Recursive released Flint, a 40B MoE coding model under Apache 2.0

If your team is weighing a self-hosted coding model, Recursive just added a candidate. The company published Flint on Hugging Face under Apache 2.0, a 39.8 billion parameter decoder-only mixture-of-experts transformer with about 2.5 billion active parameters per token. The model card lists 49 layers, 256 routed experts per layer with six selected per token, a 262,144 token context window and a knowledge cutoff of June 2026.

Recursive's own card lists three benchmark figures. Terminal-Bench 2.1 at 54.9 percent. SWE-bench Pro v2 at 89.7 percent. SWE-bench Verified at 74.2 percent. Those are the company's own numbers on its own card. TechCrunch, reporting on the launch, adds Recursive's claim that Flint ran at around 131 tokens per second on an Apple MacBook Pro M4 Pro, which the article says is roughly double the speed of comparable models at the same size, and that Flint outperformed models under 50 billion total parameters on SWE-bench Verified in Recursive's testing.

What is new, and what is not

The two interesting qualities for a working team are the open licence and the active-parameter count. Apache 2.0 means a company can take the weights, fine-tune them on its own codebase and run the result on its own hardware without paying Recursive. The 2.5 billion active parameters per token is why Flint fits on a laptop at reasonable speed: an MoE model activates only a slice of its total weights per token, so memory and compute sit closer to a dense model of a few billion parameters than one of forty.

What is not new is a vendor benchmark. Recursive's three figures are self-reported, and the company's post describing the model's development states that its own Eureka Machine automated research system contributed to "nearly every step" of Flint's creation, including pre-training and post-training. The claim that the company built the model largely with AI is a claim about process rather than quality, and the self-benchmarked quality is what you can measure yourself today.

What a team should actually do

Download the weights from Hugging Face at RecursiveSI/flint, point your usual harness at them, and run against the subset of tasks you already know the quality of your current coding model on. The repository ships quantised variants, named flint-int4, flint-int4-gguf and flint-nvfp4-gguf, plus a community unsloth GGUF, so a Mac with 32 to 64 gigabytes of memory can host one of them. Compare on bug-fix prompts from your own repository, not on a benchmark pass rate, because the SWE-Bench numbers describe public tasks and your own code is the one that matters.

What Recursive is pitching alongside

TechCrunch reports Recursive is a two-year-old company led by Richard Socher, former chief scientist at Salesforce, that emerged from stealth in May on 650 million dollars in funding at a 4.65 billion dollar valuation. Flint is positioned as the first result of the Eureka Machine. The piece reads as a vendor pitch amplified: the figures and the process claim both come from Recursive. The model weights are the test everyone can run.

Source

Primary: Hugging Face: RecursiveSI/flint model card. Reporting: TechCrunch: Recursive launches Flint, a laptop coding model it says was largely built by AI.