Tencent Hunyuan · Hugging Face new modelsThe Hy-Memory-30B-A3B-GGUF model quantization file released by Tencent's Huayan uses a hybrid precision quantization scheme with imatrix calibration and MoE perception, achieving efficient local inference on llama.cpp, where the ~8 GB expert-mixed build file size is reduced to about 24.71 GB, with performance only 1.53% lower than Q6_K.
Prime Intellect Blogprime-rl v0.10.0 has been released, defaulting to the new RL algorithm IPO to improve the stability of asynchronous RL and optimize performance. It supports new models such as Qwen3.8 Flash, and increases training throughput by 1.15 times (based on the torchtitan platform) through MoE optimization and custom kernels.
vLLM BlogvLLM achieves a downtime reduction of DeepSeek V4 Pro from 16GB200 to 20GB200 to 0.82 seconds and from 20GB200 back to 16GB200 to 1.96 seconds through asynchronous preparation and NIXL EP technology.
Inception Labs BlogMercury 2 decodes over 1000 tokens per second on standard NVIDIA GPUs, nearly 2 times faster than Gemini 3.1 Flash Lite in the query planning stage, 4.7 times faster than Claude Haiku 4.5, and 10 times faster than GPT-5 Mini. Each query step is faster than other models, allowing it to complete every step of the search process within the existing latency budget.
Inception Labs BlogMercury is launched on Azure AI Foundry, providing developers with the first commercially scalable diffusion large language model (dLLM), whose parallel generation capabilities based on diffusion architecture make its inference speed 10 times faster than similar autoregressive models.
Google Developers BlogLiteRT Tensor API is released, supporting developers to write and optimize models natively in C++ on edge devices, enabling cross-platform deployment (CPU/GPU/NPU) without script conversion. Developers can fully control the computational graph structure and directly load model weights, achieving zero-copy data transfer through shared accelerator buffers, while providing building blocks for AI coding agents.
Inception Labs BlogInception Labs has launched Mercury 2, increasing the number of free tokens from 10M to 100M. The new Inception API key includes 100M free tokens, and the free tier's request frequency has been increased 10 times. Mercury 2 performs better in real-time speech, search, AI coding sub-agents, and multi-agent workflows. The first token response time is reduced, end-to-end latency is lowered, and tool calls are more reliable.
Inception Labs BlogMercury 2 completes context compression tasks in just 150 seconds while maintaining the same quality as Sonnet 4.6, with a 5-fold speed increase and a 90% cost reduction. Augment Code uses it for enterprise-level coding AI agents, achieving an 82% reduction in compression latency and a 90% cost cut, and dynamically allocates tasks through the Prism model router, reducing total LLM expenditure by 30%.
Inception Labs BlogMercury Edit 2 is a new diffusion model (dLLM) launched by Inception Labs, used to predict the next step in code editing, with a 48% improvement in acceptance rate compared to the previous model, and more precise editing choices (improvement of 27%).
Hugging Face BlogTorchNeuron provides a new native backend for PyTorch on AWS Trainium, with native support for Transformers, Diffusers, and Kernels, making it unprecedentedly easy to run models on Trainium and share models on HF Hub.
Google Developers BlogUsing ADK and Gemini 3.8, non-blocking voice interaction is implemented through streaming tools and asynchronous functions, reducing the first response time of the sales report workflow from 14.4 seconds to 1.2 seconds, supporting zero silence, progress updates, and parallel conversations.
MiniMax NewsMiniMax M3.1 Flash is officially launched, with nearly a threefold increase in online inference speed and a 70% reduction in cost compared to the previous generation. The model performs excellently in benchmark tests for complex intelligent tasks (cowork), computer use, software development, and network security, ranking highly in the AutomationBench-AA for automated tasks, scoring 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on ALE, and 82% on CyberGym.
Ollama BlogCloudflare's 27B multimodal decision model Clef has landed on the Ollama cloud platform, users can run it locally or use cloud services. Clef transforms state and typed question patterns into decisions, supports text and image input, is compatible with Jev and System One APIs, is fine-tuned based on Qwen3.8-27B, and adopts the Apache 2.0 license. In Cloudflare's decision index evaluation, Clef outperforms Jev and performs leading in most benchmark tests.
Inception Labs BlogInception Labs launches Mercury Voice diffusion large language model, designed for voice agents, reasoning, calling tools, and following long system prompts, with a first response delay of less than 320 milliseconds, defeating GPT-6 Luna and Gemma 4 31B on proxies and voice benchmarks.
Together AI BlogTogether AI has launched Post-Training services, supporting the construction of custom models through reinforcement learning and supervised fine-tuning without complex infrastructure. The service provides flexible training loop definition capabilities and optimizes the underlying system and infrastructure to improve model quality and training speed. It supports mainstream open-source models such as Muse Spark Open, GLM 5.3, Kimi K3, and Qwen 3.8-27B, running on dedicated GPU resources to ensure efficient and numerically consistent training.
Inception Labs BlogMercury 2 is the first reasoning diffusion language model, decoding over 1000 tokens per second on standard NVIDIA GPUs, reducing inference cost from 3 seconds of silence to 300 milliseconds.
Inception Labs BlogInception Labs releases Mercury 3, this diffusion large language model surpasses GPT-6 Luna and Claude Haiku 5.5 in multiple benchmark tests, with a speed increase of 4-8 times reaching over 1,000+ tokens/second, expanding the context window to 1M tokens, with prices per million input/output tokens being $0.20 and $0.70 (50% discount at launch).
Inception Labs Blog7575SearchBlox SearchAI now integrates Inception's Mercury dLLM, providing sub-second GenAI responses for enterprise workloads. This collaboration enables SearchBlox customers to achieve real-time search in e-commerce, customer service, knowledge management scenarios with 60-90% reduced reasoning costs and scalable speed.
Inception Labs BlogMercury 2 is released on Azure AI Foundry, the world's fastest reasoning language model, built to make production AI feel instant. The model combines Inception's breakthrough diffusion architecture with Azure's enterprise-grade infrastructure, providing developers with models that offer both exceptional speed and quality.
Inception Labs BlogMercury 2 achieves a 78% task success rate on PinchBench, the fastest execution speed in its class, with a cost of $0.25 per million inputs and $0.75 per output, 4 times lower than Claude 4.5 Haiku, defining the feasible boundary for AI agents that balance performance, speed, and price.
Inception Labs BlogInception Labs has launched the world's fastest reasoning large language model, Mercury 2, based on a diffusion model architecture, achieving a generation speed of 1,009 tokens per second on NVIDIA Blackwell GPUs.
Browser UseThe browser agent operates web pages through AI models to achieve goals. After the user describes the task, the agent automatically opens pages, navigates, selects operations, and checks results to continue. It is particularly suitable for uncertain workflows that require flexibility and decision-making, composed of three parts: the model (decides actions), the suite (provides context, tools, and loops), and the browser (manages website sessions). Main uses include data extraction, research, form filling, etc., but there are drawbacks such as inconsistent results, task failure, website blocking, prompt injection, and unpredictable costs.
DeepSeek · Hugging Face new models5656DeepSeek-V4.1-Flash-1017 Released, Significantly Improving on Reasoning, Coding, Agent Tasks, and Visual Benchmarks, Reducing Execution Time and Output Length, Completing Tasks Faster and at Lower Cost. Scored 42.6 on HLE (No Tools) Benchmark, 51.8 on Terminal-Bench 4.0, 77.0 on DeepSWE v1.1, License is MIT.
Google ResearchGoogle Research releases the VERID framework, enabling AI agents to run in physics-based, safe, and resource-efficient experiments; demonstrates closed-loop intelligent agent scientific discovery on adapter and de novo protein design prototypes; through three rounds of design-test-iterate cycles, AI-driven rational and de novo design will increase confirmed target hit rates to 33.5% and 42.1% on myoglobin and α-thrombin, respectively.
Redwood ResearchScrewtape 1 is a pipeline that transcribes an agent into an evasive action-only monitor (not reviewing the agent's reasoning chain) by inserting the attack into the transcription and repeatedly refining it until the monitor no longer captures it. Use an LLM verifier to check the edited transcription and attack description to confirm whether the attack is real, obviously harmful, and likely to succeed in the original environment.
Tencent Hunyuan · Hugging Face new modelsQwen3-8B-DSpine-b16 released by Qwen, employing speculative decoding technology with parallel drafting and deep conditional injection, achieving an average acceptance length score of 10.55 on Qwen3-8B. The model supports a context length of 8192 tokens, uses bfloat16 data type and fa3 attention backend, installation requires git clone and pip install, and runs via sglang serve.
Modal BlogModal releases VM Sandboxes virtual machine sandbox, supporting complex Agent workloads, adopting a block device solution to solve the 99µs chown overhead and 6 times host system call problem added by virtio-fs, achieving sub-second cold start performance same as containers.
Dify Blog6565Dify supports selecting AI Agent or Workflow based on whether the task is repetitive, whether judgment is needed, and whether there is a clear result. Agents are chosen by the model selection tool or action, while Workflows are defined by developers. The Hybrid mode combines both. In Dify, Agents are suitable for repetitive tasks requiring explanation of requests and selection of the next step, while Workflows are suitable for SOPs with fixed steps, where Agent nodes can be embedded to handle flexible steps within the Workflow.
Tencent Hunyuan · Hugging Face new modelsHy-Memory-30B-A3B-FP8 released by Hunyuan is a compact long-term memory writing model with 3B active parameters, designed for efficient vLLM deployment, and performs better than Claude Opus 5 and DeepSeek V4.1 Flash in the cross harness macro average score of a fixed 20% benchmark matrix.
OpenAI News3535Anant Kumar uses Codex for code review and to speed up follow-up.
Ollama BlogReflection's first open-source weight model Beam has been fully launched this morning. Ollama is the first partner to provide this model for a limited time at no cost, and its Beam endpoint supports full zero-data retention (ZDR).
vLLM BlogvLLM has built a unified CuTe DSL decoding path for Kimi K3 through intelligent agent-assisted development, achieving a batch 1 performance of 840 tokens/s on 16 NVIDIA GB300 GPUs, with a 1.5-2x increase in decoding throughput in the test configuration, and a 1.8x increase in AgentX workload P90 interactivity.
LangChain BlogThe LangChain blog introduces how to use Agent Evals to test AI agent behavior, including analyzing behavior through trace logs, building test datasets containing inputs and expected outputs, running offline benchmark tests (such as Terminal-Bench) and online real-time evaluations, as well as using four types of evaluators: code inspection, LLM scoring, and manual review. Offline evaluations are used to test new prompts or models and catch regression issues, while online evaluations are used to monitor production quality and discover dataset gaps.
Comfy BlogBlack Forest Labs releases open-source FLUX 3 Image Klein model with 5.6B parameters, outperforming Qwen-Image 2.1 on text-to-image and multi-reference generation tasks, supporting both Base and Distilled versions, with the latter offering 4.3-9.3 times speed improvement, suitable for consumer-grade GPUs.
Baseten BlogMeta releases Muse Spark Open, an open-source weight model with 816B parameters designed for code agents, featuring 42.7B active parameters, a 1M token context window, and full-modal input (text, image, video, audio), performance comparable to cutting-edge lab models, using a compressed MoE architecture and hybrid global/local attention mechanisms to optimize inference efficiency, reducing model size to 540GB via NVFP4 checkpoints.
Microsoft AI NewsMicrosoft AI has completed the pre-training of a new MoE large model MAI-Base-2, with 80 billion active parameters and 2.8 trillion total parameters, surpassing all open-source models in internal evaluations. The model is 4-12 times more efficient in pre-training FLOPs than the strongest open-source base model DeepSeek-V4-Pro-Base and is used to train larger models. Microsoft plans to license the model weights through Foundry using organic data pre-training and full-stack control.
LangChain BlogLangChain introduces new features such as LangSmith AI Search, Managed Deep Agents v0.9, and LangSmith Engine v2 to help teams more flexibly build, understand, and optimize AI agents in production environments; the open-source LangChain v1.4.0 supports FastMCP and shares a case study of Jev combined with LangGraph for building efficient AI agents.
