header-langage
简体中文
繁體中文
English
Tiếng Việt
한국어
日本語
ภาษาไทย
Türkçe
Scan to Download the APP
Partnered With:

Cognition launches SWE-2 Flash: 3x faster, cost reduced to 1/5

Beating AI News Flash: AI coding company Cognition, valued at $80 billion, has released a new model, SWE-2 Flash, touted as faster and cheaper.


Compared with the previous generation SWE-2, the new model's usage cost has dropped to 1/5, and generation speed has increased from about 60 tokens per second to 200. It is mainly used for coding tasks such as code review and feature development.


SWE-2 Flash is trained on Thinking Machines' Inkling-Small, with 276 billion total parameters, but only about 12 billion parameters are activated per generation. The previous generation SWE-2 used Moonshot AI's 2.8 trillion-parameter Kimi K3. Cognition has effectively switched to a much smaller base and then improved coding capability through training.


This training also added a mid-training stage. Cognition first has the model learn trillions of tokens of code, scientific research articles, and blogs, and then conducts large-scale reinforcement learning with high-quality coding tasks. The main improvements come from richer training tasks and strict review of task requirements and result validation.


In official tests, SWE-2 Flash scored 24.0 on FrontierCode 2.0, slightly above the previous generation SWE-2's 23.8 and far above the base model's 3.2. But on Terminal-Bench 4.0, the new model scored only 20.0%, below the previous generation SWE-2's 27.3%, and far behind DeepSeek V4.1 Flash's 42.3%.


Cognition also claims that SWE-2 Flash surpasses all U.S. open-weight models in major coding benchmarks. However, Flash's own weights are not open for now. The company says that after cybersecurity-related training, the model's score on the ExploitBench vulnerability exploitation benchmark rose from 11.3 to 34.8, and it is concerned that public weights would increase the risk of cyberattacks, so it only plans to provide access to vetted partners and researchers.


SWE-2 Flash is currently available on Devin Desktop and CLI. Pro, Max, and Teams users can use it for free until November 30.

举报 Correction/Report
RelatedRecommended
Flash Trade improves user trading experience through direct trading.
Flash Trade aims to address the security and liquidity risks of Wrapped Tokens.
DeepSeek V4-Flash Official Version API Launches Public Beta, Offering 1 Million Context, Pondering Mode, and Tool Invocation Capabilities at a Very Low Price.
Correction/Report
Submit
Add Library
Visible to myself only
Public
Save
Choose Library
Add Library
Cancel
Finish