Flint

Name: Flint
Author: Recursive

Recursive · Flint · released 2026-10-21 · data checked 2026-11-04

Flint is a 39.8B-parameter Mixture-of-Experts model (2.5B active per token) from Recursive (United States), released 2026-10-21 under the Apache-2.0 license. It doesn't have enough reported benchmarks for an index score yet. It needs about 25.7 GB of memory at 4-bit with an 8K-token context.

Trained from scratch for fast agentic coding on local hardware (Apple Silicon, RTX). The card also reports 89.7 on SWE-bench Pro v2; its comparison with other models is only an image, so only two tracked benchmarks are listed and there is no index score yet. A free preview runs on OpenRouter.

Small LM Index

—

Needs 2 core benchmarks (has 1)

Parameters

39.8B

2.5B active · MoE

Price per 1M tokens

—

No hosted API on OpenRouter; self-host

Context window

262K

tokens

Benchmarks

Each bar shows this model's score. The tick marks the best medium model on that benchmark. Sources: 2 reported by Recursive. How we label sources.

SWE-bench Verified Lab-reported Coding · mini-swe-agent, mean pass@1 over 5+ attempts74.2

Class best: Qwen3.6 27B 77.2

Terminal-Bench Lab-reported Agentic · Terminal-Bench 2.1 (Terminus 2)54.9

Class best: Nanbeige4.3 33B 80.9

Specifications

Lab: Recursive (US · United States)
Architecture: Mixture-of-Experts
Total parameters: 39.8B
Active parameters: 2.5B
Size class: Medium
Context window: 262K tokens
Input modalities: text
Reasoning: Reasoning (thinking)
License: Apache-2.0
Openness: Open weights
Memory to run: ~25.7 GB at 4-bit, 8K context (more sizes)
Hugging Face: 35.5K downloads (30 days) · 1.3K likes

Weights on Hugging Face ↗

Compare with

Flint vs Seed-OSS 36B Instruct (36.2B, index 61.5)
Flint vs Hermes 4.3 36B (36.2B, index 53.2)
Flint vs MiniCPM-V 4.7 35B-A3B (35.2B, index 56.7)

Sources

https://huggingface.co/RecursiveSI/flint

Something wrong here? Report it.

Run it locally

Uses ggml-org/Flint-GGUF (Q4_K_M). Swap the tag for another size: BF16, Q4_K_M, Q8_0.

Ollama

ollama run hf.co/ggml-org/Flint-GGUF:Q4_K_M

llama.cpp

llama-server -hf ggml-org/Flint-GGUF:Q4_K_M

Then open http://localhost:8080 for a chat window.

MLX (Mac)

pip install -U mlx-lm
mlx_lm.chat --model mlx-community/flint-OptiQ-4bit

vLLM (GPU server)

vllm serve RecursiveSI/flint

Full-precision weights, for a server GPU.

Open in LM Studio

Memory needed: weights, context cache and runtime buffers.

Quantization	4K context	32K context	256K context
16-bit F16	81 GB	82 GB	88 GB
8-bit Q8_0	44 GB	45 GB	51 GB
4-bit Q4_K_M	26 GB	26 GB	33 GB

Check it against your GPU, Mac or phone →