Recursive releases Flint, an open 40B coding model for laptops

Recursive, an AI startup led by CEO Richard Socher, released its first model, Flint, on 21 October 2026. Flint is a coding model published on Hugging Face under the Apache 2.0 licence, which permits free use and modification. TechCrunch reports that the company says Flint was built largely with the help of its automated research system, the Eureka Machine, and that a small team built it from scratch in under four months.

What the model card says

The model card describes a decoder-only mixture-of-experts model with 39.8 billion total parameters and approximately 2.5 billion active per token. It has 49 layers, 256 routed experts per mixture layer with 6 selected per token, and a documented context of 262,144 tokens including generated output. It supports reasoning and multi-turn tool calls through an OpenAI-compatible interface, and the card gives June 2026 as the stated knowledge cutoff.

In its release evaluations, Recursive reports 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified. The card says agentic scores average at least five attempts, that results marked as measured by Recursive differ from developer-reported comparison scores, and that evaluation setups vary across sources. TechCrunch adds that on Recursive's own six-test average only Alibaba's Qwen3.8 27B scored higher than Flint.

The laptop claim needs a footnote

TechCrunch reports that Recursive measured about 131 tokens per second on a MacBook Pro with an Apple M4 Pro chip, against 58 to 69 for similar models on the same machine, and over 150 on an Nvidia RTX Spark. It says the company attributes a 71% size reduction with almost no quality loss to low-bit training.

The repository itself is different. It holds the BF16 checkpoint, about 80 GB to download, and the serving kit it ships is for SGLang on one B300 or H200 GPU per replica. The card says quantized variants for NVIDIA and Apple Silicon runtimes are prepared separately, and warns that results can vary with checkpoint precision and runtime. Recursive's speed and quality figures therefore may not describe the file you download first.

Limits the company states

Recursive tells TechCrunch that Flint needs more guidance than the large paid coding assistants and suggests breaking big tasks into smaller steps. The card recommends at least 32,768 output tokens for hard problems, because long reasoning can exhaust the budget before an answer finishes. It also says tool-using applications must enforce their own permissions and execution boundaries, because a structured tool-call format does not stop the model choosing an unsafe action.

TechCrunch also relays an early hands-on report from Wired's Will Knight, who used Flint on a Mac with 48 GB of memory and found it handled his small programs where other small coding models stumbled; I have not read that review directly. The card notes that hosted API terms and data practices are separate from the weight licence. Cost per task, failure rates on real repositories and independent benchmark runs are not yet published.

What to do

• Run your own repositories and tickets through the specific quantized build and runtime you would deploy, not the headline numbers.
• Keep the permission and sandbox controls you use for any coding agent; the card says the model can make unsafe tool calls.
• Wait for an independent evaluation before changing a vendor or budget decision.