Recursive's Flint Posts 89.7% on SWE-bench Pro v2 Under Apache 2.0, With No Outside Evaluator

Recursive Superintelligence released Flint, a 39.8 billion-parameter coding model with Apache 2.0 weights, on Oct. 21. Its model card reports 89.7 percent on SWE-bench Pro v2, 15.5 points above its own score on SWE-bench Verified.

All three benchmark figures on the model card are developer-reported, and the card names no training datasets. Recursive says it built the model with its own automated research system, called Eureka Machine, according to The Next Web, which published its account on Oct. 23.

What the card lists

Flint is a mixture-of-experts model, a design that routes each token to a few of 256 expert sub-networks instead of the whole network. It has 39.8 billion parameters in total and about 2.5 billion active per token. The card documents a 262,144-token context window shared between prompt and output.

Item	Card value
Licence	Apache 2.0 (hosted API terms are separate)
Download size	about 80 GB
Serving kit	one B300 or H200 GPU per replica, SGLang
Missing	Transformers modeling code

The card recommends an output budget of at least 32,768 tokens for hard problems. It warns that Flint can produce incorrect code or inappropriate tool calls and says generated code should be tested. Recursive says the model may need more guidance than larger closed coding models and works best when tasks are split into small steps, according to The Next Web.

The benchmark figures

The card reports 54.9 percent on Terminal-Bench 2.1, 74.2 percent on SWE-bench Verified and 89.7 percent on SWE-bench Pro v2, each averaged pass@1 over at least five attempts at temperature 1.0. The card does not say how Pro v2 differs from the original SWE-bench Pro. Recursive says Flint leads mixture-of-experts models under 50 billion parameters, and the card notes that test setups vary.

1. SWE-bench Pro v289.7 %
2. SWE-bench Verified74.2 %
3. Terminal-Bench 2.154.9 %

Source: RecursiveSI/flint model card, accessed 2026-10-28

No third party is named as having rerun any of the three. The card's serving kit asks for a data-centre GPU, so the laptop framing rests on Recursive's separate speed claims: more than 125 tokens a second on a MacBook Pro with an M4 Pro chip, and more than 150 on Nvidia RTX Spark laptops. The Next Web, citing TechCrunch, says Google's Gemma 4 managed 58 to 69 tokens a second on the same Mac.

The Eureka Machine claim

Chief Executive Richard Socher said Eureka Machine's low-bit work shrank Flint by 71 percent with almost no quality loss, The Next Web reported. Recursive says the system compressed the output layer to 4 bits, which it describes as against common practice, and wrote custom code for Apple and Nvidia chips. Nvidia said that code sped Flint up by 55 percent on RTX Spark machines.

Recursive says a small team of people and agents built Flint in under four months. The Next Web reports the company left stealth in May after raising $650 million, with GV, Nvidia and AMD Ventures among its backers, and that TechCrunch puts its valuation at $4.65 billion. The card's only description of training is pretraining, mid-training, supervised fine-tuning and reinforcement learning on text, code and agentic tasks.

Where it fits

Coding agents are drawing developer attention on GitHub: JetBrains' ThinkRail gained 1,075 stars in a day. Tencent's Hy-Nomas also shipped under Apache-2.0 without published scores.

Recursive has not named an independent evaluator or a date for Transformers support.