5 November 2026 · 3 min read
Cognition Launches SWE-2 Flash, a Faster and Cheaper Coding Model for Routine Agent Work
Cognition released SWE-2 Flash on 4 November, a 276B-parameter mixture-of-experts model with 12B active parameters that runs 3x faster and costs one-fifth as much as SWE-2 inside Devin. It is free through 30 November.
Cognition shipped SWE-2 Flash on 4 November 2026, positioning it as a lower-cost, higher-speed alternative to its SWE-2 coding model. The model is available immediately in Devin Desktop and Devin CLI, and it is free for Pro, Max and Teams users through 30 November [1][2].
What SWE-2 Flash brings
SWE-2 Flash is a mixture-of-experts model with 276 billion total parameters, of which only 12 billion are active at inference time. That architecture matches Thinking Machines Lab’s Inkling-Small, which Cognition used as its base [1][3]. The company says Flash runs three times faster than SWE-2 and costs one-fifth as much inside Devin [1][2].
The speed figure comes from a simulated replay: each model’s run on the Terminal-Bench 4.0 gsea-proteomics task was replayed at 200 tokens per second for Flash and 60 tokens per second for SWE-2, with transcripts shortened to equal length [2][4]. Cognition frames the cost saving as a product-level price, not a per-token rate, so it cannot be directly compared with standalone API pricing [4].
Flash is aimed at routine software-agent tasks such as automations, code review, feature development, and documentation updates [1][2][3]. It is not intended to replace SWE-2 on every benchmark, and the numbers bear that out. On FrontierCode 2.0, Flash scores 24.0% against SWE-2’s 23.8%, but it trails on Terminal-Bench 4.0 (20.0% vs 27.3%) and DeepSWE 1.1 (60.1% vs 73.0%) [2][3].
Training recipe and base-model shift
For the first time, Cognition added a large-scale mid-training stage before reinforcement learning. The team curated a dataset of a few trillion tokens from scientific articles, blogs, code, and other sources, and used token-efficiency experiments to select the most effective mixtures. Training proceeded in three stages so that the highest-quality data appeared later [2][3].
Reinforcement learning followed, with improvements attributed mainly to higher-quality and more diverse tasks rather than recipe changes. Cognition audited task instructions, environment setups, and verifiers, and added new task types to build broader reasoning skills [2][3].
The base model itself marks a departure. SWE-2 was post-trained from Moonshot AI’s Kimi K3, a 2.8-trillion-parameter model with 104 billion active parameters. SWE-2 Flash is instead built on Inkling-Small, an Apache 2.0 model released by Thinking Machines Lab in July 2026 [3][4]. That swap to a smaller, American-origin base is what makes the cost and speed claims mechanically plausible [4].
Weight access and cybersecurity
Cognition is not releasing the weights publicly. The company says it is building a secure sharing programme for trusted industry partners and research labs, with details promised in the coming weeks [1][2][3]. The stated reason is cybersecurity risk: Cognition deliberately trained for security work, raising the model’s ExploitBench score from 11.3% to 34.8%, but notes that the same capabilities could be misused by attackers [2][3]. SWE-2 Flash’s ExploitBench score is 34.8%, close to SWE-2’s 36.1% [2][4].
There is no standalone API for Flash. Developers who want to use it must do so inside Devin’s desktop or CLI tools [1][4].
What it means for developers
For teams already using Devin, SWE-2 Flash offers a cheaper way to handle the high volume of small model calls that agentic coding generates—reading files, checking tests, patching lines. Cognition is betting that as coding agents mature, cost per completed task will matter more than peak benchmark scores [1].
The restricted weight access and lack of an API limit how far developers outside Devin can evaluate or integrate the model. The planned partner programme may eventually open fine-tuning and on-premise deployment, but for now the model remains locked to Cognition’s own products [1][3][4]. The base-model change to an Apache 2.0 foundation could matter for supply-chain transparency, even though the post-trained weights are not open [4].
Sources
- Cognition launches SWE-2 Flash to make AI coding agents cheap and fast - Startup Fortune
- Introducing SWE-2 Flash
- Cognition Launches SWE-2 Flash for Faster, Lower-Cost Coding Agents
- Cognition SWE-2 Flash: Cheaper, Faster, Different Base
- cognition
- swe-2-flash
- ai-coding
- devin
- mixture-of-experts
Researched and drafted with AI from the sources above.
Learn to build with AI
Hands-on tutorials on Cursor, Claude Code, AI agents and more.