TL;DR: Recursive Flint is the first model from Richard Socher’s AI startup Recursive: a free, open-weight coding model with 40 billion total parameters that activates only about 2.5 billion per token, so it can run on a laptop. Recursive says its automated research system, the Eureka Machine, helped build nearly every part of it in under four months, and that Flint tops other mixture-of-experts models under 50 billion parameters on agentic coding benchmarks. The weights are on Hugging Face under Apache 2.0. The benchmark and speed figures are Recursive’s own.
Recursive Flint landed on Wednesday, October 21, and it is interesting for two reasons at once. It is a practical tool, a coding model small and fast enough to use offline on a MacBook. And it is a public test of a much bigger idea: that AI systems can do a large share of the research needed to build the next AI system. Here is what Recursive announced, what the numbers actually say, and who should try it.
What is Recursive Flint?
According to Recursive’s launch post, Flint is “a mixture-of-experts (MoE) model built for fast agentic coding on local hardware.” The Hugging Face model card describes it as a model for coding agents that is designed to investigate codebases, fix bugs, implement features and write tests, with support for reasoning and multi-turn tool use.
Recursive emerged from stealth in May. TechCrunch reports the company raised $650 million at a $4.65 billion valuation from backers including GV, Nvidia and AMD Ventures. It is led by CEO Richard Socher, Salesforce’s former chief scientist and the founder of You.com.
Key specs at a glance
| Spec | Recursive Flint |
|---|---|
| Total parameters | 39.8 billion (rounded to “40B”) |
| Active parameters per token | About 2.5 billion |
| Architecture | Decoder-only MoE Transformer, 49 layers, 256 routed experts (6 picked per token) plus a shared expert |
| Context | 262,144 tokens documented; Recursive says up to 512K is supported but recommends staying within 256K |
| Knowledge cutoff | June 2026 (stated) |
| License | Apache 2.0 |
| Runs with | llama.cpp, Ollama, Unsloth, LM Studio, vLLM, SGLang and Recursive’s own “rec” engine |
If terms like “tokens” and “context” are fuzzy, our explainers on what tokens are in AI and what a context window is cover the basics.
Why the 2.5 billion active parameters matter
A mixture-of-experts model splits its network into many smaller “experts” and routes each token through only a few of them. Flint stores 40 billion parameters but uses roughly 2.5 billion for any given token. That keeps the compute per token low, which is what makes it fast on consumer hardware.
Recursive says its optimized inference engine pushes Flint past 125 tokens per second on an M4 Pro MacBook Pro and past 150 tokens per second on a 64GB NVIDIA RTX Spark at short contexts, which it describes as up to twice as fast as comparable models. TechCrunch, citing Recursive’s tests, puts the MacBook figure at about 131 tokens per second, versus 58 to 69 for similar models such as Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning on the same machine.
The download is not tiny. The Ollama listing shows a 23GB quantized build with a 256K context, so a laptop with plenty of unified memory is the realistic target. Compare that with Meta’s Muse Spark Open, an 816-billion-parameter open model that TechCrunch notes needs at least a terabyte of memory.
Benchmarks: what Recursive claims
On its model card, Recursive reports these scores from its own release evaluations:
- Terminal-Bench 2.1: 54.9%
- SWE-bench Pro v2: 89.7%
- SWE-bench Verified: 74.2%
The company says Flint is the best-performing MoE model with fewer than 50 billion total parameters on agentic coding benchmarks, and that it rivals dense models of 25 to 40 billion parameters while running more than five times faster. TechCrunch adds that on Recursive’s six-test average only one model, Alibaba’s Qwen3.8 27B, scored higher.
Treat these as publisher-reported numbers. Ollama’s listing makes the same point, noting that evaluation setups and runtimes can affect scores. Independent testing will matter more than launch charts, as we saw when Cognition’s coding model won some benchmarks and lost others.
The “built by AI” part: the Eureka Machine
Recursive’s stated mission is recursively self-improving AI. Its Eureka Machine is a system that proposes research ideas, implements them, tests them and uses the results to decide what to try next. Recursive says it contributed to “nearly every part” of Flint’s stack, and gives three concrete examples:
- Architecture search. Starting from a 35B-total, 3B-active baseline, the system explored expert counts and widths and landed on the 40B/2.5B shape.
- Quantization. Against common practice, it quantized the output layer to 4 bits, then added an offline distillation step to recover accuracy. Recursive says the final NVFP4 model scores within 0.7 points of the 16-bit version, and TechCrunch reports the low-bit work shrank Flint by 71%.
- Kernels. It wrote specialized kernels for Apple Silicon and RTX Spark, including a Metal “megakernel” that fuses the whole decode pass on M4 and M5 chips.
Humans still oversaw the work. TechCrunch characterizes it as an AI research assistant supervised by people, and Recursive did not say how much researcher time the system saved. It fits a wider trend: we recently covered Anthropic saying Claude now leads a quarter of the work building the next Claude.
Who should try Recursive Flint?
Flint makes the most sense for developers who:
- Want a coding assistant that works offline and keeps source code on their own machine.
- Are tired of per-token API fees for routine tasks like writing tests or small refactors.
- Already run local models through Ollama or LM Studio and have a Mac or PC with enough memory.
- Build products and want permissively licensed weights they can adapt or fine-tune.
Recursive is candid about limits. It says Flint “may need more guidance than the larger closed-source coding models you’re used to” and recommends breaking complex work into smaller subtasks with explicit instructions. In other words, it is a fast daily driver, not a drop-in replacement for the biggest cloud agents.
How to run it
The quickest route is ollama run flint. Teams with data-center GPUs can use the BF16 checkpoint with SGLang or vLLM; the vLLM project says it supported Flint from day zero, with both checkpoints running on a single GPU. On Apple Silicon, Recursive’s own rec engine targets M4 Pro and M5 Pro/Max machines.
The bottom line
Recursive Flint is a credible new option in the fast-growing world of local AI: an Apache-licensed coding model that, by its maker’s numbers, punches above its weight and runs at laptop-friendly speeds. The bigger story is how it was made. If Recursive’s Eureka Machine really can compress months of model research, each new release should arrive faster. For now, test Flint on your own repos, keep an eye out for independent benchmarks, and see our coverage of Perplexity Computer running locally on Mac for another sign that serious AI is moving onto personal hardware.
The Teck Hustlers Editorial Team covers technology, business, finance and gaming news for readers who want the story behind the headline. Every article is researched, written and edited by our in-house team, checked against company announcements, filings and reputable reporting, and updated when new facts emerge. We focus on what each development means for founders, professionals, investors and everyday users. Have a tip or a correction? Reach us through our contact page.
