Recursive Flint Is Out: 89.7% SWE-bench, Apache 2.0, Free

On October 21, Recursive released Flint — a 40B mixture-of-experts coding model that hits 89.7% on SWE-bench Pro v2 and ships free under Apache 2.0. It runs locally: 131 tokens per second on an M4 Pro MacBook Pro, zero marginal cost per token. The thing worth paying attention to isn’t just the benchmark. Recursive says its automated research system, the Eureka Machine, contributed to nearly every stage of development — architecture design, quantization, custom kernels — and the entire model went from concept to release in under four months with a small team.

What 89.7% on SWE-bench Pro v2 Actually Means

SWE-bench Pro v2 tests whether a model can resolve real GitHub issues pulled from live repositories — not isolated coding puzzles. An 89.7% score puts Flint ahead of most sub-50B models. Among open-weight models under 50 billion parameters, only Alibaba’s Qwen3.8 27B scores higher on SWE-bench Verified (74.2%), though Qwen beats Flint on Terminal-Bench 2.1 and LiveCodeBench v6. These are self-reported numbers from Recursive. Independent verification hasn’t landed yet.

Community testing tells a more nuanced story. Developers running Flint on real tasks found it passed 12 of 18 test runs — solid for a day-one release, but not magic. Weaknesses appeared in counting operations and complex code generation. The practical advice: treat the headline benchmark as a strong signal, run 20-30 real issues from your backlog against Flint before migrating anything to production. The OpenRouter free tier makes that evaluation cost nothing.

Related: SWE-bench Pro Exposes AI Coding: 80% Score Falls to 23% — benchmark methodology context

The Eureka Machine: AI Designed Most of This Recursive Flint Model

Recursive’s Eureka Machine is an internal automated research system that runs the full ML development cycle — proposes architectural ideas, builds and tests them, analyzes results, and plans follow-up experiments without waiting for human direction. For Flint, it designed the sparse MoE architecture, built a quantization recipe that shrunk the model 71% with less than 0.7 points of quality loss, and wrote custom compute kernels for Apple Silicon and Nvidia hardware. Four months, start to finish.

According to TechCrunch’s coverage, Recursive has $650 million in funding at a $4.65 billion valuation, backed by GV, Nvidia, and AMD. The company frames the Eureka Machine as a recursive self-improvement engine — each new model is supposed to make the research system better at the next one. Whether that compounds the way they’re betting is unproven, but the four-month timeline is a data point worth watching. If AI-driven model development keeps this pace, the release cadence for frontier open-source models just changed.

Running Flint Today: Hardware Requirements and Setup

Flint launched with same-day support on Ollama, vLLM, and SGLang. Getting started takes one command:

ollama pull flint
ollama run flint

For more demanding deployments, VRAM requirements depend on quantization level. The 4-bit Q4_K_XL version needs about 27 GB — an RTX 5090 handles it cleanly. Three-bit fits in 16-20 GB. Full BF16 requires a datacenter GPU. One important gotcha: the official GGUF files have a chat template bug that breaks llama.cpp’s Jinja parser. Use the community-provided terelabs template instead. The AtomicAgent deployment guide covers the fix in detail.

No GPU at all? OpenRouter lists Flint at recursive/flint-v1.0:free — rate-limited but zero cost. That’s the fastest path to evaluation without touching hardware, and it’s live now.

Where Flint Fits in Your Stack

Flint’s Apache 2.0 license is the headline for enterprise teams. Local deployment eliminates per-token API costs, which matters when agent loops run thousands of completions per hour. Teams running privacy-sensitive codebases — defense, healthcare, finance — have a top-tier open option that doesn’t require sending code to a cloud endpoint. However, Flint isn’t the strongest model in every benchmark. If your workload is already well-served by a cloud API and you’re not tracking token costs closely, there’s no urgent reason to switch today.

Key Takeaways

• Flint scored 89.7% on SWE-bench Pro v2 — self-reported and impressive, but wait for independent verification before making production decisions
• Apache 2.0 license means free commercial use with no API key; it runs locally at 131 tokens per second on an M4 Pro MacBook Pro
• The Eureka Machine built most of this Recursive Flint model in four months — the fastest public evidence yet of AI-accelerated model development at scale
• Start with OpenRouter’s free tier (recursive/flint-v1.0:free) to evaluate on your codebase before investing in hardware
• Known gotcha: use the terelabs chat template with llama.cpp; the official GGUF template has a Jinja parsing bug