Recursive releases Flint, an open-weight coding model built to run agents on a 48 GB laptop

Recursive, the AI startup led by former Salesforce chief scientist Richard Socher, released its first model on Wednesday, October 21. Flint is an open-weight coding model published on Hugging Face under the Apache 2.0 license, and it is designed to run coding agents locally instead of through a paid cloud API.

Two claims come with it. The practical one is that a model small enough for a well-equipped laptop can handle real agentic coding work. The more ambitious one is that Flint was built largely by AI: Recursive told TechCrunch that its automated research system, the Eureka Machine, contributed to nearly every stage of development and let a small team train the model from scratch in under four months. Neither claim has been independently evaluated yet.

What was released

According to the model card, Flint is a mixture-of-experts transformer with 39.8 billion total parameters, of which roughly 2.5 billion are active per token. It has 49 layers: 35 use a 513-token sliding attention window and 14 use full attention. The documented context configuration is 262,144 tokens, shared between the prompt and the generated output, and the stated knowledge cutoff is June 2026. It handles text only, supports reasoning and multi-turn tool calls, and serves an OpenAI-compatible API.

The full BF16 checkpoint is about 80 GB and, per the card, needs one NVIDIA B300 or H200 GPU per replica. The laptop story rests on the quantized versions Recursive published alongside it: an MLX INT4 build of about 22.6 GB, a Q4_1 GGUF of 26 GB for llama.cpp on Apple Silicon, and an NVFP4 GGUF of 22.9 GB for NVIDIA hardware. The Apple Silicon GGUF requires a Mac with at least 48 GB of unified memory, and Recursive recommends at least 48 GB of GPU memory on NVIDIA. That is a high-end laptop, not a typical one. The cards include ready-made configurations for the open-source coding agents pi and OpenCode.

The numbers, and who measured them

Recursive reports 74.2% on SWE-bench Verified, 89.7% on SWE-bench Pro v2 and 54.9% on Terminal-Bench 2.1, averaged over at least five attempts. The card is candid about the caveats: some comparison scores are its own measurements while others are reported by the competing developers, and evaluation setups vary between sources. Recursive also says its Terminal-Bench figure counts solutions found through web retrieval as failures, which other developers may not do.

On speed, Recursive told TechCrunch that Flint generated about 131 tokens per second on a MacBook Pro with an M4 Pro chip, roughly double the 58 to 69 tokens per second it measured for similar local models on the same machine, and more than 150 tokens per second on NVIDIA's RTX Spark. These are vendor figures. Until independent benchmarks run the same quantized checkpoints, treat them as a starting point.

Recursive also told TechCrunch that Flint needs more guidance than the large paid coding assistants and suggested breaking large tasks into smaller steps.

What changes for developers using coding agents

If the numbers hold up, Flint gives teams a credible option for agent work where code cannot leave the machine, where connectivity is unreliable, or where per-token API costs add up across long agent loops. The trade-offs are hardware, setup effort and a model that needs tighter task scoping than frontier systems.

The model card also has concrete operating guidance that matters for anyone wiring Flint into an agent:

• Use the recommended sampling settings: temperature 1.0, top-p 0.95, top-k 20. The server applies them by default when a request omits them.
• Budget at least 32,768 output tokens for hard problems, because reasoning counts against that budget. The GGUF client configs default to a 65,536-token generation budget.
• Preserve the returned reasoning_content field when replaying conversation history, and use the checkpoint's own chat template.
• Treat the 262,144-token context as a capacity limit. The card says it does not guarantee uniform accuracy across lengths, and prompt plus output share it.
• Enforce permissions and execution boundaries in your own tooling. The card notes that structured tool-call grammar does not stop the model from choosing an unsafe or incorrect action.

The self-improvement angle

Recursive emerged from stealth in May with $650 million in funding at a $4.65 billion valuation, and its stated goal is AI that improves itself. According to TechCrunch, the company credits the Eureka Machine with a sparser model shape, low-bit training that shrank Flint by 71% with little quality loss, and custom kernels. Recursive has not said how much researcher time the system saved, and the company's own announcement page could not be opened for this article, so those attributions come from TechCrunch's reporting on what the company said.

For now, the more useful test is the plain one: whether a 2.5B-active model can carry a coding agent through real repositories on a 48 GB machine. The weights are public, so that is something developers can check for themselves.