AnswersPublished18 sources
Recursive Flint: What AI Building AI Actually Means
Released on October 21, 2026, Recursive Flint is a 40 billion parameter coding model that activates 2.5 billion parameters per token and runs locally. Nvidia’s announcement reports more than 220 tokens per second on a 64GB RTX Spark; Recursive separately reports more than 150 tokens per second on that hardware at sh...
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Released on October 21, 2026, Recursive Flint is a 40 billion parameter coding model that activates 2.5 billion parameters per token and runs locally.
- Nvidia’s announcement reports more than 220 tokens per second on a 64GB RTX Spark; Recursive separately reports more than 150 tokens per second on that hardware at short contexts.
- Recursive describes Flint’s benchmark results as best in class for certain model comparisons.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What is Recursive Superintelligence’s Recursive Flint, released on October 21, 2026, and how does its launch illustrate both the promise and. Article summary: Recursive Flint is Recursive Superintelligence’s first free, open-weight coding model, released on October 21, 2026, for running coding agents on local hardware; the company says AI helped build nearly its entire develop. Topic tags: general, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fa
Recursive Flint is Recursive Superintelligence’s first open-weight coding model, designed to run coding agents on local hardware. Its launch drew attention not only for the model’s reported speed and benchmark results, but also for the company’s claim that its Eureka Machine helped develop much of it. That makes Flint an example of AI-assisted model development—not proof that an AI system can independently and indefinitely improve itself. 1
5
8
What is Recursive Flint?
Flint is a mixture-of-experts (MoE) model with 40 billion total parameters, of which 2.5 billion are active for each token. Recursive presents it as a fast, local model for agentic coding tasks. 1
17
The model is available as downloadable weights on Hugging Face under the Apache 2.0 license, according to launch coverage. Reports describe it running locally on consumer hardware, including Apple Silicon Macs and Nvidia GPUs with 48GB of memory. 3
6
Nvidia included Flint in an announcement about local AI in Europe, describing it as optimized for RTX Spark. That promotion puts the model in the context of Nvidia’s local-AI hardware push; the available announcement excerpt does not establish that Flint was demonstrated onstage at a specific GTC Berlin session. 1
How fast is it, and how strong are its benchmarks?
The published speed claims differ. Nvidia’s announcement says Flint generates more than 220 tokens per second on a 64GB RTX Spark. Recursive’s own announcement reports more than 150 tokens per second on a 64GB RTX Spark at short contexts, and more than 125 tokens per second on an M4 Pro MacBook Pro. The sources do not make the conditions behind the two Spark figures clear enough to compare them directly. 1
17
Recursive also describes Flint’s scores on SWE-Bench Pro v2, Terminal-Bench 2.1 and SWE-Bench Verified as best-in-class among MoE models with fewer than 50 billion total parameters. That is a specific comparison, not a claim that Flint leads every coding model. The results are company-reported; the supplied reporting does not provide an independent replication of the benchmarks. 1
17
What does it mean that AI helped build Flint?
Recursive says its automated research system, Eureka Machine, assisted with nearly every stage of Flint’s development, including training, architecture work and code optimization. The company says this helped a small team build the model in under four months. These are Recursive’s claims about its development process, rather than an independently audited breakdown of which work was done by people and which by AI. 2
5
8
That distinction matters. Using AI to assist with parts of model development is different from a system autonomously choosing what to improve, carrying out the work, evaluating the results and launching another development cycle without human direction. Flint is presented as a test of AI-assisted research; the available evidence does not establish that a fully autonomous recursive-improvement loop has been achieved. 8
19
Wired journalist Will Knight described Flint as an impressive open-weight coding model and said it suggested that fully recursive self-improvement was getting closer. That is an assessment of the model’s significance, not evidence that Flint itself can independently build its successor. 8
Why the launch is drawing both interest and concern
The practical appeal is straightforward: a coding model that can run locally gives users an alternative to sending coding requests to a cloud service. Meanwhile, AI assistance in research and development could help a small team move through model-building work faster, as Recursive says happened with Flint. 3
5
The controversy concerns the more ambitious goal behind the launch: using AI to improve AI. Reporting describes the potential for recursive self-improvement as both promising and controversial, while noting that the technology has limitations. Recursive CEO Richard Socher has said the company’s system is designed to remain under human control; that assurance is a company statement, not independent verification of future systems’ safety. 8
19
Recursive’s financial backing adds context to the scale of its ambitions. TechCrunch reported that the company emerged from stealth with $650 million in funding at a $4.65 billion valuation, with GV, Nvidia and AMD Ventures among its backers. 5
Flint’s launch is therefore best read as an early, concrete example of a company using AI in model development, paired with a local coding model that has attracted attention for its reported performance. It is evidence of AI-assisted work—not proof that AI can already improve itself indefinitely or operate without human oversight. 5
8
19
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "Recursive Flint: What AI Building AI Actually Means"?
Released on October 21, 2026, Recursive Flint is a 40 billion parameter coding model that activates 2.5 billion parameters per token and runs locally.
What are the key points to validate first?
Released on October 21, 2026, Recursive Flint is a 40 billion parameter coding model that activates 2.5 billion parameters per token and runs locally. Nvidia’s announcement reports more than 220 tokens per second on a 64GB RTX Spark; Recursive separately reports more than 150 tokens per second on that hardware at short contexts.
What should I do next in practice?
Recursive describes Flint’s benchmark results as best in class for certain model comparisons.
Sources
blogs.nvidia.comBringing the Spark to Europe: Local Intelligence Powers Next ...
techcrunch.comRecursive launches Flint, a laptop coding model it says was largely built by AI | TechCrunch
wired.comThis Startup Is All In on Recursive Self-Improvement. Its Latest Model Is Scary Good
recursive.comIntroducing Flint - Recursive