Is Recursive Flint free to use commercially?

Recursive Flint's weights are released under the Apache 2.0 license on Hugging Face, which allows free use, modification and commercial use. Recursive notes that any hosted API terms and data practices are separate from the weight license. Running Flint locally means there are no per-token API fees, and code never leaves the machine.

What hardware do I need to run Flint locally?

Recursive's own rec engine runs Flint on an Apple Silicon Mac with an M4 Pro, M5 Pro or M5 Max and at least 48 GB of unified memory, with about 30 GB of free disk. Recursive also reports more than 150 tokens per second on a 64GB NVIDIA RTX Spark. The full BF16 checkpoint is about 80 GB and is served on one B300 or H200 GPU.

How long is Flint's context window?

Flint's Hugging Face model card lists a 262,144-token context window shared by the prompt and the output. Recursive's launch post says the model supports up to 512k tokens but recommends staying within 256k for the best results. The model card also warns that the context size is a capacity limit, not a promise of equal accuracy at every length.

How does Flint compare to larger coding assistants?

Recursive says Flint rivals dense models with 25–40 billion parameters on coding while running more than five times faster. TechCrunch reports that only Qwen3.8 27B scored higher on Recursive's six-test average. Recursive also admits Flint needs clearer instructions than large paid coding assistants, and suggests breaking big tasks into smaller steps.

Recursive · 2026-10-21 · major

Flint — Recursive's open 40B coding model runs on a laptop

Flint is Recursive's first model: a 40B mixture-of-experts coding model with 2.5B active parameters, open under Apache 2.0. It runs at 125+ tokens/s on an M4 Pro MacBook Pro, and Recursive says its automated research system built most of it.

Recursive Flint launch graphic from the official announcement

An open-weight coding model small and fast enough to run agents on your own Mac or RTX Spark.

Key specs

SWE-bench	74.2%
Speed (m4 pro)	125+ tok/s

Quick facts

Maker	Recursive (CEO Richard Socher)
Parameters	39.8B total, ~2.5B active per token
Context window	262,144 tokens (model card)
License	Apache 2.0
Terminal-Bench 2.1	54.9%
Runs with	llama.cpp, Ollama, LM Studio, vLLM, SGLang

What is it?

Flint is the first model from Recursive, the AI startup led by former Salesforce chief scientist Richard Socher. It is an open-weight mixture-of-experts model with 39.8B total parameters that is built for coding agents: it explores codebases, fixes bugs, implements features and writes tests. The weights are on Hugging Face under Apache 2.0.

How does it work?

Each token uses only about 2.5B of the model's parameters: every MoE layer picks 6 of 256 experts plus a shared expert, and 35 of the 49 layers use a 512-token sliding window instead of full attention. Recursive says its Eureka Machine, an automated research system, proposed and tested ideas at almost every stage, including the sparser architecture, low-bit training and custom kernels, and that a small team built the model from scratch in under four months.

Why does it matter?

Recursive reports 125+ tokens per second on an M4 Pro MacBook Pro and 150+ on a 64GB RTX Spark, and says the model leads MoE models under 50B parameters on SWE-bench Pro v2, Terminal-Bench 2.1 and SWE-bench Verified. That makes private, offline agentic coding with no per-token bill realistic on one machine. The benchmark numbers are Recursive's own and not yet independently checked.

Who is it for?

developers who want a local coding agent

Frequently asked questions

Is Recursive Flint free to use commercially?: Recursive Flint's weights are released under the Apache 2.0 license on Hugging Face, which allows free use, modification and commercial use. Recursive notes that any hosted API terms and data practices are separate from the weight license. Running Flint locally means there are no per-token API fees, and code never leaves the machine.
What hardware do I need to run Flint locally?: Recursive's own rec engine runs Flint on an Apple Silicon Mac with an M4 Pro, M5 Pro or M5 Max and at least 48 GB of unified memory, with about 30 GB of free disk. Recursive also reports more than 150 tokens per second on a 64GB NVIDIA RTX Spark. The full BF16 checkpoint is about 80 GB and is served on one B300 or H200 GPU.
How long is Flint's context window?: Flint's Hugging Face model card lists a 262,144-token context window shared by the prompt and the output. Recursive's launch post says the model supports up to 512k tokens but recommends staying within 256k for the best results. The model card also warns that the context size is a capacity limit, not a promise of equal accuracy at every length.
How does Flint compare to larger coding assistants?: Recursive says Flint rivals dense models with 25–40 billion parameters on coding while running more than five times faster. TechCrunch reports that only Qwen3.8 27B scored higher on Recursive's six-test average. Recursive also admits Flint needs clearer instructions than large paid coding assistants, and suggests breaking big tasks into smaller steps.

Try it

vllm serve "RecursiveSI/flint"