Recursive Flint: A Coding Model That Runs on a Laptop
Recursive Superintelligence shipped its first model on 21 October: Flint is a 40B-class coding MoE that wakes only about 2.5B parameters per token, with an ~80GB BF16 checkpoint under Apache 2.0. Its published evaluations put it at 89.7% on SWE-bench Pro v2, 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1, with quantized builds prepared for NVIDIA and Apple Silicon
Tool Interface
Interactive tool will be available soon
Features
- ✓ A decoder-only MoE with 39.8B total parameters and roughly 2.5B active per token, designed for coding agents: investigate codebases, fix bugs, implement features and write tests
- ✓ 2 dense layers followed by 47 MoE layers, 256 routed experts per layer with 6 selected per token plus an always-on shared expert, and 35 sliding-window layers alongside 14 full-attention layers
- ✓ Documented context of 262,144 tokens including prompt and output, a 150,144-token vocabulary and a stated knowledge cutoff of June 2026
- ✓ Published scores: 89.7% on SWE-bench Pro v2, 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1, with agentic numbers reported as average pass@1 over at least five attempts
- ✓ An ~80GB BF16 checkpoint runs one replica on a single B300 or H200, and the repo ships an SGLang serving kit; separate quantized variants target NVIDIA and Apple Silicon
- ✓ Apache 2.0 weights with no revenue gate, no OEM clause and no service-operator trigger, plus an open-source local engine called rec
How to Use
- Start on Hugging Face and pick the checkpoint: BF16 is about 80GB, so choose BF16 or a quantized build such as NVFP4 based on your VRAM
- For serving, build the image from the sglang/ kit in the repo, pull the weights with hf download and start one replica on a single B300 or H200
- For local use, run the official rec engine; on Apple Silicon budget enough unified memory, and consider quantized builds when you want more throughput
- Read the limitations before wiring it into an agent: structured tool-call grammar does not stop the model from choosing an unsafe or wrong action, so your application has to enforce permissions and execution boundaries
FAQ
Who released Flint and when?
Recursive Superintelligence released it on 21 October 2026 as the company's first model and a real-world test of its automated research system, the Eureka Machine. The model card and weights are at huggingface.co/RecursiveSI/flint and the release note is at recursive.com/updates/introducing-recursive-flint.
What are the size and architecture?
39.8B total parameters with about 2.5B active per token, counting the output head (the active body alone is roughly 2.2B). 49 layers: 2 dense followed by 47 MoE. Each MoE layer has 256 routed experts with 6 selected per token plus an always-active shared expert. 35 sliding-window layers with a 513-token window and 14 full-attention layers, 32 query heads and 4 key/value heads.
What do the published benchmarks say?
The model card lists Terminal-Bench 2.1 at 54.9%, SWE-bench Pro v2 at 89.7% and SWE-bench Verified at 74.2%. Settings were temperature 1.0, top-p 0.95 and top-k 20, with Terminus 2 for Terminal-Bench and mini-swe-agent for SWE-bench; agentic scores are average pass@1 over at least five attempts. Note that Flint's Terminal-Bench score counts solutions that used web retrieval as unsuccessful.
What hardware does it need?
The model card states that Flint runs one replica on a single B300 or H200 GPU, that the BF16 checkpoint is about 80GB, and that the repo includes an SGLang kit. For local machines Recursive points to its rec engine, with quantized variants published separately for NVIDIA and Apple Silicon.
Any licensing catches?
The weights are Apache 2.0, explicitly with no revenue gate, no OEM clause and no service-operator trigger. Hosted API terms and data practices are separate from the weight licence, so check both before you build on it.