Recursive Flint: The Free Local AI Coding Model Largely Built by AI Itself

What Is Recursive Flint?

Richard Socher's startup Recursive — backed by $650 million from Google Ventures and Nvidia — released Flint this week: a free, open-weight coding model designed to run entirely on your own hardware. It's a decoder-only mixture-of-experts (MoE) transformer with 39.8 billion total parameters but only ~2.5 billion active per token, which is the trick that makes laptop-class speed possible. The weights are on Hugging Face under Apache 2.0, so there are no API fees, no rate limits, and no code leaving your machine.

Flint is purpose-built for agentic coding workflows: give it a repository task — investigate a bug, implement a feature, write tests — and it reasons through the work, uses tools, and works across multiple turns. It supports up to a 512k-token context window (Recursive recommends staying within 256k for best results), with a documented configuration of 262k tokens.

Speed is the headline: with Recursive's optimized inference stack, Flint generates more than 125 tokens per second on an M4 Pro MacBook Pro and over 150 tokens/sec on a 64GB NVIDIA RTX Spark at short contexts — roughly twice as fast as comparable models in its class, and more than 5× faster than dense 25–40B models with similar coding scores.

Benchmarks: How Good Is It Actually?

Recursive reports that Flint leads all MoE models under 50 billion total parameters on agentic coding benchmarks (publisher-reported, pass@1 averaged over at least five attempts):

• SWE-bench Verified: 74.2% — resolving real GitHub issues, edging Qwen 3.5 (69.2%) and Qwen 3.6 (73.4%)
• SWE-bench Pro v2: 89.7% — far ahead of Qwen 3.5 (68.6%) and Laguna XS 2.1 (38.6%)
• Terminal-Bench 2.1: 54.9% — with solutions obtained through web retrieval counted as unsuccessful, a stricter adjustment than some rivals apply
• LiveCodeBench v6: 80.3% and IFEval: 95.1%

The nuance worth knowing: larger dense models like Qwen 3.8 still beat Flint on some tests (73.0 on Terminal-Bench, 90.3 on LiveCodeBench). Flint's claim is the speed-to-capability ratio, not absolute frontier performance. The NVFP4-quantized version — the one that actually fits on consumer hardware — lands within 0.7 points of the full 16-bit model across benchmarks.

The Eureka Machine: An AI That Builds Models

The back story is the part that made headlines. Recursive says Flint was built from scratch in under four months by a small team of people and agents, with its automated research system — the "Eureka Machine" — contributing to nearly every part of the stack. Eureka automates the full research loop: proposing ideas, implementing them, testing them, and using results to decide what to try next, running many directions in parallel.

Concrete examples from the build:

• Architecture: Eureka searched expert configurations and widths starting from a 35B/3B-active baseline, arriving at the 40B/2.5B-active design as the best speed/performance trade-off it tested.
• Quantization: It quantized the output layer to 4 bits — against common practice — then recovered accuracy with an offline distillation stage using the 16-bit model as teacher, nearly halving training steps versus the baseline.
• Kernels: It generated specialized kernels for Apple Silicon and RTX hardware, including a fused Metal megakernel for the entire decode pass on M4 and M5 chips.

Wired framed it as recursive self-improvement in the wild; skeptics call it sophisticated AutoML with good marketing. Either way, arriving the same week OpenAI disclosed agents tampering with their own safety monitors gives the debate uncomfortable timing.

How to Run Flint on Your Own Machine

Flint ships with broad runtime support — quantized builds are already on Ollama's library (ollama run flint:40b, a ~23GB download), and it works with LM Studio, llama.cpp, Jan, Unsloth, vLLM, SGLang, and Docker Model Runner. Recursive also released rec, its own lightweight inference engine optimized for M4 Pro and M5 Pro/Max Macs.

• On a Mac: the rec engine or LM Studio with the NVFP4 build; 125+ tokens/sec on M4 Pro
• On a server: the BF16 checkpoint runs on one B300 or H200 GPU per replica via the bundled SGLang kit
• Recommended settings: temperature 1.0, top-p 0.95, top-k 20, and at least 32k output tokens for hard problems

If you want an agent harness around it, open CLI tools like Aider pair naturally with a local OpenAI-compatible endpoint, so you can run a fully offline coding agent with no subscription at all.

Local Flint vs Cloud Coding Tools

Flint won't replace Claude Code, Cursor, or GitHub Copilot on hard tasks — Recursive itself says Flint "may need more guidance than the larger closed-source coding models you're used to." But it changes the economics for a specific set of workflows:

What Flint Isn't

Flint is a model, not a product. There's no IDE integration, no hosted plan, no support SLA — you bring the harness. Transformers modeling code wasn't included at release, so the BF16 checkpoint is served through SGLang or vLLM rather than stock Hugging Face pipelines. Benchmark figures are publisher-reported, and evaluation setups vary across labs, so treat cross-model comparisons as directional. And "built by AI" here means AI automated most of the pipeline under human direction — a meaningful milestone for automated AI R&D, not an autonomous lab.

The bigger signal is price pressure: a capable, genuinely fast local coding model at $0 shifts the baseline for what every paid coding tool has to beat, the same week DeepSeek open-sourced its production inference engine. The gap between "good enough and free" and "frontier and expensive" keeps narrowing.

FAQ