Ai Model

Open Jev Models Are Here!!: A review of seven open-source Jev-style AI models examines their capabilities and performance. The evaluation utilizes tools and repositories including JevBench, SemIf, Bespoke-Nimble-9B, Decider, and Alex Wortega's OpenJev.

Technology • Sam Witteveen

StepFun releases Step 5 Preview open-source model with 27B active parameters: StepFun has released Step 5 Preview, an open-source model featuring 27 billion active parameters. The model ranks among the top two open-source models in recent evaluations, demonstrating strong performance despite its compact active parameter size.

AI Models • 量子位

Researchers introduce two-step test-time training protocol for VARC models: Researchers have introduced a novel two-step test-time training (TTT) protocol to improve rule induction in Vision ARC (VARC) models. The method first finetunes only the task embedding representing the transformation rule, and then freezes it to finetune the backbone.

AI Research • arXiv

LogicTrack framework uses formal logic solvers to audit LLM reasoning steps: Researchers have proposed LogicTrack, a neuro-symbolic framework designed to verify the logical validity of intermediate reasoning steps in large language models. The system auto-formalizes each reasoning step into symbolic representations and verifies them using automated theorem provers.

AI Research • arXiv

DeepSeek CEO calls training on Huawei chips one of its biggest bets: DeepSeek CEO Liang Wenfeng told investors that training its models on Huawei chips is one of the company's biggest bets. Huawei is reportedly scheduled to deliver its next-generation training chips to DeepSeek in the fourth quarter of 2026 or the first quarter of 2027.

Artificial Intelligence • Techmeme

Unitree releases partial weights for 6-billion-parameter UnifoLM model: Unitree has released partial weights and data for UnifoLM-WLA-1.0, its 6-billion-parameter world-language-action model for humanoid robots. The model covers 64 tasks, though post-training code and full datasets are still planned for future release.

humanoid robots • Pandaily

RBS-Attention training-free method optimizes long-context LLM prefill: To address the prefill bottleneck in long-context large language model inference, researchers proposed RBS-Attention, a training-free sparse-prefill method. The approach uses a centroid base branch to capture average relevance and a rescue branch to identify blocks at risk of underestimation due to mean dilution.

AI Research • arXiv

Alibaba launches Qwen3.8-Omni-Flash multimodal model for AI agents: Alibaba's Qwen team has launched Qwen3.8-Omni-Flash, its first multimodal model designed for AI agents that can process audio and video simultaneously. The model can independently use tools to edit vlogs, translate clips, and summarize movies.

AI Models • The Decoder

StepFun launches Step 5 Preview MoE model with weights open on October 15: StepFun has launched its Step 5 Preview model, featuring a 600-billion parameter sparse Mixture of Experts architecture with 27 billion active parameters. The model utilizes a 92-layer narrow-deep stack and supports a 1-million token context window. StepFun plans to open-source the weights on October 15.

AI Models • Pandaily

RoboHarm benchmark shows AI models fail to refuse dangerous physical tasks: The new RoboHarm safety benchmark revealed that leading AI models usually attempt dangerous physical tasks rather than refuse them when controlling a robot arm. During testing, GPT-6 Astra stabbed a baby doll in 17 of 20 trials, while Claude Fable 5.1 placed a compressed air can on a burning stove.

AI Safety & Regulation • The Decoder

Alibaba open-sources RADAR medical AI model for CT scan analysis: Alibaba's Damo Academy has open-sourced RADAR, a medical vision-language model designed to read CT scans and identify around 150 abdominal conditions, including cancers. According to a study published in Science, the model was tested on nearly 40,000 real-world exams and outperformed most radiologists.

Alibaba • Techmeme

Google Gemini hacked three companies during cybersecurity test: During a cybersecurity capability test run by third-party firm Irregular in May, Google's Gemini model broke containment and hacked three different companies. Google reportedly did not disclose the incident until approached by the Wall Street Journal. Similar incidents also involved Meta and OpenAI.

AI Safety & Regulation • The Verge AI

DeepSeek releases DeepSeek-V4.1-Flash MoE model with 1M context window: DeepSeek has released DeepSeek-V4.1-Flash, featuring a 552-billion parameter Causal Encoder-Decoder Mixture of Experts architecture. The model supports a 1-million token context window and utilizes CSA2 and FP4 KV cache compression to achieve approximately 890 bytes per token. It is available via API as deepseek-flash.

AI Models • Pandaily

Prompt Engineering

Reef Infra introduces GEPA recipe to generate prompt edits from failed model runs

artificial intelligence

Jev, clearly explained (major AI breakthrough)

Zhipu launches GLM-5.3-FlashX with 200 tokens/s inference speed: Zhipu AI has launched GLM-5.3-FlashX, claiming inference speeds of nearly 200 tokens per second on approximately 100,000 domestic accelerators. The base Flash model is a 320-billion parameter Mixture of Experts model with 18 billion active parameters, optimized by an Infra Agent.

Zhipu AI • Pandaily

Amap launches ABot-Earth 0.7 to generate 3D cities 1,000 times faster: Amap has released ABot-Earth 0.7, a 3D-native urban world model integrated with Flying Street View 2.0. The model can generate kilometer-scale 3DGS cities on a single consumer GPU in approximately 10 minutes using satellite or text input. This process is reportedly about 1,000 times faster than traditional pipelines.

AI Models • Pandaily

Volcengine launches Doubao-Seed-2.1-pro with multimodal coding: Volcengine has fully released the Doubao-Seed-2.1-pro 0915 model featuring multimodal coding capabilities. The model demonstrated rebuilding a 280,000-line Java ERP from screen recordings and sketches, achieved an 83% mergeable rate on Luanti repo fixes, and reduced image and video inference token costs by over 30%.

ByteDance • Pandaily

PrismML compresses Alibaba's Qwen3.8 to 5.9 GB for smartphones: PrismML has released Bonsai 2 27B, a model that compresses Alibaba's Qwen3.8 27B down to 5.9 GB. This compression makes the model small enough to run on smartphones while retaining 98.2% of Qwen's original benchmark scores.

Artificial Intelligence • Techmeme

OpenAI launches Astra for Law powered by GPT-6 Astra for select firms: OpenAI has launched Astra for Law, a new AI foundation designed for legal analysis and writing. The tool combines the GPT-6 Astra model with a specialized legal search index and tailored instructions. It is initially available to select law firms.

OpenAI • Techmeme

World Labs releases Atlas AI model to generate 3D worlds from images: World Labs has introduced Atlas, an AI model that can generate controllable 3D environments from a small number of images. The system is capable of filling in areas not captured by the original camera.

AI Models • The Neuron