Recursive Flint

// Overview

Flint is the first model from Recursive (Recursive Superintelligence, Inc.), announced on the Recursive blog on October 21, 2026. It is a mixture-of-experts language model for coding agents — investigating codebases, fixing bugs, implementing features and writing tests — with 39.8B total parameters and approximately 2.5B active per token, which Recursive rounds to the size designation 40B-A2.5B. It takes text in and gives text out, supports reasoning and multi-turn, OpenAI-compatible tool calls, and its weights are released under the Apache 2.0 license.

The architecture is built for speed on small machines. Of its 49 layers, the first 2 are dense and the other 47 are MoE layers that each route a token to 6 of 256 experts plus an always-active shared expert. 35 layers use a sliding window of 513 tokens (the current token and the 512 before it) and only 14 use full attention, with RoPE on the sliding-window layers and no positional encoding on the full-attention ones. The model card documents a 262,144-token context shared between prompt and output, a 150,144-token vocabulary and a stated knowledge cutoff of June 2026; Flint was trained from scratch through pretraining, mid-training, supervised fine-tuning and reinforcement learning.

Recursive's model card reports 54.9% on Terminal-Bench 2.1 (with Terminus 2), 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified (with mini-swe-agent), each the average pass@1 over at least five attempts. Its published chart compares the BF16 and NVFP4 checkpoints with sparse MoE models of 33–35B total and 3B active parameters (Qwen 3.5, Qwen 3.6, Laguna XS 2.1), where the BF16 checkpoint leads on five of the six benchmarks shown (Qwen 3.6 edges it on LiveCodeBench v6, 80.4 to 80.3), and with dense 27–31B models (Qwen 3.8, Muse Glimmer, Gemma 4), which still score higher on some tests — Qwen 3.8 posts 73.0 on Terminal-Bench 2.1 and 90.3 on LiveCodeBench v6 against Flint's 54.9 and 80.3. The NVFP4 checkpoint lands within 0.7 points of the 16-bit model across those benchmarks.

Recursive reports more than 125 tokens/s on an M4 Pro MacBook Pro and more than 150 tokens/s on a 64GB NVIDIA RTX Spark at short contexts, and says its Eureka Machine, an automated research system, proposed and tested ideas at most stages of the build. The weights run in vLLM, SGLang and Docker Model Runner, with quantized builds for llama.cpp, Ollama, LM Studio and Jan, and Recursive's own rec engine targets Apple Silicon Macs. The repository ships an SGLang serving kit for one B300 or H200 GPU per replica; Transformers modeling code was not included at release.

// Benchmarks

This model's scores

1. IFEval95.1%
2. SWE-bench Pro v289.7%
3. IFBench89.2%
4. LiveCodeBench v680.3%
5. SWE-bench Verified74.2%
6. Terminal-Bench 2.154.9%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• Ahead of the 33–35B sparse MoE models on Recursive's chart on five of six benchmarks, including 54.9% on Terminal-Bench 2.1 against 44.9% for Qwen 3.6
• Only about 2.5B parameters active per token, with most layers on a 513-token sliding window — Recursive reports 125+ tokens/s on an M4 Pro MacBook Pro
• Strong instruction following for its size: 95.1 on IFEval and 89.2 on IFBench on Recursive's chart
• An NVFP4 checkpoint within 0.7 points of BF16 on the published benchmarks
• Apache 2.0 weights with no usage threshold, plus OpenAI-compatible tool calls and a ready SGLang serving kit

// Best for

• Running a private, offline coding agent on a Mac or a 64GB RTX Spark through Ollama, LM Studio, llama.cpp or Recursive's rec engine
• Bug fixing, feature work and test writing inside agent harnesses that speak the OpenAI tool-call format
• Self-hosted coding-agent serving on one B300 or H200 GPU with the bundled SGLang kit
• Fine-tuning or studying a permissively licensed sparse MoE coding model