Recursive Flint: Free Laptop Coding Model Built Mostly by AI
Recursive Flint is a free Apache 2.0 coding model with 40B total and 2.5B active parameters, which Recursive says runs over 125 tokens per second on a MacBook.
- Recursive says its automated research system, the Eureka Machine, helped build nearly every part of Flint in under four months.
- The model card reports 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1, all measured by Recursive itself.
- The public repo holds a full-precision checkpoint of about 80 GB; laptop-friendly quantized builds are handled separately.
Recursive Flint is the first model from Richard Socher's self-improving-AI startup, and it is a free, open-weight coding model built to run on your own laptop. Released on October 21 under Apache 2.0, it has 40 billion total parameters but activates only about 2.5 billion per token, which is why Recursive says it hits more than 125 tokens per second on an M4 Pro MacBook Pro.
What actually happened
Recursive came out of stealth in May with $650 million at a $4.65 billion valuation, according to ToolSift's daily briefing, and its stated goal is AI that improves AI. Flint is the first public product of that bet.
On its launch page, Recursive describes Flint as a mixture-of-experts model for "fast agentic coding on local hardware." It claims Flint leads MoE models under 50 billion total parameters on SWE-Bench Pro v2, Terminal-Bench 2.1 and SWE-Bench Verified, and rivals dense models of 25 to 40 billion parameters while running more than five times faster.
Speed is the headline. Recursive says Flint does more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 on a 64 GB NVIDIA RTX Spark at short contexts, up to twice as fast as comparable models. TechCrunch, as relayed by Models at Work, reported a measured figure of about 131 tokens per second against 58 to 69 for similar models.
How Recursive Flint was built
The fun part is the process. Recursive's Eureka Machine is a system that proposes research ideas, implements them, tests them and uses the results to pick the next experiment. The company says it contributed to data, optimizers, training recipes, infrastructure and inference.
A few concrete examples from the launch post: the Eureka Machine searched expert configurations and landed on the 40B total, 2.5B active layout. It quantized the output layer to 4 bits, which the post calls "against common practice," then recovered accuracy with an extra distillation step. The final NVFP4 model scores within 0.7 points of the 16-bit version on Recursive's benchmarks. It also wrote custom kernels for Apple Silicon and NVIDIA RTX Spark.
ToolSift reports that low-bit training shrank the model by 71% with almost no quality loss, and that Flint trails only Qwen3.8 27B on Recursive's six-test average. Context goes up to 512k tokens, with Recursive recommending you stay under 256k.
Humans still steered. Recursive frames Flint as an early public test of AI doing much of the research, not a lab running itself.
The catch
Every benchmark number here comes from Recursive. Models at Work flagged that the card itself says setups vary across sources. It also noted a practical gap: the Hugging Face repo holds only the BF16 checkpoint, about 80 GB, and its serving scripts target a single B300 or H200 GPU. In the card's words, quantized builds for laptops are prepared separately. So the laptop speed claim may not describe the file you download first.
Recursive is also upfront that Flint "may need more guidance than the larger closed-source coding models you're used to." Breaking work into smaller tasks is the suggested workflow. Recursive Flint is a strong local option, not a drop-in replacement for frontier cloud models.
For other fast, cheap models we have covered, see our MiniMax M3.1 Flash piece and our DeepSeek V4.1 Flash breakdown.
What to do as a reader
If you code and care about privacy or API bills, Recursive Flint is worth a test. Start with the official Recursive Flint announcement, which links docs, and check that your hardware matches whichever build you choose. Recursive says it works with llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang, plus its own lightweight engine called rec for M4 Pro and M5 machines.
Run it on your own repo before believing any leaderboard. Self-reported scores are a starting point, and your codebase is the benchmark that actually matters.
If you are here for the bigger AI narrative, the interesting signal is a small team shipping a competitive model fast with heavy automation. Whether that compounds, as Recursive hopes, is still an open question. Ser, that is a story to watch, not a ticker to chase.
Sources: Recursive launch post, Models at Work briefing and ToolSift daily LLM news.
Not financial advice. DYOR, ser.
