The short version
Flint makes a useful coding agent possible without sending a single line of code to a third-party cloud. Released by Recursive on October 21, 2026, the open-weight model has 40 billion parameters but activates only about 2.5 billion per token, so it runs on a 48GB laptop. For companies planning generative AI integration into their engineering workflow, it widens the choice between paying per token and keeping code in-house.
What did Recursive release?
Recursive, the AI lab led by former Salesforce chief scientist Richard Socher, came out of stealth in May with $650 million in funding at a reported $4.65 billion valuation. Flint is its first product: a model built specifically for coding agents that read a codebase, fix bugs, implement features and write tests, with support for reasoning and multi-turn tool use.
The weights are published on Hugging Face together with dozens of quantized variants for Nvidia and Apple hardware. Flint works with the runtimes most teams already use — Ollama, LM Studio, llama.cpp and Jan on desktops, vLLM and SGLang on servers. According to TechCrunch, Recursive’s six-benchmark average placed it behind only Alibaba’s Qwen3.8 27B among comparable local models, and on SWE-bench Verified it led every model under 50 billion parameters.
How does a 40B model fit on a laptop?
The trick is sparsity. In a mixture-of-experts design only a small slice of the network works on each token, so Flint computes like a 2.5-billion-parameter model while storing the knowledge of a much larger one. Recursive also trained it in low precision and compressed the output layer to 4 bits, which it says cut the model’s size by 71% with almost no quality loss, and wrote custom kernels for Apple and Nvidia chips.
The result, per TechCrunch, is about 131 tokens per second on an M4 Pro MacBook Pro — roughly twice the 58–69 tokens per second it measured for Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning. That speed matters for agents, which make many model calls per task: a slow local model turns a two-minute fix into a coffee break.
What does “built by AI” mean here?
Recursive says its automated research system, the Eureka Machine, proposed, built and tested ideas at nearly every stage — architecture, pretraining, post-training, inference and kernel optimization — letting a small team build Flint from scratch in under four months. Socher told WIRED, as TechCrunch reported, that the long-term goal is “full recursive self-improvement” with humans overseeing and able to stop the process. For buyers, the practical takeaway is narrower: model development cycles are compressing, so the gap between a niche open model and a usable one is closing faster than procurement cycles assume.
Where does Flint fall short?
Recursive is candid that Flint needs more guidance than large paid assistants and recommends breaking big tasks into smaller steps. Its model card warns that long reasoning chains can exhaust the output budget before finishing, that results vary by quantization and runtime, and that generated code must be tested before deployment. The benchmark figures are vendor-reported. A 48GB machine is also not a standard developer laptop, so the “runs locally” promise has a hardware price.
What it means for US & EU software teams
For regulated teams in FinTech, HealthTech and the public sector, code privacy has been the main blocker for AI coding agents. A model that runs entirely on company hardware removes the data-transfer question for source code and secrets in prompts, which simplifies GDPR and client-contract reviews and suits on-premise or air-gapped environments. The Apache 2.0 license also allows fine-tuning on internal code without vendor terms.
The sensible architecture is a routed one: a local model like Flint for routine edits, test generation and work on sensitive repositories, with a frontier cloud model reserved for hard, multi-file reasoning. That cuts per-token spend and reduces lock-in to any single assistant vendor. Before rolling it out, benchmark Flint on your own repositories, budget for 48GB-class machines or a shared GPU server, and keep CI tests and human review as the gate for anything an agent writes.
Frequently asked questions
What is Recursive Flint?
Flint is the first model from AI startup Recursive, released on October 21, 2026. It is an open-weight mixture-of-experts coding model with 39.8 billion total parameters, of which about 2.5 billion are active per token. It is designed for coding agents that investigate codebases, fix bugs, implement features and write tests, and it is free to download from Hugging Face under the Apache 2.0 license.
Can Flint really run on a laptop?
Yes, on a well-equipped one. Recursive says Flint runs at more than 125 tokens per second on a MacBook Pro and above 150 tokens per second on an Nvidia RTX Spark, and TechCrunch reported about 131 tokens per second on an M4 Pro machine. Apple Silicon users need an M4 Pro, M5 Pro or M5 Max chip with at least 48GB of memory. Server deployments are documented for vLLM and SGLang.
How good is Flint at coding compared with paid assistants?
Recursive reports 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1, and says Flint beats other models under 50 billion parameters on SWE-bench Verified. A WIRED reviewer said it handled his small coding requests every time. Recursive itself notes that Flint needs more guidance than large paid assistants and recommends splitting big tasks into smaller steps.
What does it mean that Flint was built by AI?
Recursive says its automated research system, the Eureka Machine, contributed to nearly every stage of development: architecture, pretraining, post-training, low-bit training that shrank the model by 71% with little quality loss, and custom kernels for Apple and Nvidia chips. A small team built the model from scratch in under four months. Humans still oversaw the process.
Should companies use Flint for proprietary code?
Flint is a strong candidate for workloads where source code must not leave company machines, because it runs fully offline and the Apache 2.0 license permits commercial use and modification. Teams should still benchmark it on their own repositories, keep tests and code review in the loop, and treat it as one model in a routed setup rather than a drop-in replacement for frontier assistants.
Sources
Hugging Face — RecursiveSI/flint model card (October 21, 2026)
TechCrunch — Recursive launches Flint, a laptop coding model it says was largely built by AI (October 21, 2026)
The Next Web — Recursive releases Flint, a laptop coding model largely built by its own AI (October 21, 2026)