Recursive Flint, released on 21 October, is a free coding model built to run on a laptop rather than in the cloud. Its maker, the AI startup Recursive, says the model was designed and trained with heavy help from its own automated research system. For developers who pay for every token or work on unreliable internet, the more practical news is simpler: a capable coding assistant you can download once and use offline.
Here is what Flint is, what the company claims for it, where the claims stop, and how to try it.
What Recursive Flint is
Flint is a mixture-of-experts model with 40 billion parameters in total, of which about 2.5 billion are active for each token, according to Recursive's announcement. That design is why it can run quickly on consumer hardware: most of the model sits idle on any given step.
- Context window: up to 512,000 tokens, though Recursive recommends staying within 256,000 for best results.
- Licence: Apache 2.0, which allows free commercial use and modification.
- Where to get it: the weights are on Hugging Face under RecursiveSI/flint.
- Tools it works with: llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang, plus Recursive's own lightweight engine, called rec.
The final release uses 4-bit NVFP4 quantisation. Recursive says this version scores within 0.7 points of the full 16-bit model across its benchmarks.
How fast is it?
Recursive says Flint generates more than 125 tokens per second on a MacBook Pro with Apple's M4 Pro chip and more than 150 tokens per second on a 64GB Nvidia RTX Spark at short context lengths. TechCrunch reports about 131 tokens per second on the M4 Pro in Recursive's tests, against 58 to 69 for similar models such as Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning.
Two caveats apply. These are the company's own figures, and they are measured at short contexts. Recursive does not publish speeds for long prompts, which is where local models usually slow down.
Benchmarks
Recursive calls Flint "the best-performing MoE model with fewer than 50 billion total parameters on agentic coding benchmarks," naming SWE-Bench Pro v2, Terminal-Bench 2.1 and SWE-Bench Verified. It also says Flint rivals dense models of 25 to 40 billion parameters while running more than five times faster.
The exact scores appear only in charts on the announcement page, not in its text. TechCrunch notes that on Recursive's own average of six tests, only Alibaba's Qwen3.8 27B scored higher. Independent results will matter more than launch charts, so treat the ranking as a claim for now.
"Largely built by AI": what that means
Recursive was founded in 2025 by a team led by Richard Socher, formerly chief scientist at Salesforce. It came out of stealth in May with $650 million in funding at a $4.65 billion valuation, TechCrunch reports, with GV, Nvidia and AMD Ventures among its backers.
The company says a small team of people and agents built Flint from scratch in under four months, using its automated research system, the Eureka Machine. According to TechCrunch, the system contributed to:
- a sparser model shape,
- low-bit training that shrank the model by 71% with almost no loss in quality, and
- custom kernels that speed up computation, including a Metal kernel for Apple's M4 and M5 chips.
What Recursive has not said is how much time the system saved its researchers, or how much human checking went into each step. "Built by AI" here means AI did much of the search and engineering inside a process people designed and supervised.
Where Flint falls short
Recursive is open about the limits. "While Flint is very capable, it may need more guidance than the larger closed-source coding models you're used to," the announcement says. The company recommends breaking big jobs into smaller tasks and giving explicit instructions.
Early hands-on testing is encouraging but thin. TechCrunch cites Wired's Will Knight, who used Flint on a Mac with 48GB of memory to build a to-do list app and a web research tool, and found that it handled his requests every time. That is two projects, not a broad evaluation.
Why a local coding model matters for African developers
Most strong coding assistants are cloud services billed per token or per seat, often in dollars. A model that runs on your own machine changes the cost picture:
- No usage bills. After the download, there are no per-token charges.
- Works offline. Power cuts and patchy connections no longer stop the assistant, as long as the laptop is charged.
- Code stays private. Nothing is sent to a remote server, which helps teams handling client or government code.
The catch is hardware. Flint's published speeds come from recent, high-memory machines, and older laptops will run it more slowly, if at all. The download is also large, so plan it on a good connection.
Flint joins a busy month for open models. Meta's much larger Muse Spark Open arrived last week, but at 816 billion parameters it needs server-class memory. Flint aims at the other end: one developer, one laptop.
How to try it
- Download the weights from Hugging Face (RecursiveSI/flint), or pull it through a tool you already use, such as Ollama or LM Studio.
- On Apple Silicon, consider Recursive's rec engine, which it says is tuned for M4 Pro and M5 chips.
- Start with small, well-defined tasks, and keep prompts under 256,000 tokens.
- Compare its output against your current assistant on real work before switching.
