5 billion active per token, a documented context configuration of 262,144 tokens and the Apache 2.0 license. Recursive says Flint scores 74.2% on SWE-bench Verified in its own release evaluations.

#What changed?

Flint is Recursive's first model. The card describes it as "Recursive’s mixture-of-experts language model for coding agents" and says "It is designed to investigate codebases, fix bugs, implement features, and write tests." It "supports text input and output, reasoning, and multi-turn tool use," and the card says "Flint supports OpenAI-compatible tool calls."

Recursive rounds the size to "40B-A2.5B", and its announcement and WIRED say 40 billion. Recursive's blog says Flint "supports context windows of up to 512k tokens," while the model card documents 262,144 tokens. The card adds that the context figure "is a capacity limit, not a guarantee of uniform accuracy across all sequence lengths or tasks."

The card says Flint "was trained from scratch through pretraining, mid-training, supervised fine-tuning, and reinforcement learning." Recursive says "Our small team of people and agents built it from scratch in under four months," using its Eureka Machine, which it calls "our system for large-scale automated AI research and development."

#Who can use it and when?

Anyone can download the weights, which are not gated on Hugging Face. The card says "Hosted API terms and data practices are separate from the weight license." None of our sources gives hosted API pricing. Richard Socher, whom WIRED identifies as Recursive CEO and cofounder, wrote on X that you can "try it for free on @opencode or @pidotdev."

The SGLang kit says "80 GB GPUs are not supported." The model card says "Transformers modeling code is not included yet."

#How does it compare?

All scores below come from Recursive. The model card says: "In our release evaluations, it scores 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2, and 74.2% on SWE-bench Verified." It notes that "Evaluation setups vary across sources." No independent benchmark results were found.

Recursive's blog claims Flint "leads MoE models with fewer than 50 billion total parameters on agentic coding benchmarks." It also says "The final NVFP4 model scores within 0.7 points of the 16-bit model across our benchmarks." The blog also cautions that Flint "may need more guidance than the larger closed-source coding models you’re used to."

On speed, Recursive says Flint generates "more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 tokens per second on a 64GB NVIDIA RTX Spark at short contexts." NVIDIA, which WIRED lists as an investor in Recursive, says in its blog that Recursive's system created "specialized kernels that boosted Flint’s generation speed on RTX Spark by 55% without changing the model’s weights."

WIRED tested Flint on a Mac with 48 GB of RAM. Its reviewer wrote that "Flint nailed my coding requests every time." WIRED reported that "Recursive has raised $650 million at a $4.65 billion valuation from Google Ventures, Greycroft, Nvidia, and AMD Ventures."

For another recent open model release, see our coverage of DeepSeek V4.1 Flash. More is in our Developers hub.

#How to get it

1. Download the weights from Hugging Face, or pick a quantized version that fits your hardware.
2. On a Mac, Recursive's rec engine needs "an M4 Pro, M5 Pro, or M5 Max" with at least 48 GB of unified memory and about 30 GB of free disk space. Recursive says rec is "optimized for Flint on M4 Pro and M5 Pro/Max."
3. Other tools: Recursive says "Flint works with tools and frameworks including llama.cpp, Ollama, Unsloth, LM Studio, vLLM, and SGLang."
4. Settings: the card's evaluations use temperature 1.0, top-p 0.95 and top-k 20, and it recommends an output budget of "At least 32768 tokens, including reasoning" for hard problems.

The card warns that the model "can still produce incorrect or harmful content, faulty code, or inappropriate tool calls," and that "Tool-using applications should enforce their own permissions and execution boundaries." For one way to limit what a local coding agent can reach, see our guide to controlling an AI coding agent's network access on a Mac.

#What is still unclear?

• Independent benchmarks. Every score so far is Recursive's own; WIRED's test is one reviewer's experience.
• Hosted API pricing and terms. Recursive says they are separate from the weight license but has not published them in our sources.
• Framework support. Ollama and vLLM support are Recursive's claims; we did not verify them.
• Context. The blog's 512k figure and the card's 262,144 tokens differ.