Alibaba announces Qwen 4 at the Apsara Conference
Alibaba has officially announced the release of Qwen 4 during the Apsara Conference.
Alibaba Qwen releases Qwen3.8-Omni-Flash, an agentic omni-modal model with 1M context
Alibaba's Qwen team has released Qwen3.8-Omni-Flash, its first omni-modal model designed around agentic capabilities for audio-video understanding and tool use. The model accepts text, images, audio, and video inputs to return text outputs, featuring a 1M-token context window and native function calling.
Alibaba's Qwen Team releases Qwen3.8-Flash-Next, a 125B MoE previewing Qwen4
Alibaba’s Qwen team has released Qwen3.8-Flash-Next, an open-weight multimodal Mixture-of-Experts model designed to preview the architecture for the upcoming Qwen4 series. The checkpoint pairs a 125B backbone with a 51B N-gram embedding table and a 4B multi-token prediction module, activating only 6B parameters per token.
VideoX-Qwen introduces data-centric framework for instruction-based video editing
Researchers present VideoX-Qwen, an integrated framework combining scalable data construction with model training to advance general-purpose, instruction-driven video editing. The system utilizes a production pipeline that organizes specialized models to generate directional editing records, resulting in over 1.2 million high-quality samples across addition, removal, replacement, and attribute tasks.
Qwen releases Qwen3.8-Omni-Flash, a native multimodal agentic model with 1M token context
Alibaba Cloud has introduced Qwen3.8-Omni-Flash, a natively multimodal agentic model designed for real-world productivity tasks that substantially improves multimodal understanding and reasoning compared to previous omni models.
Alibaba plans AI model with 5 trillion to 10 trillion parameters and unveils new chip
Alibaba has announced plans to develop an artificial intelligence model containing between 5 trillion and 10 trillion parameters. Alongside this roadmap, the company also unveiled a new custom chip designed to support its infrastructure needs.
Alibaba Qwen releases unified 7B Qwen-Image-2.1 model for image generation and editing
Alibaba's Qwen team has released Qwen-Image-2.1, a unified text-to-image generation and image editing model that consolidates previous separate checkpoints into a single 7B parameter diffusion transformer.
Qwen releases open-weight Qwen-Image-2.1 image generation model
Qwen has released the open-weight Qwen-Image-2.1, a unified model designed for both image generation and editing.
Alibaba Qwen Team releases Qwen3.8-LiveTranslate with Interleave architecture
The Alibaba Qwen Team has released Qwen3.8-LiveTranslate, a next-generation real-time simultaneous interpretation model that listens to live speech and returns translated text and audio while the speaker is still talking. The release introduces a new Interleave architecture designed to reduce latency and improve translation quality.
Alibaba open-sources medical AI model that can detect cancer and nearly 150 conditions
Alibaba has released an open-source medical artificial intelligence model capable of detecting cancer and approximately 150 other medical conditions.
Qwen3.8 Max (0902) scores 45 on the Artificial Analysis Intelligence Index
The Qwen3.8 Max (0902) model has achieved a score of 45 on the Artificial Analysis Intelligence Index, reclaiming the top spot on China's leaderboard.
TRL v1.14 AsyncGRPOTrainer enables LoRA-only sync across Hugging Face Jobs
TRL v1.14 introduces LoRA support to the AsyncGRPOTrainer, allowing it to train a Low-Rank Adaptation adapter and synchronize only that small file to vLLM inference workers. This architecture decouples training and generation jobs on separate machines by using a shared Storage Bucket as a filesystem bridge, eliminating the need for NCCL or direct network communication between nodes.
Study identifies FragileTokens that fail contextual copying despite passing isolation probes
A study characterizes "FragileTokens," vocabulary entries in open-weight language models that successfully copy when isolated but exhibit errors when embedded in surrounding text. The research highlights that literal identity preservation is not guaranteed by standard isolation tests, as tokens can be deleted, substituted, or truncated within sequences.
ExecCritic uses role-specific RL to improve coding agents via test-guided repair
ExecCritic introduces a framework that separates test construction from source-code repair to prevent false confidence in coding agents. The system employs a Test agent and a Repair agent, both backed by Qwen-3.5-35B-A3B, trained separately using reinforcement learning.
ExecCritic uses role-specific RL to improve coding agents via test-guided repair
Researchers introduce ExecCritic, a framework that combines a test-verify-revise scaffold with role-specific reinforcement learning to enhance coding agents. The system separates test construction from source-code repair, using a Test agent to generate repository-native tests and a Repair agent to revise code based on execution feedback.
Speculative Uncertainty method cuts software engineering agent errors using draft-model gating
The authors present Speculative Uncertainty (SU), a technique that recovers a predictive failure signal for black-box LLM agents by analyzing their output tokens alone, without requiring access to logits, weights, or activations. By inverting speculative decoding, a small open-weight draft model scores the agent's trajectory in a single forward pass to extract phase-aware features and calibrate them against a verifiable objective.
VisCAD releases foundation model suite with multimodal industrial CAD intelligence
Researchers present VisCAD, a foundation model suite designed to provide broad generalization and strong capability for realistic industrial product design. The suite addresses the challenges of part-level generation from diverse inputs like renders and text, as well as assembly-level generation involving interacting parts.
AntLing open sources Ling-3.0-flash-Fin, a finance-enhanced model
Ant Group has open sourced Ling-3.0-flash-Fin, the first finance-enhanced model in the Ant Ling family, developed with leading financial institutions and domain experts.
OpenEnv reproduces Surya Narreddi's watercolour painting model using TRL
An engineer has open-sourced a reproduction of Surya Narreddi’s viral project, which trains a coding model to paint watercolours using JavaScript and reinforcement learning. The implementation utilizes the TRL library and OpenEnv to create an end-to-end pipeline on Hugging Face for training, scoring, and inference.
CANOPY enables Qwen3-14B to top AppWorld using outcome-only RL
The paper introduces CANOPY (Coverage-ANchored On-PolicY RL), a protocol that addresses signal starvation and policy drift in long-horizon reinforcement learning for small open models. By scaling same-task exploration and keeping updates KL-anchored, the method allows agents to learn effectively from end-of-task verification alone.