Cognition SWE-1.7 Release Notes: Frontier Intelligence for Software Engineering

Cognition has launched SWE-1.7, a model designed to provide frontier-level intelligence for software engineering at a significantly lower cost. Built from a Kimi K2.7 base, the model is optimized for long-horizon asynchronous tasks and is now available in Devin via Cerebras at 1,000 tokens per second (TPS).

Performance Benchmarks

SWE-1.7 demonstrates high pass rates on agentic coding benchmarks, competing closely with models like GPT-5.5 and Opus 4.8.

Benchmark	SWE-1.7	Kimi K2.7 Code	GPT-5.5	Opus 4.8	Opus 4.7	GLM-5.2	Composer 2.5	SWE-1.6
FrontierCode 1.1 Main	42.3%	30.1%	43.0%	46.5%	38.5%	24.5%	25.6%	9.4%
Terminal-Bench 2.1	81.5%	72.7%	84.2%	86.9%	83.0%	81.0%	76.0%	39.7%
SWE-Bench Multilingual	77.8%	73.5%	76.8%	84.4%	80.5%	74.5%	71.6%	58.3%

Technical Architecture and Training

SWE-1.7 was developed through broad improvements to the Reinforcement Learning (RL) pipeline, challenging the notion of a "post-training ceiling" by achieving significant gains over its Kimi K2.7 base.

Preserving Entropy and Stabilizing Training

To prevent entropy collapse—where a model stops exploring and reward plateaus—Cognition implemented top-p sampling in rollouts. To resolve the resulting training-inference mismatch, they developed sampling distribution replay, which records the tokens available during sampling and renormalizes probabilities in the trainer. This process reduces gradient noise by focusing optimization on tokens with high learning signals.

Multi-cluster Training and Fault Tolerance

Cognition utilized a distributed RL architecture spanning four datacenters across three continents.

• Weight Updates: To minimize staleness and latency, the trainer sends compressed weight deltas via cloud object storage, reducing transfer sizes by over 99%. Cross-continental updates for a 1T parameter model complete in 1–2 minutes.
• Fault Tolerance: Inference engines are stateless and managed by NVIDIA Dynamo; failures are handled by rerouting trajectories to healthy workers. The trainer uses asynchronous checkpointing to local disks and peer replication to ensure recovery takes only seconds.

Self-Compaction for Long-Horizon Tasks

To handle tasks that exceed the raw context window, SWE-1.7 employs self-compaction. The model is trained to summarize its working state and resume from that summary, allowing rollouts to reach up to six hours in duration.

To prevent the model from becoming overly verbose (a common trend in reasoning models), Cognition uses an alternating length penalty. The model alternates between "unconstrained phases" (optimizing for success) and "budget phases" (penalizing solutions that exceed a cost budget of tokens, turns, and time).

Data Quality and Verifier Rigor

The training data was curated to maximize learning signal by focusing on tasks the model solves only a low fraction of the time. To prevent reward-hacking and cheating, Cognition implemented network-restricted sandboxes, stripped git histories, and assigned a reward of 0 to any trajectory attempting to cheat.

Model Behaviors and Observations

SWE-1.7 exhibits distinct behavioral shifts compared to its base model, Kimi K2.7 Code:

• Condensed Chain-of-Thought: Due to the alternating length penalty, the model uses fewer function words and shorter sentences in its reasoning process.
• Deep Codebase Exploration: SWE-1.7 performs significantly more tool calls, file reads, and searches before acting. It is more likely to investigate root causes and probe edge cases via small Python scripts rather than guessing.
• Increased Change Scope: A side effect of increased reasoning is that the model often touches more files and writes more test cases than strictly required for a task.

Community Perspectives and Critiques

Discussion among developers and researchers has raised several points of skepticism regarding the reported benchmarks:

"What are the chances that CursorBench ranks Cursor's model highest, and Cognition's bench ranks Cognition's model highest? ... their training data and benchmarks come from the same dataset (Devin/Cursor interaction logs) so they naturally overfit."

Other critics pointed to discrepancies between these results and third-party evaluations, such as those on artificialanalysis.ai, suggesting that the benchmarks may be cherry-picked or that the model's performance may not generalize beyond the specific evaluation sets used.