Unsloth shrinks Recursive’s Flint coding model to a file as small as 11.4 GB

Unsloth has posted GGUF versions of Flint, a 39.8-billion-parameter coding model from Recursive, in sizes from 11.4 GB to 79.6 GB. Only about 2.5 billion parameters are active per token, and the license is Apache 2.0.

Flint was built by Recursive Superintelligence, Inc., which publishes it as RecursiveSI/flint. Unsloth, which the card names as the host of the quantized files, did not train the model. It converted the weights so they can run in local tools such as llama.cpp.

Ten file sizes for one coding model

According to the model card, Flint is a decoder-only mixture-of-experts Transformer with 49 layers, two of them dense. Each MoE layer has 256 routed experts, six of which are selected per token, plus a shared expert that is always active. Thirty-five layers use sliding-window attention with a 513-token window, and 14 use full attention. The context length is 262,144 tokens, shared between prompt and output, and the stated knowledge cutoff is June 2026.

Unsloth lists ten files. The smallest is a 1-bit build, UD-IQ1_M, at 11.4 GB. The 4-bit UD-Q4_K_XL is 25.1 GB, the 8-bit Q8_0 is 42.3 GB, and the unquantized BF16 weights are 79.6 GB. The card shows two charts of size against quantization error and top-1 accuracy, but this article does not reproduce their numbers.

Recursive describes Flint as a model for coding agents: investigating codebases, fixing bugs, implementing features and writing tests. It supports reasoning and multi-turn tool use. The card says training ran through pretraining, mid-training, supervised fine-tuning and reinforcement learning. For the full-precision model, Recursive says one B300 or H200 GPU per replica is enough with the SGLang serving kit.

The recommended sampling settings are temperature 1.0, top-p 0.95 and top-k 20, with an output budget of at least 32,768 tokens for hard problems. The card warns that the model can produce incorrect or unsafe code and tool calls, that long reasoning can use up the output budget, and that results vary with precision and runtime.

Scores that come only from the developer

Recursive reports 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified. The Terminal-Bench run used Terminus 2 and the SWE-bench runs used mini-swe-agent. Agentic scores are averaged over at least five attempts. The card itself notes that evaluation setups differ between sources, and says Flint’s Terminal-Bench score counts solutions retrieved from the web as failures.

Nothing on the card shows these numbers were reproduced by anyone else, and the Unsloth page does not say whether the quantized files score the same as the BF16 original. That matters most for the 1-bit and 2-bit builds, which are the ones that fit on a single consumer machine.

A first model whose claims are still company-reported

Flint is Recursive’s first model, released on October 21, 2026, according to AI Weekly’s summary. The same summary says the company claims the top spot among mixture-of-experts models under 50 billion total parameters on agentic coding tests. It also reports that OpenHands has not independently assessed those benchmark claims. With no earlier Recursive models to compare against, the Hugging Face pages are the only track record there is.

Secondary coverage adds a speed figure. Forkast and other outlets say the company benchmarked more than 125 tokens per second on an M4 Pro MacBook Pro, and that the company says it built the model in under four months. Both figures are Recursive’s own, and neither appears on the Unsloth page.

The pattern of a small active-parameter count paired with aggressive quantization is not unique to Flint. Our recent coverage of Liquid AI’s LFM3-1.7B-Base shows the same push toward models that run on modest hardware, though that model targets fine-tuning rather than coding agents.

Who can run it today

Anyone can download the files now. Apache 2.0 allows commercial use, and the Recursive card says its hosted API terms and data practices are separate from the weight license. The Unsloth page had roughly 149,888 downloads in the last month and about 300 likes when it was read, which signals strong interest but is not evidence of quality.

The 4-bit file at 25.1 GB suits machines with plenty of unified memory or a large GPU. Buyers of smaller hardware would have to accept the lower-bit builds, whose quality loss the card does not state in plain numbers. One more caveat from a secondary listing: Transformers modeling code was reportedly not included at release, so llama.cpp-compatible tools are the practical route for now.