Breaking

Cognition Ships SWE-1.7, a Coding Model Built on China's Kimi K2.7

July 9, 2026 at 16:00 EDT

Software Dev & Coding
AI Agents
Foundation Models

Cognition, the maker of Devin, formally released SWE-1.7 on July 8, 2026, a software-engineering model trained on top of Kimi K2.7, the open-source base model from China's Moonshot AI. The company says trustworthiness-focused training lets it approach US frontier models while running at a fraction of the cost.

July 8, 2026 · Cognition

SWE-1.7: Near-Frontier Coding, at a Fraction of the Cost

Cognition's new agentic coding model — built on China's open-source Kimi K2.7 and hardened with its own RL and trustworthiness training — approaches US frontier models while running at roughly $1.97 per task.

42.3%

FrontierCode 1.1 (Main) — near GPT-5.5's 43.0%

$1.97

Avg. cost per task on FrontierCode Main

~1,000

Tokens/sec inference on Cerebras

FrontierCode 1.1 (Main) — real-world coding agents

Higher is better · columns drawn to scale

46.5%

Claude
Opus 4.8

43.0%

GPT-5.5

42.3%

SWE-1.7

30.1%

Kimi K2.7
(base)

24.5%

GLM-5.2

Cognition's RL layer lifts the Kimi K2.7 base from 30.1% → 42.3%, closing most of the gap to US frontier models — a challenge to the idea of a fixed "post-training ceiling."

81.5%

Terminal-Bench 2.1

77.8%

SWE-Bench Multilingual

~1T

Est. parameters, 4 data centers / 3 continents

Trustworthiness training

Addressing concerns over Chinese base models — refusing requests that touch civil rights and privacy.

87%

of tasks other models refuse over human-rights concerns were completed by the base Kimi K2.7

→ matches US models

After hardening, SWE-1.7 correctly refuses surveillance requests on evaluations

PRAISED

Favorable real-world reception — developers generated large volumes of PRs during free access, and the technical write-up on RL stabilization drew praise.

SKEPTICISM

Questions over whether benchmarks were cherry-picked in Cognition's favor, alongside references to past demo controversies.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.