Recursive says Flint does more than 125 tokens a second on an M4 Pro MacBook Pro and more than 150 on a 64GB NVIDIA RTX Spark at short contexts. It claims Flint leads MoE models under 50B total parameters on SWE-Bench Pro v2, Terminal-Bench 2.1 and SWE-Bench Verified. All of those numbers are Recursive's own.
Context goes up to 512K tokens, but Recursive recommends staying under 256K. The weights are on Hugging Face, and it runs on llama.cpp, Ollama, LM Studio, vLLM and SGLang. Recursive also released rec, its own inference engine tuned for M4 Pro and M5 Pro/Max chips.
The company says a small team plus its automated research system (it calls it the Eureka Machine) built Flint from scratch in under four months, including an odd choice: quantizing the output layer to 4 bits. Recursive's own advice is that Flint needs more guidance than big closed models, so break complex tasks into smaller steps.
On OpenRouter the free endpoint has a 262K context window, supports tool calling and JSON output, and is rate limited (a third-party catalog lists 200 requests a day; I couldn't confirm that on OpenRouter itself). I haven't run it yet.