Recursive releases Flint, a free laptop coding model it says was built largely with AI help

Recursive, an AI startup that wants to build AI that can improve itself, released its first model on Wednesday 21 October. Flint is a coding model that Recursive says is built to run on local hardware. Its weights are free to download, and the company says an automated research system it calls the Eureka Machine helped develop nearly every part of it.

Several things in that paragraph are the company's own account, so this piece keeps what Recursive says apart from what has shipped.

What shipped

Flint's weights are on Hugging Face under the Apache 2.0 licence, according to both Recursive's announcement and the model's Hugging Face page. TechCrunch describes that licence as an open source licence that allows free use and modification.

Recursive describes Flint as a mixture-of-experts model with 40 billion total parameters, of which it activates 2.5 billion per token. The Hugging Face model card gives the total as 39.8 billion and the active figure as approximately 2.5 billion. It says "40B-A2.5B" is the rounded designation. The card lists a documented context configuration of 262,144 tokens. Recursive's announcement says Flint supports context windows of up to 512k tokens and recommends staying within 256k for the best results.

Recursive says it is also releasing rec, a lightweight inference engine optimised for Flint on M4 Pro and M5 Pro/Max chips. It lists llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang as tools and frameworks Flint works with.

The model card adds detail on what the main repository contains. It holds the BF16 checkpoint, the configuration, the tokeniser, the chat template and an SGLang serving kit, and the card puts the download at about 80 GB. The card says that kit runs on one B300 or H200 GPU per replica, and that quantised variants for NVIDIA and Apple Silicon runtimes are "prepared separately".

What Recursive claims about performance

Flint's own figures below come from Recursive's testing; the model card says some of the comparison scores are reported by other developers.

• Speed. Recursive says Flint generates more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 on a 64GB NVIDIA RTX Spark, at short contexts. TechCrunch puts the M4 Pro figure from Recursive's tests at about 131. The company says that is up to twice as fast as comparable models. TechCrunch reports that it is roughly twice as fast as similar models on the same machine, naming Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning.
• Benchmarks. Recursive says Flint leads mixture-of-experts models with fewer than 50 billion total parameters on SWE-Bench Pro v2, Terminal-Bench 2.1 and SWE-Bench Verified. It also says Flint rivals the coding performance of dense models with 25 to 40 billion parameters while running more than five times faster. The model card gives its release scores as 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified.
• A model that did better. TechCrunch notes that on Recursive's own average across six tests, one model, Alibaba's Qwen3.8 27B, scored higher than Flint.
• Compression. Recursive says its final NVFP4 version of Flint scores within 0.7 points of the 16-bit model across its benchmarks.

The model card is explicit that comparing scores is not straightforward. It says evaluation setups vary across sources and that Flint's Terminal-Bench score counts solutions obtained through web retrieval as unsuccessful. Developer-reported scores, it says, "may not apply the same adjustment".

Recursive also tempers the pitch. Its announcement says that for complex tasks users should break the work into smaller subtasks and give explicit instructions, because Flint "may need more guidance than the larger closed-source coding models you’re used to". TechCrunch reports that Wired's Will Knight used Flint on a Mac with 48GB of memory to build several small programs, including a to-do list app and a tool that searches the web for research. According to TechCrunch, he found it handled his coding requests every time, where smaller coding models had often hit errors or run slowly compared with Claude or Codex.

The Eureka Machine claim

The part of the launch that goes beyond a new model is how Recursive says it was made. The company describes the Eureka Machine as a system for large-scale automated AI research and development. It says the system proposes ideas, implements them, tests them and uses the results to decide what to try next. Recursive says a small team of people and agents built Flint from scratch in under four months, and that the Eureka Machine helped develop nearly every part of the model-building stack, from data and training recipes to infrastructure and inference optimisation.

Recursive's announcement gives three examples:

• Architecture. It says the Eureka Machine started from a baseline of 35 billion total and 3 billion active parameters and explored different expert configurations and widths. The 40-billion, 2.5-billion-active design was projected to give the best mix of speed and coding performance among the configurations it tested.
• Quantisation. Against common practice, the company says, the system quantised the output layer to 4 bits, then added an offline distillation stage to recover accuracy.
• Kernels. It says the system produced specialised kernels for the target hardware, including a Metal megakernel that fuses the entire decode forward pass on M4 and M5 chips.

TechCrunch notes that Recursive did not say how much time the Eureka Machine saved its researchers. It places the work alongside OpenAI's statement last month that by mid-August its AI agents were logging 3.1 workdays for every one worked by a human researcher.

Why the framing matters

TechCrunch calls Flint an early public test of recursive self-improvement, the idea that AI can build better AI, which in turn builds better AI still. It describes this as one of the industry's most debated ideas, with some researchers fearing that AI improving itself could slip out of human control. For now, TechCrunch says, Recursive's version is more limited: an AI research assistant overseen by people that produced a model small enough for a laptop.

Recursive's stated goal is to build "recursively self-improving superintelligence" to automate knowledge discovery. It says Flint is its starting point and that each new model should make the Eureka Machine a better researcher, helping build the next one faster. TechCrunch reports that Socher told Wired the longer-term goal is full recursive self-improvement, with humans overseeing the process and able to stop it if something goes wrong.

My reading is that Flint itself is a modest, checkable thing: a downloadable model with published weights, benchmark scores and a licence. The larger claim, that automated research did much of the building, rests on the company's own description of its process.

The company behind it

TechCrunch reports that Recursive was founded last year and emerged from stealth in May with $650 million in funding at a $4.65 billion valuation, from backers including GV, Nvidia and AMD Ventures. It says the company is led by chief executive Richard Socher, Salesforce's former chief scientist, and that its $410 million compute deal with Amazon Web Services was struck in June.

Sources

Mehta, A. (2026) 'Recursive launches Flint, a laptop coding model it says was largely built by AI', TechCrunch, 21 October 2026, available at: https://techcrunch.com/2026/10/21/recursive-launches-flint-a-laptop-coding-model-it-says-was-largely-built-by-ai/ (Accessed: 22 October 2026).

Recursive (2026) 'Introducing Flint', Recursive, available at: https://recursive.com/updates/introducing-recursive-flint (Accessed: 22 October 2026).

Recursive Superintelligence, Inc. (2026) 'RecursiveSI/flint — model card', Hugging Face, available at: https://huggingface.co/RecursiveSI/flint (Accessed: 22 October 2026).

Picture credits

• Top picture: Richard Socher, chief executive of Recursive. Richard Socher at World Economic Forum Annual Meeting 2025, photo: World Economic Forum, CC BY 3.0 (source)
• A MacBook Pro laptop. MacBook Pro 13" M1 2020, photo: DYVER, CC BY-SA 4.0 (source)