DevCheck Beating AI News: The $80 billion AI programming company Cognition launches its new model SWE-2 Flash, emphasizing faster speed and lower cost.

Compared to the previous generation SWE-2, the new model reduces usage costs to 1/5 and increases generation speed from around 60 tokens per second to 200, primarily targeting programming tasks such as code review and feature development.

SWE-2 Flash is trained on Thinking Machines' Inkling-Small foundation with a total of 276 billion parameters, but only activates approximately 12 billion parameters per generation. The prior SWE-2 used Kimi K3, a 2.8 trillion-parameter model from Moonshot AI. Cognition essentially replaced the base model with a much smaller one and enhanced its programming capabilities through targeted training.

This training also introduced an additional intermediate phase. Cognition first had the model learn trillions of tokens from code, research papers, and blogs, then applied large-scale reinforcement learning using high-quality programming tasks. The main improvements stem from richer training tasks and stricter validation of task requirements and outcomes.

In official tests, SWE-2 Flash scored 24.0 on FrontierCode 2.0—slightly higher than the previous SWE-2's 23.8—and far surpassing the base model's 3.2. However, on Terminal-Bench 4.0, the new model achieved only 20.0%, falling short of the prior SWE-2's 27.3% and significantly behind DeepSeek V4.1 Flash's 42.3%.

Cognition also claims that SWE-2 Flash outperforms all U.S.-based open-weight models in major programming benchmarks. However, the Flash model's weights remain closed for now. The company states that after security-focused training, the model's score on ExploitBench—a vulnerability exploitation test—rose from 11.3 to 34.8. Concerned about increased cyberattack risks if weights were publicly released, they plan to grant access only to vetted partners and researchers.

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

Dynamic Insight Beating AI News: Moonshot AI has officially open-sourced the complete model weights of Kimi K3.1 Pro.

The model was already launched for applications and APIs in late October. The newly released Hugging Face repository contains 96 weight shards with a total volume of approximately 1.56 TB.

Kimi K3.1 Pro retains the customized licensing terms from its predecessor K3, allowing free use, modification, distribution, and commercialization, but includes two special restrictions:

1. Large-scale API service providers must sign a separate agreement. If a company operates model inference or fine-tuning services and its total revenue—along with that of affiliated companies—exceeds $20 million over a continuous 12-month period, it must sign an additional agreement with Moonshot AI before commercial use.

2. Ultra-large-scale commercial products must prominently label the model name. If a product using this model has more than 100 million monthly active users or generates over $20 million in monthly revenue, it must clearly display "Kimi K3.1-Pro" on its interface. Internal use only, or invoking the model through Moonshot AI's official products or certified inference partners, exempts users from these signing and labeling requirements.

To reduce network attack risks associated with open-weight models, Moonshot AI has proactively removed network attack-related content from mid-stage training data and reinforcement learning tasks, intentionally limiting the model's vulnerability exploitation capabilities.

According to internal testing, the unmodified version of the model achieved a success rate of up to 90.5% within three attempts in vulnerability exploitation tasks, while the final released version dropped to 81.0%, still significantly higher than the previous K3’s 33.3%.