Cognition's SWE-2 Flash: coding model at a fifth of the price, weights stay locked

What's going on

Cognition, the US company behind the AI coding agent Devin (which also bought the code editor Windsurf in 2025), unveiled SWE-2 Flash on November 4, 2026. The coding model is meant to handle everyday software engineering work quickly and cheaply: automations, code reviews, routine feature work and documentation updates. Cognition says it costs a fifth as much as its bigger sibling SWE-2 and runs three times faster.

SWE-2 Flash is available in Devin Desktop and the command-line version, Devin CLI. It's free until November 30, for users on the Pro, Max and Teams plans according to Startup Fortune. There's no standalone API that developers could use to call the model from their own code. The weights aren't being released for now.

A new American base model

SWE-2 Flash builds on Inkling-Small, a model from Thinking Machines Lab, the company founded by former OpenAI CTO Mira Murati. Inkling-Small came out on July 30, 2026, under the permissive Apache 2.0 license. It's a mixture-of-experts (MoE) model: it consists of many specialized subnetworks, and only a small slice of them runs for each word. Think of a big team where you only ever ask the relevant specialists. Of its 276 billion parameters, just 12 billion are active per token (a chunk of a word).

The full SWE-2, released on September 10, was based on Kimi K3 from Chinese lab Moonshot AI, with 2.8 trillion parameters, 104 billion of them active. The OrcaRouter blog says this switch is the real news. It's Cognition's first coding model that isn't built on a Chinese base model, and Cognition is pitching it as a "fully American coding model." With roughly eight times fewer active parameters, OrcaRouter says the lower price is technically plausible and not just marketing, since running costs depend mostly on active parameters. But OrcaRouter also notes that "fully American" describes a supply chain rather than a competitive edge: anyone can download Inkling-Small.

How it was trained

For the first time, Cognition added a large mid-training phase before reinforcement learning (learning by being rewarded for solved tasks). In mid-training, the base model keeps training on curated data before the actual specialization starts. To find the best data mix, the team built a held-out coding dataset of realistic software scenarios. It then measured how many tokens each candidate mix needed to reach the same validation loss as a baseline run. The result was "a few trillion tokens" drawn from scientific papers, blogs, code and other sources. Training ran in three stages, with the rarest and highest-quality data saved for last.

In the reinforcement learning phase that followed, Cognition says the gains came mostly from better and more varied training tasks, not from changes to the method. The team reviewed instructions, environments and verifiers, the automated checkers that decide whether a task has been solved, and added new task types. According to OrcaRouter, Cognition hasn't published compute figures, exact token counts or ablations separating the contribution of mid-training from reinforcement learning.

Benchmarks: a big jump over the base, just behind SWE-2

All numbers come from Cognition itself. FrontierCode is even an in-house benchmark.

Flash edges out SWE-2 only on FrontierCode 2.0. It comes close on bounded tasks but falls further behind on long, multi-step tasks like DeepSWE.

Cognition says Flash beats every American open-weight model, "even ones three times larger." OrcaRouter checked the math. The strongest of those rivals, Muse Spark Open (616B parameters, 42B active), trails by just 0.1 points on DeepSWE 1.1 at 60.0. On Terminal-Bench 4.0 (20.0 vs. 8.0) and FrontierCode 2.0 (24.0 vs. 6.2), though, the gap is wide. China's DeepSeek V4.1 Flash (763B, 16B active) beats SWE-2 Flash on four of eight benchmarks in the same table, for example 77.0 vs. 60.1 on DeepSWE 1.1 and 42.3 vs. 20.0 on Terminal-Bench 4.0.

The performance claims also deserve a careful read. "A fifth of the price" refers to the cost of a task inside Devin, not a per-token price. The speed figure of 200 vs. 60 tokens per second comes from a simulated replay of a single Terminal-Bench run, not a live measurement. And unlike at the SWE-2 launch, OrcaRouter says Cognition didn't disclose the agent harness, the reasoning effort, or whether results counted the best of several attempts.

Why the weights stay locked

Cognition specifically trained the model for cybersecurity to help defenders. That pushed its ExploitBench score, which measures how well a model can exploit security vulnerabilities, from 11.3 to 34.8 percent. But attackers could use the same skills too, Cognition writes, pointing to a recent wave of cyberattacks. So the weights aren't public. Instead, Cognition is setting up a program to give trusted industry partners and research labs access. They'd be able to fine-tune the model and run it on their own hardware, including on-prem in their own data centers. Details are due "in a few weeks."

OrcaRouter sees a contradiction here. The customer who, by Cognition's own account, needs the weights to run the model in-house doesn't get them, and the stated reason is a capability Cognition trained in itself. On top of that, Flash sits just 1.3 points behind SWE-2 on ExploitBench, and Cognition already sells SWE-2.

Our take

SWE-2 Flash represents a specific bet. According to Startup Fortune, a coding agent makes dozens to hundreds of small model calls per session: read a file, check a test, change a line. Most of those don't need a frontier model. They need a fast, cheap one. So Cognition is building out a lineup of models for different jobs, where what counts is the cost per completed task, not peak benchmark scores.

The limits are clear. Every number is a vendor claim, and none of it has been independently verified. Without an API, a neutral comparison with GPT or Claude models is barely possible. And even though the base model is freely available, SWE-2 Flash isn't an open-weight model for individuals: you can only try it inside Devin. The training is still impressive. With good data and verifiers, a model with 12 billion active parameters gets close to a flagship ten times its size.