AI News Daily

Cognition releases SWE-2 Flash coding model

Cognition released the SWE-2 Flash coding model, trained on Thinking Machines' Inkling-Small (276B total parameters, 12B active parameters), adding a mid-training stage before reinforcement learning for the first time in its training pipeline. The company said Terminal-Bench 4.0 rose from 0.9% to 20.0%, FrontierCode 2.0 rose from 3.2% to 24.0%, and ExploitBench rose from 11.3% to 34.8%. The model is now available on Devin Desktop and the CLI, priced at one-fifth of the larger sibling SWE-2 and three times faster (200 tok/s versus 60 tok/s); its weights are not yet public, with plans to share them securely with trusted industry partners and research institutions through the program.

• Cognition releases SWE-2 Flash coding modelCognition
• Cognition releases SWE-2 Flash coding model and discloses training detailsCognition
• Cognition releases SWE-2 Flash coding model, built on Inkling-Small with new…Cognition
• Cognition releases coding model SWE-2 Flash, built on Inkling-SmallCognition

China-region Apple Intelligence and Siri launch with the iOS 27.3 developer preview

Apple released the iOS 27.3 developer preview, and China-region "Apple Intelligence and Siri" arrived as promised; after upgrading, users of iPhone 15 Pro and later models can find the feature page in Settings. The first batch of features focuses on basic high-frequency scenarios, including Live Translation in Messages, FaceTime calls, and the Phone app; rewriting, proofreading, and summarizing in Writing Tools; visual intelligence recognition in the camera; and the Clean Up tool in the Photos app. The China-region version is a collaboration with Alibaba, and users can invoke Qwen in Siri, Writing Tools, and Visual Intelligence; by logging in to a Qwen account or using "Sign in with Apple," they can access it within the system without jumping to a standalone app.

Baidu releases physical AI world model Apollo World

Baidu unveiled its world model for physical AI, Apollo World, at its AI Day in Shenzhen. The model can turn complex road interactions and rare driving scenarios into repeatable, controllable simulations for training and testing autonomous driving systems. It supports AI traffic agents that make decisions autonomously and respond by slowing down, changing lanes, and yielding, generates simulated camera, LiDAR, and other sensor data in real time, and covers a complete trip in a virtual city from accepting a ride request to drop-off. Apollo World is already being used for testing and validation of Apollo Go.

• Baidu Apollo discloses world model Apollo World for the first time, proposing a "2+3" evaluation standardLeiphone
• Baidu discloses Apollo World world model for the first time, virtual road testing covers 100,000 km per dayEnglish text is being prepared.
• Baidu discloses Apollo World world model, Apollo Go fully driverless operations exceed one million kilometersSina Tech
• Baidu discloses ApolloWorld, a world model for physical AICheDongXi

Together AI releases Together Compiler, boosting inference throughput by 1.3-3x

Together AI has released Together Compiler (TKCC), which uses high-throughput agents to continuously generate inference optimizations for different traffic patterns, models, and hardware. The company says it achieved 1.3-3x TPS improvements on Kimi K3, GLM 5.3, MiniMax M3.1-Flash, and DeepSeek v4.1-Flash in under 24 hours, and reached SOTA on SemiAnalysis's InferenceX low-latency benchmark with MiniMax M3.1-Flash at 1111 TPS and DeepSeek v4.1-Flash at 1205 TPS. The system uses a resident engine and a two-tier fast/slow engine to balance experimental throughput and measurement fidelity; its experimental throughput is 252 times that of baseline agent systems, and it is already used in online serving.

OpenAI's automated vulnerability chain discovers 1,724 Linux kernel vulnerabilities

OpenAI's automated vulnerability research pipeline identified 2,308 Linux kernel vulnerability candidates, dynamically reproduced and confirmed 1,724 vulnerabilities, and generated 948 local privilege escalation exploits. The pipeline is divided into five stages: vulnerability discovery, dynamic verification, exploitability assessment, patch generation, and human review. The full cost of the second scan across the entire process was approximately $1.21 million in API credits. Estimated by bug bounty standards, these 948 exploits are worth $18.9 million to $66.3 million, 16 to 55 times the cost. The agent used 10 new vulnerability exploitation techniques verified by third-party kernel vulnerability experts, and OpenAI is submitting vulnerability information and patches to Linux kernel maintainers.

• OpenAI's automated vulnerability chain discovers 1,724 Linux kernel vulnerabilitiesSina Tech

Report: DeepSeek plans to pretrain its next-generation V5 model on Huawei Ascend chips

According to STAR Market Daily, DeepSeek plans to pretrain its next-generation V5 model on Huawei Ascend chips. Previously, Ascend only handled DeepSeek's inference and V4-Pro post-training tasks. X.PIN believes that DeepSeek's open-sourcing of its full Ascend compute and communication stack (DeepGEMM, FlashMLA, DeepEP) on September 30 is related to this.

Anthropic biology team uses Claude to map the first 3D structure of bacterial vault particles

Anthropic's biology research team used Claude to map the first 3D structure of vault particles in bacteria and the viruses that infect them; these barrel-shaped protein structures had previously been found only in animal and plant cells. Claude participated in the actual wet-lab workflow, cutting particle isolation time from 3 days to about 90 minutes and producing the first 3D structure model within 34 hours, whereas manual analysis would originally have taken weeks to months. Anthropic also noted that Claude made some simple errors during the process, which the team handled in collaboration with machine learning researchers.

Nathan Lambert reviews Ripple-65B: small teams can also build solid models

Nathan Lambert commented on Ripple-65B Preview, released by Ricursive Intelligence, calling it an ambitious project and expressing hope that more US labs will follow suit and build solid models with small, focused teams. He noted that the model follows the olmo series in its approach and that it is meaningful to see a project he was involved in scaled up. The quoted tweet said Ripple-65B is a 65B dense large language model pretrained from scratch that surpasses larger models such as Muse Spark Open-816B, Beam-501B, Inkling-975B, and Nemotron 3 Ultra-550B on agentic coding tasks among US open-weight models.