Recursive Flint: Richard Socher's New Lab Ships 40B Apache 2.0 Coder Claiming 74.2% SWE-bench Verified
Recursive Superintelligence — the AI lab founded by Richard Socher, former chief scientist at Salesforce — released its first model on October 22, 2026. The model is called Flint. It is a mixture-of-experts coding model with 39.8 billion total parameters, 2.5 billion active per token, and an Apache 2.0 license. Weights are on Hugging Face. OpenRouter is serving it free at launch.
Architecture
Flint uses a decoder-only MoE design with 256 routed experts per layer, 6 selected per token, plus a shared expert. Context window is 262,144 tokens. The knowledge cutoff is June 2026. The full BF16 checkpoint is approximately 80GB; quantized builds go down to 11.4GB and score within 0.7 points of the 16-bit model on published benchmarks, per Recursive.
Inference speed: Recursive reports approximately 125 tokens per second on an M4 Pro MacBook Pro and 150 tokens per second on a 64GB NVIDIA RTX Spark at short contexts, using its own optimized inference engine. Cline, a popular coding IDE extension, reports about 24GB memory requirement for its test setup, though Recursive has not published an official figure.
Benchmark Claims
All figures are self-reported by Recursive and have not been independently reproduced.
| Benchmark | Score | Harness |
|---|---|---|
| SWE-bench Verified | 74.2% | mini-swe-agent, no internet |
| SWE-bench Pro v2 | 89.7% | mini-swe-agent, no internet |
| Terminal-Bench 2.1 | 54.9% | Terminus 2, retrieval counted as failure |
The SWE-bench Pro v2 score at 89.7% is high relative to the other two numbers. Recursive notes that web retrieval was counted as a failure in all runs, and scores are averages over at least five attempts. The model card warns that other developers’ setups may differ, making cross-model comparisons unreliable.
For reference, independently verified SWE-bench Verified scores on established models: Kimi K3 93.4%, Claude Fable 5 95.0%, DeepSeek-V4.1-Flash around 79% from DeepSeek’s own report. Flint at 74.2% self-reported would place it roughly between GPT-5.1 and GPT-5.4 on the independently tracked leaderboard — plausible for a 40B model, though confirmation is pending.
What Recursive Says Built This
Socher’s team says most of Flint’s development was automated by an internal system they call the Eureka Machine, described as a self-improvement loop. The claim is that the model iteratively generated and refined its own training data. There is no further technical detail in the model card.
The Laptop Claim
Recursive’s marketing leads with laptop deployment. The practical picture is narrower: the main release artifact is an 80GB BF16 checkpoint requiring a high-end workstation or data-center GPU. Quantized builds targeting laptop use are listed as separate artifacts prepared after the BF16 release. The 125 tokens/sec figure from the model card likely reflects the quantized path, not the BF16 weights.
That caveat aside, a 40B-total, 2.5B-active MoE at Apache 2.0 with verified quantizations is a legitimate entry point for local agentic coding if the benchmark numbers hold. The model is also available via OpenCode and Pi in addition to OpenRouter.
Competitive Position
At 74.2% SWE-bench Verified, Flint competes with models like Ornith-1.0 (82.4%), DeepReinforce’s earlier open release, and the Qwen3.8-27B line (76.8% claimed on vendor card). The 2.5B active-parameter design makes it significantly cheaper to run than dense models at similar total size. Whether the self-reported scores hold on third-party harnesses will determine how it ranks in practice.
Independent evaluations are expected on swebench.com and vals.ai in the coming days.