Model release · Recursive

Flint

Model released

Oct 22, 2026 · Date only

Review updated

Oct 22, 2026 10:49 Thu · Beijing Time (UTC+8)

Recursive says its Eureka Machine research system contributed to Flint’s data, training and inference optimization. The company reports scores of 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified. These results depend on the stated evaluation setup and are not general performance guarantees. With Recursive’s optimized inference engine and short contexts, an M4 Pro MacBook Pro exceeds 125 tokens per second. Recursive states a maximum context of 512K and recommends staying within 256K for coding and other tasks. The Hugging Face release is configured for 262,144 tokens shared by input and output; OpenRouter lists 262K. On OpenCode Console, flint-40b-free has no input, output or cached-read charges during a limited-time trial. Access uses a Console account and API; quotas, the end date, any credit-purchase exemption and regional eligibility are unspecified. This is not an unlimited Go offer or a promise of permanent free access. Data from the free endpoint is retained for 30 days and may be used to improve the model, unlike a local deployment that keeps data on the user’s device.