AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Recursive shows first result of its AI-driven research process with Flint

Collected Oct 22, 2026

Recursive, the AI startup co-founded by Richard Socher and including former DeepMind researcher Tim Rocktäschel, has released Flint, its first language model. Flint is a 40-billion-parameter mixture-of-experts model aimed at agentic coding that runs on local hardware, and it was built in roughly four months. Recursive says an internal AI system called the Eureka Machine was involved in almost every stage of development, proposing ideas, building them, testing them and planning follow-up experiments from the results. The company came out of stealth in May 2026, having raised $650 million at a $4.65 billion valuation, at a point when it had no concrete technical results to show.

The architecture is a sparse mixture of experts: of the 40 billion total parameters, only about 2.5 billion are active per token. That design is the basis of Recursive's speed claims. The company says Flint competes with dense models in the 25 to 40 billion parameter range while running more than five times faster, and that inference is up to twice as fast as similar models. On a MacBook Pro with an M4 Pro chip, Recursive reports more than 125 tokens per second at short context lengths; on an Nvidia RTX Spark with 64 GB of memory it reports over 150. Mixture-of-experts matters here because only a fraction of the network is computed for each token, which cuts compute per generated token at the cost of holding all the weights in memory — part of why the memory figures (48 GB on the Mac, 64 GB on the RTX Spark) are quoted alongside the speed numbers.

On benchmarks, Recursive claims Flint leads coding evaluations including SWE-Bench Pro v2, Terminal-Bench 2.1 and SWE-Bench Verified among MoE models under 50 billion parameters, naming Qwen3.6-35B-A3B and Laguna S 2.1 as comparisons. It also says Flint leads all six of its coding and instruction-following benchmarks among sparse MoE models of similar size, though dense models such as Qwen 3.8 score higher in some tests. These results have not been independently verified, so they should be treated as vendor claims for now. The pattern the company describes — strong results against sparse peers, mixed results against larger dense models — is a familiar trade-off in this size class.

Recursive gives three concrete examples of the Eureka Machine's influence. Starting from an architecture with 35 billion total and 3 billion active parameters, the system tested different expert configurations and settled on 40 billion total and 2.5 billion active as the best balance of speed and coding performance. For quantization, it departed from common practice by reducing the output layer to four bits as well, adding a distillation step to recover accuracy; Recursive says this needed roughly half as many training steps as the prior method, and that the finished NVFP4 model trails the 16-bit version by only 0.7 points across benchmarks. The system also generated specialized compute kernels for Apple Silicon and Nvidia RTX Spark; on Apple's M4 and M5 chips a Metal "megakernel" fuses the entire forward pass during token generation into one operation. In practice, that kind of kernel-level fusion reduces launch overhead and memory traffic, which is where much of the gain at short context lengths would come from.

On the product side, Flint supports context windows up to 512,000 tokens, with Recursive recommending a cap at 256,000. Weights are on Hugging Face under the Apache 2.0 license, and supported runtimes include llama.cpp, Ollama, LM Studio, vLLM and SGLang. Recursive also shipped rec, its own inference engine for Apple Silicon. Because everything runs locally, code does not need to be sent to external servers — a meaningful draw for teams with data-residency or confidentiality constraints. Recursive acknowledges the model needs more guidance on complex tasks than large proprietary coding models and recommends breaking bigger jobs into smaller steps, which fits the agentic-coding use case it targets.

The broader pitch is recursive self-improvement: each model built with the Eureka Machine is supposed to make the machine a better researcher, so the next generation can be produced faster and more cheaply, with the same approach eventually applied to other scientific fields. Flint is a first public data point but does not demonstrate that loop — that would require showing a model built this way measurably improves the Eureka Machine itself. Outside voices are mixed: former Google DeepMind research lead Oriol Vinyals is said to consider AI-assisted self-improvement realistic while doubting an intelligence explosion, and finding good research ideas and reliably judging results remain hard problems. Other labs are pursuing similar automation — Anthropic says Claude handles a large share of its own development work, and OpenAI deploys agents across its research pipeline — while a study of more than 700 task logs found agents do much of the work but humans still make most critical research calls.

Why it matters: Developers who want a capable coding model on a laptop or workstation now have an openly licensed, fast local option in a size class that fits consumer-grade memory, assuming the unverified benchmarks hold up. For product teams, the more consequential claim is the process: if AI-proposed experiments can reliably shorten model development, the cost and cadence of shipping specialized models could change — though on the evidence here that remains a hypothesis, not a demonstrated result.

Read at The Decoder

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

The AI startup Recursive is unveiling Flint, its first language model, a 40-billion-parameter MoE for local coding developed by former DeepMind researchers. It was created in four months using the company's in-house AI system, "Eureka Machine," which independently tests and optimizes architectures. Recursive has yet to prove whether this approach will actually lead to self-improving AI. The article Recursive shows first result of its AI-driven research process with Flint appeared first on The Decoder .