Flint: What's New, Hardware Needed, How to Run It Free

Flint packs 39.8 billion parameters but only activates about 2.5 billion per token thanks to its mixture-of-experts design. It’s built specifically for coding agents that investigate codebases, fix bugs, and write tests.

According to the model card, Flint is a decoder-only mixture-of-experts transformer with a total of 39.8 billion parameters. Only about 2.5 billion of those are active for each token, which includes the output head. The architecture uses 49 layers, two dense layers followed by 47 MoE layers, and supports a context window of 262,144 tokens. It was trained from scratch on a mix of text and code, with supervised fine-tuning and reinforcement learning to handle instruction following and agentic tasks.

The model introduces a large expert pool: each MoE layer has 256 routed experts, and the routing mechanism picks six of them for every token, plus a shared always-active expert. Attention is split between 35 sliding-window layers that look back 512 tokens and the current token, and 14 full-attention layers that see the whole context. This combination aims to keep inference efficient while still providing the capacity needed for complex coding tasks.

In the release evaluations, Flint scored 54.9 percent on Terminal-Bench 2.1, 89.7 percent on SWE-bench Pro v2, and 74.2 percent on SWE-bench Verified. These numbers are reported with temperature 1.0, top-p 0.95 and top-k 20, and they reflect average pass-at-1 over at least five attempts per task.

According to the model card, Flint excels at investigating codebases, fixing bugs, implementing features, and writing tests. However, long reasoning chains can consume the output budget and raise latency, especially when the 256 K token context is approached. The model still can produce incorrect or harmful code, and tool-using applications need to enforce their own safety checks. Performance may also vary with checkpoint precision and the inference runtime used.

Flint is a good fit for hobbyists who have a high-end GPU such as an RTX 4090, RTX 5090, or an Apple M4 Max, where it runs at several hundred tokens per second. Developers looking for a free, open-source model that can act as a coding assistant will find its specialized training useful. Researchers interested in mixture-of-experts behavior can experiment with its 256-expert routing. In all cases, users should plan to review generated code and enforce safety constraints.

What do you need to run it at home? At 4-bit it takes about 27 gigabytes of memory. The smallest home machine that holds it fully is an RTX 5090. On an RTX 3060 12 GB with 32 gigabytes of RAM it writes about 59 tokens per second. On an RTX 4090 with 64 gigabytes of RAM it writes about 172 tokens per second.

You can try it for free through OpenRouter, without a GPU. The weights are open: you can download them from Hugging Face under the Apache 2.0 license. Arena has not ranked it yet: we add the score as soon as the votes arrive.

Compare Flint with every free model, and check what your own GPU can run, on FreeLLMArena.