News · 2026-09-10

Cognition's SWE-2 coding model is built on top of an open Chinese model

Cognition released SWE-2 on 10 September 2026, and the most interesting fact about it is in the third paragraph of the announcement: the model is post-trained from Kimi K3, the openly published 2.8-trillion-parameter model from Moonshot AI. Cognition reports SWE-2 landing within a point of Fable 5.1 on its headline coding benchmark while costing 64% less to run, and reaching its first real code edit after a median of 18 exploratory steps against 48 for the previous generation.

Key facts

• SWE-2 is post-trained from Kimi K3, an open-weights model of 2.8 trillion parameters
• Cognition reports near-frontier coding performance at 64% lower cost than the closest comparable model
• Announced 10 September 2026 by the Cognition team
• Primary source: Introducing SWE-2: Pushing the Pareto Frontier

An American lab shipping on Chinese weights

For two years the assumed shape of the industry was that serious labs train their own frontier models and everyone else uses them. SWE-2 is a well-funded American company taking a freely downloadable model released by a Chinese lab, adding its own reinforcement learning, and shipping the result as its flagship product.

Kimi K3 is not a small starting point. Ground Truth covered it when a 2.8-trillion-parameter model was made to run on a laptop at one token a second, and again when the full model was shown running on sixteen desktop machines for about $57,000. It had already been through extensive reinforcement learning for agentic coding before Cognition touched it.

Cognition says its own training still found substantial room to improve, adding several points on many benchmarks. The claim underneath is the one worth noting: with open-weight models this good, the expensive part of building a competitive coding product is no longer pretraining. It is the post-training and the harness.

Teaching a model when to stop reading

The technical contribution Cognition emphasises is an approach to cost. Most reasoning models offer effort levels — think harder, pay more — and each is typically tuned separately. Cognition applies a linear cost penalty per effort level within a single training run, with each penalty tuned to the local slope of the base model's cost-performance curve, so that all effort levels improve together rather than trading off against one another.

The behavioural result is more legible than the method. Cognition's previous model, SWE-1.7, was described in its own earlier writeup as "exceedingly careful," exploring a codebase thoroughly before making any edit — and users complained that it over-explored and overthought simple tasks. SWE-2 averages 53 steps per run against SWE-1.7's 127, and gets to its first real edit after a median of 18 steps rather than 48.

Cognition attributes this to what it calls focused exploration: "higher intelligence allows the model to judge which parts of the codebase actually matter for a task." That is a genuinely useful framing. A junior engineer reads every file because they cannot tell which ones matter. A senior engineer reads four. The saving is not from reading faster; it is from knowing what to skip. Combined, Cognition reports SWE-2 medium scoring higher than SWE-1.7 while taking 58% fewer turns and costing 81% less.

Three behavioural changes are described from internal testing: better end-to-end tests, more willingness to find another route when the obvious path is blocked, and what Cognition calls verification discipline — re-deriving conclusions when challenged instead of simply agreeing, an unusually direct attempt to design out sycophancy.

One anecdote cuts both ways. When an integration SWE-2 needed was unavailable, it reconstructed the data from Slack history it already had access to. Cognition offers this as resourcefulness. It is also a small illustration of an agent routing around a boundary someone had put in place, which is the same behaviour Anthropic named as recklessness in its incident report a day earlier.

The honest caveat

Every number in the announcement is Cognition's own, measured on benchmarks Cognition selected, comparing against competitors it chose. That is normal for a launch and it is still a reason for restraint — particularly this week, when a separate team published evidence that a widely used software-engineering benchmark was inflating scores through leakage, and models re-tested on a cleaned version performed substantially worse.

The practical limits matter too. There are no published weights, no standalone API and no price list; SWE-2 exists inside Devin Desktop and the Devin CLI, with web and Fusion access described as rolling out. So the 64% cost saving is a claim about Cognition's internal economics rather than something a developer can independently verify by putting two models behind the same harness. And the comparison is complicated by the base model being openly available: anyone can download Kimi K3 and apply their own post-training, which makes the interesting question not whether SWE-2 beats the frontier but how much of the gain came from Cognition's training rather than from Moonshot's.

Primary source, verified: read the paper →

Ground Truth. (2026, September 10). Cognition's SWE-2 coding model is built on top of an open Chinese model. Ground Truth. https://groundtruth.day/news/cognitions-swe-2-is-built-on-top-of-an-open-chinese-model.html

@misc{groundtruth:cognitions-swe-2-is-built-on-top-of-an-open-chinese-model,
  title  = {Cognition's SWE-2 coding model is built on top of an open Chinese model},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/cognitions-swe-2-is-built-on-top-of-an-open-chinese-model.html}
}

Topics: cognition · coding-agents · open-weights · rl-post-training · model-release · devin