Recursive Flint is an open-weight coding model from Recursive. Its weights are published under Apache 2.0, while the model card’s clearest hardware reference applies to a specific BF16 checkpoint and SGLang setup: about 80 GB to download and one B300 or H200 GPU per replica. Other local runtimes are available, but the documented configurations differ by checkpoint and software.
What Recursive publishes about Flint
Flint is a decoder-only mixture-of-experts (MoE) model, an architecture that routes each token through selected parts of a larger network. It has 39.8 billion parameters in total and about 2.5 billion active per token, including the output head. Its model card and files on Hugging Face specify the architecture and deployment details:
| Attribute | Published detail | Scope |
| Architecture | 49 layers: 2 dense and 47 MoE | Decoder-only Transformer |
| Parameters | 39.8 billion total; about 2.5 billion active per token | Active count includes the output head |
| Attention and experts | 35 sliding-window layers with a 513-token window, and 14 full-attention layers; 256 routed experts per MoE layer, with 6 selected per token plus a shared expert | Model-card architecture |
| License | Apache 2.0 | Applies to the published weights |
| BF16 checkpoint | About 80 GB | Download size |
| SGLang setup | One B300 or H200 GPU per replica | The model card’s listed serving setup |
Recursive also lists llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang for local use. Its rec engine is optimized for Flint on M4 Pro and M5 Pro/Max. Quantized variants are prepared separately for NVIDIA and Apple Silicon runtimes, so the BF16 file size and SGLang hardware configuration are not descriptions of every local route. Recursive’s license for the weights also does not set the terms for hosted APIs.
Benchmark scores and context limits
Recursive’s model card reports 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified. These are the developer’s evaluation results. The listed methodology uses Terminus 2 for Terminal-Bench and mini-swe-agent for SWE-bench; the average pass@1 results use at least five attempts, and Terminal-Bench penalizes solutions that rely on web retrieval.
The model card configures a context of 262,144 tokens, counting both the prompt and generated output. Recursive’s announcement separately describes context windows of up to 512K tokens and recommends staying within 256K. These figures are published in different contexts.
Recursive also reports more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 tokens per second on a 64 GB NVIDIA RTX Spark at short context lengths. Those are company-reported results for its optimized engine and specified hardware.
What a local coding test found
In a separate test, AICodeKing ran Flint with llama.cpp on an M4 Max and reported about 110 tokens per second at short context, 52 at 8K, 32 at 32K and 9 at 115K. The test used a llama.cpp pull-request branch, and the creator reported that Flint’s speed fell as context grew. AICodeKing also reported that a link-shortener API task passed 14 of 14 test cases. Those results belong to that creator’s setup and task, not a standardized benchmark.
How the Eureka Machine contributed
Recursive says its Eureka Machine contributed to almost every part of Flint’s development. People set research goals, reviewed results and retained the ability to stop the process. Flint’s development involved AI assistance under human direction.
