What is Recursive Flint?

Flint is a free coding model from Recursive, released Oct 21, 2026 under Apache 2.0. It's a mixture-of-experts model with 39.8B parameters, about 2.5B active per token, built for coding agents on local hardware.

How much memory does Flint need?

At least 48 GB, per Recursive. The 4-bit builds are 22.6 to 26 GB, and the rest goes to the long context. That means a 48GB Mac or an NVIDIA setup with 48 GB of GPU memory.

Can Flint run on a 24GB RTX 4090?

Not comfortably. The 22.9 GB NVFP4 file barely fits and leaves almost no room for context, so you'd need a short context or partial CPU offload. Recursive recommends 48 GB.

Is Flint free for commercial use?

Yes. The weights are under the Apache 2.0 license, which allows commercial use and changes. Recursive says hosted API terms are separate from the weight license.

Can Flint replace Claude Code or Codex?

Not fully. It has no usage caps and keeps code on your machine, but Recursive says it needs more guidance than big closed models. It works best as a fallback when your $20 or $20 plan hits its limit.

Recursive says more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 on a 64GB NVIDIA RTX Spark at short contexts. Those are vendor numbers, not independent tests.

Recursive Flint: can it replace a paid coding plan?

Updated 2026-10-22. From Recursive's model card and launch post; plan prices from the official pricing pages.

From Recursive's model card and launch post, checked 2026-10-22. Benchmark numbers are Recursive's own, not independent tests.

Question	Answer
Who made it	Recursive, the AI lab led by Richard Socher
Released	Oct 21, 2026, free on Hugging Face
License	Apache 2.0, so commercial use is allowed
Size	39.8B parameters in total, about 2.5B active per token
Context	262,144 tokens documented; Recursive says up to 512K, best within 256K
Memory you need	At least 48 GB (Mac unified memory or NVIDIA GPU memory), per Recursive
Coding scores	74.2% SWE-bench Verified, 89.7% SWE-bench Pro v2, 54.9% Terminal-Bench 2.1 (vendor-reported)
Independent rankings	Not on LMArena's boards yet
Hosted API price	Recursive mentions hosted API terms but lists no per-token price on the pages we checked

What hardware you need to run Flint

Plan on 48 GB of memory. Recursive's INT4 GGUF build is 26.0 GB and "requires an Apple Silicon Mac with at least 48 GB of unified memory". Its NVFP4 build for NVIDIA is 22.9 GB, and Recursive recommends "at least 48 GB of GPU memory, or unified memory on Spark". The MLX INT4 weights are about 22.6 GB.

The weights alone would squeeze onto a 24GB or 32GB card, but the full 262K-token context needs room for the KV cache on top. That's why the official guidance is 48 GB. On a 24GB RTX 4090 you'd have to cut the context a lot or offload part of the model to system RAM, which is much slower.

Works well: a MacBook Pro or Mac mini with an M4 Pro or better and 48 GB or more, a 64GB NVIDIA RTX Spark or DGX Spark, or an RTX PRO 6000. Our 48GB MacBook Pro page shows what else runs on that machine.
Tight: 32GB Macs and single 24GB to 32GB GPUs. The 4-bit weights may load with a short context, but expect trial and error.
No: 16GB laptops and 8GB to 16GB GPUs. Look at smaller models in the local model finder instead.

Our estimate per quant and per machine is on the Flint VRAM requirements page.

How fast is Flint on a laptop?

Fast, by Recursive's numbers. Its launch post says Flint generates "more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 tokens per second on a 64GB NVIDIA RTX Spark at short contexts", up to twice as fast as similar local models. TechCrunch reported 131 tokens per second on the M4 Pro in Recursive's tests. Both are vendor-reported.

The speed comes from the design: only about 2.5B of the 39.8B parameters do work for each token, so the machine reads far less memory per token than a dense 30B model. You still need memory for all 39.8B, which is the catch.

Flint vs Codex and Claude Code on a $20 plan

	Flint (local)	Codex on ChatGPT Plus	Claude Code on Claude Pro
Monthly price	Free (your hardware and power)	$20	$20
Usage limit	None beyond your machine	Estimated 15 to 160 local GPT-6.1 Sol messages per 5 hours; weekly limits may apply; shared with ChatGPT Work	Included; shares one pool with Claude chat; at least 5x Free per 5-hour session
Reset window	n/a	5-hour + weekly	5-hour session + weekly
Code leaves your machine?	No	Yes	Yes
Works offline?	Yes	No	No

Plan prices and limits from OpenAI's Codex pricing page and Anthropic's pricing page via our data, checked 2026-10-22.

On raw quality the paid plans still come out ahead. They run frontier models that sit at the top of LMArena's coding board, and Recursive itself says Flint "may need more guidance than the larger closed-source coding models you're used to" and suggests breaking big tasks into smaller steps.

Where Flint wins is the limit. A heavy Codex or Claude Code day can use up a 5-hour window by lunch, and the weekly cap can run out days before it resets. A local model keeps going. For how those windows work, see Codex limits explained and Claude Code limits explained.

When Flint makes sense, and when it doesn't

Yes, as your fallback: you already own a 48GB Mac or a big NVIDIA box and keep hitting Codex usage limit reached or Claude usage limit reached late in the week. Point your coding agent at Flint until the reset.
Yes, for private code: client work you can't send to a cloud model. Nothing leaves your machine.
Yes, for small, clear tasks: tests, bug fixes and single features with explicit instructions, which is how Recursive suggests using it.
No, as a reason to buy hardware: a new 48GB machine costs far more than a year of a $20 plan ($240). Buy it for other reasons, not to dodge a limit.
No, for big refactors: long, vague, multi-repo jobs are where frontier models still pull ahead.

How to run Flint locally

Get the build for your machine from Recursive on Hugging Face: INT4 (MLX or GGUF) for Apple Silicon, NVFP4 GGUF for NVIDIA.
Load it in a runtime that supports it. Recursive lists llama.cpp, Ollama, LM Studio, Unsloth, vLLM and SGLang, plus its own engine, rec, tuned for M4 Pro and M5 Pro/Max.
Connect a coding agent. The model cards ship ready-made configs for the open-source agents pi and OpenCode, pointing at a local server on your machine.
Use Recursive's settings: temperature 1.0, top-p 0.95, top-k 20, and leave a big output budget (it recommends at least 32,768 tokens for hard problems, reasoning included).

Comparing other open models first? The best local model board and the Qwen3 Coder 30B A3B requirements page are good next stops.

Recursive Flint: can a free local model replace a $20 coding plan?

Recursive Flint: what we know and what we don't