Recursive's AI-Built Coding Model Ships Open Weights — and a Benchmark Table It Grades Itself

Recursive, the roughly one-year-old startup built around the idea that AI should do AI research — it emerged from stealth only this past May — put out its first public model last week: Flint, a 40-billion-parameter coding model released under an unrestricted Apache 2.0 license, with full weights on Hugging Face and claimed support across the major local-inference runtimes out of the gate. That's a real contribution to open-weight AI — a usable, locally-runnable coding model with no license fine print. What's less settled is everything Recursive says makes it special.

The pitch: a model that mostly built itself

Recursive's headline claim is that an internal system it calls the Eureka Machine — described as a loop that proposes architecture and training changes, implements them, and tests the results — handled most of the engineering between a baseline design and the shipped model. The company says Eureka searched expert counts and layer widths to land on Flint's final shape, pushed the output layer down to 4-bit precision against the usual engineering advice, and wrote custom low-bit kernels for Apple and Nvidia hardware. None of that is independently checkable. The only source for the Eureka Machine's existence and its contributions is Recursive itself, in the same documents that report Flint's scores.

The numbers don't compare what they appear to compare

Recursive's release chart puts Flint ahead of every mixture-of-experts model under 50 billion parameters on agentic coding, including Alibaba's Qwen3.6-35B-A3B. But the comparison stitches together different measurement regimes: Flint's Terminal-Bench score uses version 2.1 of the benchmark, while the competitors' are drawn from their own model cards on version 2.0. Recursive's own documentation flags that its SWE-bench Pro v2 figure — Flint's best number, at 89.7% — is not comparable to the public SWE-bench Pro leaderboard at all. And on the one score that is apples-to-apples, Terminal-Bench 2.1, Flint posts its weakest result, well behind the dense Qwen3.8 model Recursive includes in the same table for context. The 2.5-billion-parameter 'active' figure Recursive advertises also includes the output head; strip that out and the active compute per token is closer to 2.2 billion, a small but telling rounding in the direction that makes the model look lighter.

The speed claims hold up less well at scale

The laptop pitch — over 125 tokens per second on an M4 Pro MacBook, more than 150 on Nvidia's RTX Spark — is the most consumer-facing part of the launch, and it's also the part with the most fine print. Those figures are measured at short context lengths, and Flint's own model card lists a documented context window of 262,144 tokens while Recursive's blog post advertises 'up to 512K' elsewhere. Throughput at the long contexts a 40-billion-parameter coding agent would actually chew through on a real repository is not in the release materials at all. For a company whose entire premise is that automated research finds efficiency gains humans miss, the gap between the number in the headline and the number that would actually describe a workday is a strange one to leave unaddressed.

What's actually new here

Strip away the self-graded scoreboard and there's a less flashy but sturdier story: a well-funded new entrant — Recursive raised $650 million earlier this year from GV, Nvidia, and AMD Ventures at a $4.65 billion valuation — chose to open-source its first model rather than gate it behind an API, and benchmarked itself against Chinese open models (Qwen3.6, Qwen3.8) as the default competitive bar rather than against OpenAI or Anthropic. That's consistent with where open-weight coding models have been heading all year: the competition that matters now runs on Hugging Face, not behind a paywall. Recursive's actual achievement is shipping a locally-runnable, no-asterisks-licensed coder. Whether an AI system built most of it, and whether it truly leads its weight class, are claims the company has not given anyone outside its own walls a way to check.