Overview
SWE-2 Flash is a coding model from Cognition, the company behind Devin, released on 4 November 2026. It is the smaller sibling of SWE-2: Cognition describes it as one-fifth the price and three times the speed of SWE-2, and built it for the everyday share of agent work — automations, code review, feature development and documentation updates — that smaller models can increasingly handle. In the launch post's generation-speed replay, SWE-2 Flash runs at 200 tokens per second against 60 for SWE-2.
Where SWE-2 was post-trained from Moonshot AI's Kimi K3, SWE-2 Flash starts from Inkling-Small, the 276B-total, 12B-active open-weight model from Thinking Machines Lab, and Cognition calls it a fully American coding model. For the first time in its recipe, Cognition added a large mid-training stage before reinforcement learning: a few trillion tokens of scientific articles, blogs, code and other sources, fed in three stages so the scarcest, highest-quality data came last. The RL stage that followed used more diverse, more carefully audited tasks than SWE-2's. Cognition reports this lifted Terminal-Bench 4.0 from 0.9% for the base model to 20.0%, and FrontierCode 2.0 from 3.2% to 24.0% — slightly above SWE-2's 23.8% on that benchmark.
Cognition also trained for cybersecurity work, raising the model's ExploitBench score from 11.3% to 34.8%. Citing the risk that the same skills help attackers, it did not publish the weights; instead it said it is building a program to share them securely with trusted industry partners and research labs. At launch, SWE-2 Flash is reached through Devin Desktop and the Devin CLI.
| Released | 2026-11-04 |
|---|---|
| License | Proprietary |
| Weights | API only |
| Parameters | 276B total / 12B active (inherited from Inkling-Small) |
| Architecture | Mixture-of-Experts; mid-trained and then reinforcement-learned by Cognition on top of Thinking Machines' Inkling-Small |
| Modalities | Text |
| Status | Available in Devin Desktop and Devin CLI from November 4, 2026; Cognition offered it free through November 30, 2026. Weights are not public. |
Benchmarks
SWE-2 Flash against the field, as published by Cognition at launch
| Benchmark | SWE-2 Flash | Inkling-Small | Beam 1.0 | Muse Spark Open | GLM 5.3 Flash | DeepSeek V4.1 Flash 1017 | GPT-6 Luna | Claude Haiku 5.5 | SWE-2 |
|---|---|---|---|---|---|---|---|---|---|
| FrontierCode 2.0 | 24% | 3.2% | 2.9% | 6.2% | 14.2% | 19.3% | 21.9% | 27.7% | 23.8% |
| FrontierCode 1.1 Main | 45.5% | 21.5% | 22.8% | 24.5% | 31.8% | 43.9% | 42.4% | 46.4% | 50% |
| DeepSWE 1.1 | 60.1% | 17.4% | 43.3% | 60% | 63.4% | 77% | 66.6% | 59.3% | 73% |
| Terminal-Bench 4.0 | 20% | 0.9% | 4% | 8% | 41.8% | 42.3% | 16.4% | 39.2% | 27.3% |
| Terminal-Bench 2.1 | 89.2% | 64.7% | 80.1% | 76.4% | 84.3% | 91% | 82% | 94.4% | 92.8% |
| SWE-Atlas QA | 50% | 27.4% | 33.9% | 54% | 56.5% | 71% | 45.2% | 56.5% | 56.3% |
| SWE-Bench Pro v2 Hard | 89.9% | 56.5% | 76.5% | 82.4% | 88.2% | 86.3% | 80.4% | 90.2% | 90.6% |
| ExploitBench | 34.8% | 11.3% | 19.5% | — | 36% | 31.1% | 44.7% | 38% | 36.1% |
This model's scores
- FrontierCode 2.024%
- FrontierCode 1.1 Main45.5%
- Terminal-Bench 2.189.2%
- Terminal-Bench 4.020%
- DeepSWE 1.160.1%
- SWE-Bench Pro v2 Hard89.9%
- ExploitBench34.8%
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Strengths
- Large gains over its base model: Terminal-Bench 4.0 from 0.9% to 20.0% and FrontierCode 2.0 from 3.2% to 24.0%, per Cognition
- Matches its much larger sibling on FrontierCode 2.0 (24.0% vs 23.8% for SWE-2) at one-fifth the price
- Fast: 200 tokens per second in Cognition's replay, against 60 for SWE-2
- Ahead of Inkling-Small, Beam 1.0, Muse Spark Open, GLM 5.3 Flash and DeepSeek V4.1 Flash on both FrontierCode 2.0 and FrontierCode 1.1 Main in Cognition's table
- Trained for defensive security work: 34.8% on ExploitBench, up from 11.3% for the base model
Best for
- High-volume Devin work where cost and latency matter more than peak capability — automations, code review, documentation updates
- Feature development tasks routed to a cheaper model, keeping SWE-2 for the hardest long-horizon jobs
- Interactive work in Devin Desktop and the Devin CLI, where 200 tokens per second shortens the wait
- Security-focused engineering, given the model's ExploitBench training
How to access
| Provider | Model ID |
|---|---|
| Devin (Cognition) ↗ | — |
FAQ
Is SWE-2 Flash open source?
No. Its base model, Inkling-Small, is open-weight, but Cognition did not publish SWE-2 Flash's weights, citing cyber risk. It said it is building a program to share them securely with trusted industry partners and research labs.
What model is SWE-2 Flash built on?
Inkling-Small from Thinking Machines Lab, a Mixture-of-Experts model with 276B total and 12B active parameters. Cognition added a mid-training stage of a few trillion tokens, then reinforcement learning, on top of it.
How does SWE-2 Flash differ from SWE-2?
SWE-2 is post-trained from the 2.8T-parameter Kimi K3; SWE-2 Flash from the much smaller Inkling-Small. Cognition puts SWE-2 Flash at one-fifth the price and 3x the speed. It edges SWE-2 on FrontierCode 2.0 (24.0% vs 23.8%) but trails it on FrontierCode 1.1 Main (45.5% vs 50.0%), DeepSWE 1.1 (60.1% vs 73.0%) and Terminal-Bench 4.0 (20.0% vs 27.3%).
Where can I use SWE-2 Flash?
In Devin Desktop and the Devin CLI, from launch day. Cognition offered it free through November 30, 2026. The launch post does not mention a standalone API.
What is SWE-2 Flash's context window?
Cognition does not state one in the launch post, so we do not list a figure.