Cognition · 2026-07-08 · major
SWE-1.7 — Cognition's coding model runs on Devin at 1000 tok/s via Cerebras
Cognition released SWE-1.7, a coding model trained from Kimi K2.7 with reinforcement learning across four datacenters. It scores 81.5% on Terminal-Bench 2.1 and 77.8% on SWE-Bench Multilingual, and runs on Devin at 1000 tok/s via Cerebras.
SWE-1.7 is Cognition's new coding model — near GPT-5.5 on agentic coding, running at 1000 tokens per second on Devin.
Quick facts
| Maker | Cognition |
|---|---|
| Base model | Kimi K2.7 (post-RL) |
| Training | RL across four datacenters on three continents, with Fireworks inference |
| Serving | Cerebras at ~1000 tokens per second |
| Availability | Devin (Web, Desktop, CLI) |
| FrontierCode 1.1 Main | 42.3% (Opus 4.8 46.5%, GPT-5.5 43.0%) |
| Terminal-Bench 2.1 | 81.5% (Opus 4.8 86.9%, GPT-5.5 84.2%) |
| SWE-Bench Multilingual | 77.8% (Opus 4.8 84.4%, GPT-5.5 76.8%) |
Benchmarks
| Claude Opus 4.8 | 46.5% | |
|---|---|---|
| GPT-5.5 | 43% | |
| SWE-1.7 | 42.3% | |
| Claude Opus 4.7 | 38.5% | |
| Kimi K2.7 Code | 30.1% | |
| Composer 2.5 | 25.6% | |
| GLM-5.2 | 24.5% | |
| SWE-1.6 | 9.4% |
| Claude Opus 4.8 | 86.9% | |
|---|---|---|
| GPT-5.5 | 84.2% | |
| Claude Opus 4.7 | 83% | |
| SWE-1.7 | 81.5% | |
| GLM-5.2 | 81% | |
| Composer 2.5 | 76% | |
| Kimi K2.7 Code | 72.7% | |
| SWE-1.6 | 39.7% |
| Claude Opus 4.8 | 84.4% | |
|---|---|---|
| Claude Opus 4.7 | 80.5% | |
| SWE-1.7 | 77.8% | |
| GPT-5.5 | 76.8% | |
| GLM-5.2 | 74.5% | |
| Kimi K2.7 Code | 73.5% | |
| Composer 2.5 | 71.6% |
What is it?
SWE-1.7 is a coding-focused frontier-tier model that Cognition trained from a Kimi K2.7 base using its own reinforcement-learning pipeline. The model becomes Devin's default engine on web, desktop, and CLI, and serves through Cerebras at roughly 1000 tokens per second — fast enough that a long agent session finishes in a fraction of the wall-clock a GPU-served frontier model needs.
How does it work?
Four training changes drove the jump from SWE-1.6: top-p sampling to prevent entropy collapse, multi-cluster RL training across four datacenters on three continents with compressed weight deltas that sync 1T-parameter updates cross-continent in 1–2 minutes, tighter data curation, and self-compaction that keeps agent trajectories coherent over long horizons. Cognition also observed condensed chain-of-thought and more thorough codebase exploration as emergent behaviors from the training setup.
Why does it matter?
Cognition needed a model that could match GPT-5.5 and Anthropic Opus on agentic coding at Devin's price point, and SWE-1.7 gets close on their published numbers — 42.3% on FrontierCode 1.1 Main, 81.5% on Terminal-Bench 2.1, 77.8% on SWE-Bench Multilingual — while running fast enough that a Devin agent no longer stalls waiting for tokens. It also validates that RL from an open base model can still push capability meaningfully beyond a strong post-RL starting point.
Who is it for?
Devin subscribers, coding-agent teams comparing frontier-tier models on cost and latency.
Frequently asked questions
- What is SWE-1.7 built from?
- Cognition trained SWE-1.7 from a Kimi K2.7 base — an already RL-post-trained open model — and then applied its own reinforcement-learning pipeline. Four training changes drove the gain: top-p entropy preservation, multi-cluster distributed training with compressed weight deltas, higher-quality data curation, and self-compaction for long agent trajectories.
- How does SWE-1.7 compare to GPT-5.5 and Claude Opus on coding?
- SWE-1.7 posts 42.3% on FrontierCode 1.1 Main, 81.5% on Terminal-Bench 2.1, and 77.8% on SWE-Bench Multilingual. Those numbers sit near GPT-5.5 (43.0 / 84.2 / 76.8) and below Claude Opus 4.8 (46.5 / 86.9 / 84.4) on Cognition's published table, and well above SWE-1.6 (9.4 / 39.7 on the two benchmarks reported).
- Where does SWE-1.7 run?
- SWE-1.7 is available today inside Devin on web, desktop, and the CLI, served through Cerebras at about 1000 tokens per second. Cognition uses SWE-1.7 as Devin's default coding engine, so existing Devin subscribers pick it up without configuration.
- Is SWE-1.7 open weights?
- SWE-1.7 is hosted by Cognition rather than released as open weights. Access is via a Devin subscription on cognition.com — there is no separate API endpoint or Hugging Face checkpoint listed in the launch post.
Try it
Open Devin on web, desktop, or CLI — SWE-1.7 is the default coding model.