Cognition launches new SWE-2 model, Rivaling Fable 5.1 and GPT-Astra (cognition.com)

Article summary

Cognition has introduced SWE-2, a new coding model that achieves high performance on coding tasks while being more cost-efficient than its predecessors. SWE-2 is post-trained from the Kimi K3 model and uses a novel approach to optimize the cost-performance tradeoff. The model has shown significant improvements in efficiency and intelligence compared to previous models. It is available for use through the Devin platform, with some users having free access to try it out.

Main themes

• AI model performance
• cost-efficiency
• benchmarking
• adoption barriers
• company valuation
• innovation and growth
• model generalizability
• closed weights and benchmaxxing

What commenters say

• The model's performance on certain benchmarks is questionable, with some suggesting it may be overfitting to specific tasks.
• The use of a bespoke platform to interact with the model is seen as a significant barrier to adoption by some users.
• There is skepticism about the model's ability to generalize to new problems, given its performance on different benchmarks.
• Some commenters are critical of the company's valuation and fundraising, suggesting that it may be overhyped or unsustainable.
• Others defend the company's approach and suggest that its valuation may be justified by its potential for growth and innovation.
• The model's closed weights and potential for benchmaxxing are seen as limitations by some, while others argue that these are not necessarily significant drawbacks.
• There is debate about the significance of the model's performance on different benchmarks, with some arguing that it is not necessarily indicative of its overall quality or usefulness.