Recursive SI's Flint Runs a 40B Coding Model on Your Laptop
Recursive's first model is a 40B mixture-of-experts coder that activates just 2.5B parameters, runs on a MacBook, and was largely designed by an automated research agent.
- Recursive released Flint, a 40B-total, 2.5B-active MoE coding model under Apache 2.0.
- Scores 54.9% Terminal-Bench 2.1, 89.7% SWE-bench Pro v2, 74.2% SWE-bench Verified on company benchmarks.
- Runs at 125+ tok/s on an M4 Pro MacBook, 150+ tok/s on an Nvidia DGX Spark.
- Built in under four months using Recursive's Eureka Machine, an automated AI research loop.
- 256 experts, 6 active plus a shared expert, 262k context, 4-bit output head via distillation.
- Supported in vLLM, SGLang, Ollama, LM Studio, llama.cpp, Unsloth; free hosted trial on opencode.
Recursive Superintelligence, the AI lab led by Richard Socher, released its first model, Flint, on October 21, 2026. The Apache 2.0 release is a coding-focused mixture-of-experts model with 39.8 billion total parameters and about 2.5 billion active for each token. Quantized builds can run on sufficiently equipped laptops, while the full 16-bit checkpoint requires data-center-class memory.
The company says a small group of researchers and automated agents built Flint in under four months. Its in-house research system, the Eureka Machine, reportedly proposed and tested decisions across architecture, data, optimization, infrastructure, and inference. That process makes Flint an early public case study in agent-assisted model development, although most supporting evidence comes from Recursive.
Under the hood: 40B stored, 2.5B active
| Component | Specification |
|---|---|
| Parameters | 39.8 billion total, about 2.5 billion active per token |
| Expert routing | 6 of 256 routed experts, plus one shared expert |
| Layers | 49 total: 2 dense layers followed by 47 MoE layers |
| Attention | 35 sliding-window layers with a 513-token window, interleaved with 14 full-attention layers |
| Heads | 32 query heads and 4 key/value heads |
| Context | 262,144 tokens |
| Knowledge cutoff | June 2026 |
| License | Apache 2.0 |
Flint’s feed-forward layers contain hundreds of specialist subnetworks, but the router selects only six for each token and also invokes a shared expert. The active parameter count drives much of the per-token computation and memory traffic. The total parameter count still determines checkpoint size and how much memory the complete model needs.
Attention alternates between inexpensive local windows and periodic access to the full sequence. The sliding-window layers use rotary position embeddings, while the full-attention layers omit them. Per-head QK RMSNorm stabilizes attention scores, and learned attention sinks give heads a designated place to allocate otherwise diffuse attention. Together, these choices reduce long-context costs while preserving occasional global token access.
Vendor benchmarks await replication
Recursive reports the following agentic coding results and describes Flint as the strongest MoE model below 50 billion total parameters in its comparisons:
| Benchmark | Score | Relevant detail |
|---|---|---|
| Terminal-Bench 2.1 | 54.9% | Web-retrieval solutions counted as failures |
| SWE-bench Pro v2 | 89.7% |
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
