Yuandong Tian's Recursive_SI Releases Flint-40B-A2B: Model Built in Under 4 Months

Recursive_SI, a company Yuandong Tian helped found, has unveiled its first model. Flint-40B-A2B is a fast model for agentic programming, designed to run on local hardware. It uses a mixture-of-experts (MoE) architecture: 40B parameters in total, of which about 2B are active per token. About five months passed between the startup's May announcement and the model's release, and development from scratch took less than four. AI was involved throughout: the model was built with Eureka Machine, the company's internal automated research system.

What the name 40B-A2B means

In MoE models, the A2B suffix is conventionally read as the number of active parameters. For Flint, that means the following. You have to store all 40B weights, but the compute per token is roughly that of a 2B model. For local deployment, this is a convenient combination. The bottleneck on home and office machines is usually memory and its bandwidth, not peak compute. If the model reads only a small fraction of its weights from memory at each step, generation is noticeably faster than with a dense model of comparable size.

Simple arithmetic shows where the limit lies. With 4-bit quantization, 40B parameters take up about 20 GB; with 8-bit, about 40 GB. The first fits on a workstation with a powerful GPU or a laptop with a large pool of unified memory. The second already requires a serious build. For an agent that reads a repository, runs tests, and edits files dozens of times per session, response speed matters more than a few extra benchmark points. That is exactly what Flint is betting on.

A researcher that works around the clock

The most interesting part of the release is how the model was made. The company calls Eureka Machine its own automated research system, and according to Recursive, it took part in developing the model from start to finish. The idea is that part of the routine in a modern ML cycle shifts to AI: choosing experiments, launching and analyzing runs, testing hypotheses. Humans remain the ones who set the tasks and act as arbiters.

For a startup, this is a matter of survival, not an experiment for its own sake. A small team has no army of researchers like the big labs do. If automation compresses the "hypothesis — experiment — conclusion" loop, then four months from a blank slate to a working model looks like a realistic goal. The company's name, Recursive_SI, hints at a more ambitious aim: AI that helps improve AI.

Key release parameters:

• model: Flint-40B-A2B;
• architecture: MoE, about 40B parameters in total and about 2B active per token;
• purpose: agentic programming;
• runtime environment: local hardware;
• development time: less than four months from scratch;
• creation tool: Eureka Machine, Recursive's in-house automated research system.

What to look for when choosing a local coding agent

Local coding models appeal to teams that do not want to send their source code to external APIs, whether because of security requirements, token costs, or simply latency. Flint targets exactly this niche, where every model has specific rivals among open MoE solutions.

The practical test is simple. First, measure real speed on your own hardware, in tokens per second at long context, because agentic workflows quickly inflate the conversation history. Second, run the model not on synthetic tasks but on your own repositories: fixing a bug, refactoring, writing tests. Third, see how well it handles tool chains when there are many steps and one error drags down all the rest.

It is also worth watching what Recursive shows next. If Eureka Machine has really cut the model development cycle to a few months, the startup's next releases should come quickly. In that case, the main product will be not any single model but the system that produces them.