Flint

Recursive's 39.8B mixture-of-experts coding model with about 2.5B active parameters, built for fast local agents. Apache 2.0.

In short. Flint is a free coding model for AI agents that fix bugs, write tests and build features. Only a small part of the model works on each word, so it answers fast, and Recursive offers versions that run on a Mac or a single NVIDIA card with 48 GB of memory. It is released under Apache 2.0, which allows commercial use.

Summary

Recursive Superintelligence, Inc. announced Flint on 18 October 2026 and published the weights on Hugging Face under Apache 2.0. It is a mixture-of-experts language model for coding agents with 39.8 billion parameters, of which about 2.5 billion are active per token. Recursive says it trained the model from scratch in under four months with its automated research system, the Eureka Machine. Flint reads and writes text, reasons, and calls tools over several turns. The card lists 49 layers, 256 routed experts with 6 chosen per token plus one shared expert, and an attention mix of 35 sliding-window layers of 513 tokens and 14 full-attention layers. The documented context is 262,144 tokens, the blog says up to 512k and advises staying within 256k, and the stated knowledge cutoff is June 2026.

The BF16 checkpoint is about 79.6 GB, and the SGLang kit in the repository targets one H200 or B300 GPU per replica. For local use there are three 4-bit builds, an MLX INT4 of 22.6 GB for Apple silicon, a Q4_1 GGUF of 26.0 GB for llama.cpp on Metal and an NVFP4 GGUF of 22.9 GB for llama.cpp on NVIDIA GPUs. Recursive asks for at least 48 GB of GPU or unified memory for all of them, and the same figure is the requirement of its own rec engine for M4 Pro, M4 Max, M5 Pro and M5 Max Macs. It says Flint generates more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 on a 64 GB NVIDIA RTX Spark at short contexts. The documentation also covers Ollama, Unsloth, LM Studio, Transformers, vLLM and the OpenCode, Pi, Cline and OpenHands harnesses.

Recursive's release evaluations report 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified, with temperature 1.0, top-p 0.95 and top-k 20. In its chart that is ahead of the Qwen3.5 and Qwen3.6 models in the 33B-to-35B mixture-of-experts group on all three, with 54.9% against 38.7% and 44.9% on Terminal-Bench, but the dense Qwen3.8 in the chart's 27B-to-31B group scores higher on Terminal-Bench (73.0%) and SWE-bench Pro v2 (94.9%). Recursive's argument is speed, since it says Flint runs more than five times faster than dense models of 25 to 40 billion parameters with similar scores. Some comparison scores were measured by Recursive and the rest are developer-reported, and Flint's Terminal-Bench score counts solutions found through web retrieval as failures. In the 30 days to 11 November 2026, Hugging Face counted about 253,000 downloads across the four repositories, most of them the GGUF builds, and the main repository has about 1,400 likes.

Good for

  • Local agentic coding on a Mac with 48 GB or more of unified memory, or on one NVIDIA card with 48 GB
  • Fast decoding, since only about 2.5 billion of the 39.8 billion parameters are active for each token
  • Long coding sessions with a 262,144-token window and tool calls that carry reasoning content
  • Teams that want an Apache 2.0 coding model to self-host and fine-tune, with BF16 weights and three 4-bit builds

Watch out for

  • Even the 4-bit builds are 22 to 26 GB, and Recursive asks for at least 48 GB of GPU or unified memory, while the BF16 checkpoint is 79.6 GB
  • Recursive's own chart has the dense Qwen3.8 ahead on Terminal-Bench 2.1 and SWE-bench Pro v2, so the pitch is speed, not the top score
  • The blog says context up to 512k, but the card and the quantized builds document 262,144 tokens
  • Recursive says complex tasks need explicit guidance and long reasoning can use up the output budget, and the card asks for at least 32,768 output tokens on hard problems
  • The card says Transformers modeling code is not included yet, and all benchmark scores are the developer's

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the Flint (BF16) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Flint (BF16) BF16 39.8B 2.5B active 92 GB 46 GB 48 GB Workstation
Flint INT4 (MLX, Apple silicon) INT4 39.8B 2.5B active 92 GB 46 GB 48 GB Workstation
Flint INT4 GGUF (Q4_1, Metal) INT4 39.8B 2.5B active 92 GB 46 GB 48 GB Workstation
Flint NVFP4 GGUF (NVIDIA CUDA) NVFP4 39.8B 2.5B active 92 GB 46 GB 48 GB Workstation

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software. The developer states a minimum of 48 GB of GPU memory.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Flint?
Flint is a free coding model for AI agents that fix bugs, write tests and build features. Only a small part of the model works on each word, so it answers fast, and Recursive offers versions that run on a Mac or a single NVIDIA card with 48 GB of memory. It is released under Apache 2.0, which allows commercial use.
Can I use Flint commercially?
Yes. Flint is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Flint open source?
Flint is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Flint on my own computer?
Only on a workstation. The smallest variant, Flint (BF16), needs about 48 GB of memory at 4-bit precision, so plan for a 48 to 96 GB GPU, several graphics cards, or a Mac with plenty of memory.
When was Flint released?
Flint was released by Recursive on October 18, 2026.
How long a context window does Flint have?
Flint supports a context window of up to 256K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Qwen3.6 Alibaba Qwen 27.8B – 36B 3B active 256K Apache 2.0 (Permissive) 19 GB Gaming GPU
Ornith 1.5 DeepReinforce 9.65B – 403B 17B active 256K MIT (Permissive) 6.66 GB Laptop or phone
Nemotron 3.5 Lightning NVIDIA 31.6B 3B active 1M OpenMDW-1.1 (Permissive) 22 GB Gaming GPU
Qwen3.8-27B Alibaba Qwen 27.8B 256K Apache 2.0 (Permissive) 19 GB Gaming GPU