Dedicated Coding Model
Drop the general-purpose model,
and get 10x faster.
Coding agent Devin has switched to a purpose-tuned model called SWE-1.7. Running on Cerebras chips, it hits roughly 1,000 tokens per second — a speed tier general-purpose models couldn't reach — and Pro plan users get it at no extra cost.
Built on Kimi K2.7,
with another layer of RL
Not built from scratch — an already RL-trained base model got a further, coding-specific round of reinforcement learning stacked on top.
Cognition, the company behind Devin, announced SWE-1.7 on its blog in July 2026. It's built on Moonshot AI's Kimi K2.7 Code, a base model that had already undergone large-scale reinforcement learning (RL). Cognition layered on additional RL tuned for agentic software engineering, along with more stable long RL runs, higher-quality training data, and a multi-cluster rollout infrastructure. It also adopted a technique called "self-compaction," which lets the agent summarize its working state to keep going on long tasks that exceed the raw context window.
On benchmarks, it trails Claude Opus 4.8 by a small margin, but beats GPT-5.5 on SWE-Bench Multilingual and comes within roughly one point on FrontierCode 1.1 Main, according to VKTR. It doesn't fully match frontier models across the board, but it's designed to sit favorably on the cost-performance Pareto curve.
Why a dedicated model
matters right now
The more agentic work scales up, the more the weight of "speed per response" matters.
General-purpose LLMs have long forced an awkward compromise between speed, cost, and accuracy. Dedicated tuning is a hard investment to justify — which is exactly why it kept getting pushed down the roadmap. SWE-1.7 breaks that compromise by moving off GPUs entirely and onto Cerebras's dedicated inference silicon. In long-running agentic tasks that call tools hundreds of times, the per-response speed difference translates directly into how long a task feels.
Who benefits, and how
Engineers get the most direct benefit. Pro plan users get this speed at no extra cost, which should meaningfully cut the wait-time friction of large refactors or long autonomous runs. Casual, one-off users, on the other hand, will barely notice — the speed gain only matters when you're running an agent for long, high-frequency stretches.
For anyone evaluating vendors, "cost per unit of speed" is now a real axis alongside Claude Code and GitHub Copilot's MAI-Code. That the model is built on a Kimi-family open base is also one more data point for how fast open base models are becoming production-ready.
Speed doesn't replace
accuracy — it shortens
the distance to the right answer.
What to do next
Try it on long-running tasks first
Large dependency updates or sweeping refactors — tasks where wait time used to be the bottleneck — are where the difference is most noticeable.
Don't overweight a small benchmark gap
The gap to Opus 4.8 is reported as small, but real-world impact varies by task type. Validate on your own codebase at small scale first.
Compare vendors on cost-per-speed, not accuracy alone
Weigh Claude Code and GitHub Copilot's dedicated models on all three axes — speed, cost, and accuracy — not accuracy in isolation.
It really is "slightly behind" on accuracy
SWE-1.7 currently trails Opus 4.8 by a small margin on benchmarks, so calling it "on par with the frontier" would be premature. How much accuracy you're willing to trade for speed depends on how much is riding on the task. There's also a structural risk in the Cerebras dependency: any shift in Cerebras's supply or pricing strategy could ripple into Devin's own cost and speed.