Weekly LLM rankings, capability comparison and benchmarks. Track GPT, Claude, Gemini and domestic model updates.
Hugging Face released DSpark draft-model checkpoints for three LFM2.5 models. Using speculative decoding, they raise high-end GPU throughput by up to 3.18x and on-device throughput by up to 2.87x without changing output quality. The draft models have about 300M parameters, cut average latency by 57% in LFM2.5-2.6B agent (tool-calling) scenarios, are open-source, and support llama.cpp and SGLang.
Alibaba has officially launched Qwen-UI Agent, a real-world-centric GUI agent foundation model covering mobile, desktop, web, and DeepSearch environments.
Liquid AI released four Q4_0 GGUF checkpoints - LFM2.5-230M, 350M, 1.2B-Instruct, and 2.6B - trained with QAD (quantization-aware distillation, a "slimming" technique where a smaller model is taught by a larger one). They recover 97% of the average BF16 accuracy loss while keeping native Q4_0 memory and speed.
OpenAI launched GPT-5.1, with significant gains in complex reasoning, code generation, and image understanding; API latency and cost are further optimized, and enterprise usability is enhanced.
Anthropic released the Claude 4 family, supporting an ultra-long context window and stronger agent capabilities, with standout performance in code refactoring and long-document analysis.
Google launched Gemini 3, strengthening native multimodal capability and long-video understanding, and opened a finer-grained video-analysis API to developers.
Alibaba open-sourced the Qwen3 model series, covering multiple parameter scales, with strong Chinese and coding ability, and support for tool calling and reasoning-mode switching.
DeepSeek released V3.1, delivering capabilities close to top closed-source models at extremely low inference cost, and continuously pushing the open-source AI frontier.
Gemini 3.7 Flash is now available to Pro and Ultra users inside the Gemini chat. The update improves reasoning and accuracy on multi-step tasks, such as intelligently consolidating dozens of files and emails into a single master document. Gemini Spark also now runs on 3.7 Flash, improving tool calls against Google Workspace apps to make the personal AI assistant more precise.
Qwen has open-sourced the Qwen3.8 model series. Qwen3.8-27B is a native multimodal (image/voice/text) dense model whose 27B parameters already reach the level of Qwen3.7-Plus, with native 262K context extendable to 1M tokens via YaRN, under the Apache 2.0 license. The Max-class Qwen3.8-2.4T-A95B open weights were released at the same time.
Xiaohongshu Technology has open-sourced dots3-note Preview, the lightest model in the dots3 family. With 280B total parameters and 16B activated, it supports 512K context and understands text, vision, and voice, optimized for complex reasoning and long-horizon agent tasks.
Zhipu has released GLM-5.3, built on the same base as GLM-5.2 but pushed to the top of the open-weight coding league through extreme post-training scaling. Coding ability is up 50% over the previous generation, and it ranks first among free-to-use models on Terminal Bench 3.0. It also matches Mythos 5 on white-box code review and scores 84.5 on CyberGym. GLM-5.3 is live now in ZCode and AutoClaw, with an API coming soon and full weights open-sourcing free within two weeks.
DeepSeek-V4-Pro-0813 is now live on SiliconFlow with Day-0 support, offering a 1M context window and low/high/max inference-intensity tiers, with more focus on coding, tool calling, and agent workflows, still under the MIT license. Pricing is input $1.32/M, output $3.96/M, cache-hit $0.44/M. The sibling DeepSeek-V4-Flash-0731 targets everyday production scenarios that value speed and cost.
Zhipu has released GLM-5.3, built on the same base as GLM-5.2 but raising the intelligence ceiling through extreme post-training scaling. Coding improved 50% over the previous generation, ranking first among open-source models on public benchmarks like Terminal Bench 3.0 and approaching Claude Code. On security tasks it matches Mythos 5 in white-box code review and scores 84.5% on CyberGym. Weights open in two weeks; available now on ZCode and AutoClaw.
Alibaba's Qwen team open-sourced the Qwen3.8-2.4T-A95B model weights, its first Qwen-Max-level model released for free use, with 2.4T parameters, 95B active per token, and a context expandable beyond one million tokens.
Microsoft has released MAI-Thinking-1, its first self-built reasoning model, now available on Microsoft Foundry and announced by AI CEO Mustafa Suleyman.
LTX-2.5 produces a 10-second 720p video in 6.8 seconds on two GB200 GPUs with native ComfyUI support, and is free for organizations under 10 million USD ARR.
Cursor and SpaceXAI launched Grok 4.6, boosting long-running agent and interactive vision skills, matching GPT-5.6 Sol on the Artificial Analysis Intelligence Index.
xAI released Grok 4.6, upgrading from 4.5 with stronger long-running agent and visual abilities, matching GPT-5.6 Sol on the nine-benchmark Intelligence Index.
SGLang announced day-zero support for NVIDIA Nemotron 3.5 Lightning, an open-source mixture-of-experts (MoE) model with 30B total parameters and 3B activated parameters, supporting context lengths up to 1M tokens. BF16 and NVFP4 weights are available on Hugging Face. The model supports three speculative decoding techniques—MTP, DFlash, and DSpark—and can plug into agent workflows through an OpenAI-compatible API.
NVIDIA released Nemotron 3.5 Lightning, a customizable open-source 30B mixture-of-experts (MoE) model designed for resident AI agents. Compared with similar open-source models, its token generation speed is up to 4x faster and task completion time is reduced by 30%. The model uses open weights, supports task-specific fine-tuning, and runs on RTX PCs, DGX Spark, and Jetson devices.
Ant Ling (Bailing) has open-sourced Ling-3.0-tiny, a native hybrid reasoning model with 7.9B total parameters that activates only 1.3B parameters during inference, and simultaneously releases three versions in BF16, FP8 and INT4.
SGLang, in collaboration with Meta Superintelligence Labs, provides day-zero support for Muse Glimmer, a 30B-parameter multimodal model with a 128k+ token context window.
Meta introduced Muse Glimmer, an open-weight 30B-parameter model optimized for local, always-on agent workflows. It outperforms leading same-size models on key agent benchmarks and standardized evaluations, and is designed to run entirely on consumer hardware such as Macs or PCs with high-performance GPUs, released under the permissive Apache 2.0 license.
Scale AI announced it will soon release open-weight versions of Muse Spark 1.2, and is releasing Muse Glimmer—a 30B-parameter agent model under the Apache 2.0 license. Muse Glimmer runs on 24GB of VRAM without losing agent reliability.
OpenAI released GPT-5.6-Cyber, a cybersecurity-specific model available via Daybreak Red, for authorized vulnerability research, exploit validation, and security testing. The model addresses the shrinking cyber defense window and gives security researchers a dedicated toolset.
NVIDIA NemotronLabs open-sourced VoiceChat-11B, an end-to-end full-duplex speech-to-speech model with about 448ms turn-taking latency and live tool calling. It runs on a single 80GB GPU under a permissive license, though NVIDIA labels the checkpoint research-only.
Ant Group's Lingwan lab has officially open-sourced Ling-3.0-flash, a new-generation native hybrid-reasoning model. It adopts a MoE architecture with 124B total parameters and 5.1B activated parameters, and offers multiple quantization versions including FP8, FP4, and INT4. It supports three modes—API calls, single-machine deployment, and high-performance clusters—balancing inference cost and effectiveness.
Google DeepMind, with several institutions, launched WeatherNext Cyclones, a cyclone-forecast model that leads the industry in track, intensity, and wind-field structure accuracy. The model extends the useful forecast horizon from 2 days to 3 days, an average of 24 hours earlier, with an improvement roughly equivalent to a decade of meteorological progress.
NVIDIA released Cosmos 3, an open physical-AI foundation model built on a hybrid AI core architecture that integrates visual reasoning, world generation, and action prediction.
ChatGPT launched an improved GPT-5.6 Sol with better accuracy and consistency, while expanding free-user access. Free users can also use GPT-5.6 Luna for unlimited everyday conversations.
NVIDIA has released Alpamayo 2 Super, a 34B-parameter vision-language-action (VLA) model designed specifically for long-tail events in autonomous driving. The weights are released under the Linux Foundation's OpenMDW-1.1 license, while the code is under Apache 2.0, making it commercially usable from day one.
Alibaba's Tongyi Qianwen has released Qwen-Image-3.0-Pro and Standard, now available on Qwen Cloud. The model ranks first among Chinese models and second among mainstream models on the Arena text-to-image leaderboard, supporting 4.5k-token prompts, 10px-level text rendering, and 12 languages. Pro starts at $0.04 per image, while Standard starts at $0.03 per image.
FLUX 3 Video from Black Forest Labs is now open to everyone on OpenRouter — a unified family covering video, audio, image and motion prediction, jointly trained on a single architecture.
ByteDance Seed has released SeedRealtime, which natively fuses audio, video, and text within a unified architecture to achieve real-time interaction of "seeing, listening, and speaking" at the same time. Compared with cascaded models, it halves the rhythm problems in audio-video conversations, and has been fully rolled out on the Doubao App, pioneering the large-scale deployment of audio-video full-duplex technology.
SenseTime released SenseNova U1, a publicly available, free-to-use model that performs reasoning and image generation in a single unified pipeline. Its infographic mode turns one prompt into structured slides, while its interleaved mode generates text and images step by step, as shown in a six-step dragon-drawing demo. The model is live on SenseNova Studio, HuggingFace and GitHub.
Ant Group's Linghu team (AntLingAGI) released Ling-3.0-flash with publicly available code and free-to-use weights. Official BF16 and FP8 compressed model versions are now available, so users can choose the most suitable version based on their hardware, performance requirements, and deployment needs.
NVIDIA Alpamayo 2 Super is now commercially available, built on Cosmos 3 Super Reasoner, using trial-and-error post-training for trajectory prediction, causal reasoning, meta-actions, auto-labeling and VQA.
Tencent Hunyuan releases Hy ASR 3.0 preview, a MoE-based ASR model with WER of 3.34% (Mandarin), 2.62% (English) and 3.12% (Cantonese), supporting context correction, hot-word injection and noisy-whisper scenarios; available on Tencent Cloud and Yuanbao.
SenseTime launched SenseNova U1.5-Lite-Preview, a lightweight native unified multimodal model on the NEO-Unify architecture that reaches closed-source quality at just 8B MoT parameters.
MiniMax has open-sourced H3, a general video model that unifies text, image, video and audio understanding and generates up to 2K, 15-second clips with native 32 kHz stereo audio.
Qwen launched Qwen3.8-Max, its strongest model yet, with 2.4T parameters and 95B active, and will open-source Qwen-Max-level weights next week for the first time.
DeepSeek released the open-weight model DeepSeek V4 Flash 0731, scoring 50 on the Artificial Analysis Intelligence Index and ranking among the top three open, freely usable models. The model uses the MIT license, with 284B total parameters (13B active) and roughly 167GB in FP4/FP8 mixed precision, matching the V4 Flash architecture and pricing, and is now available on the official API.
MiniMax officially launched H3, an all-in-one multimodal generation model that understands text, images, video and audio, and generates video up to 2K resolution, 15 seconds long with native stereo sound. H3 excels at instruction following, text and brand rendering, and V2V motion transfer. At 2K its per-second price is below one third of mainstream models, and at 768p below half of mainstream 720p pricing. The team plans to open-source the weights soon to support the community and accelerate hardware compatibility.
DeepSeek-V4-Flash's official API is now in public beta. The team has greatly upgraded its agent capabilities—standardized agent benchmark scores now far exceed V4-Pro-Preview. The official V4-Flash now natively supports the Responses API format and is fully adapted to Codex. See the configuration details in the official API documentation.
Google DeepMind has introduced Gemini Robotics 2, a next-generation physical AI described as "One brain. For any robot." It brings full-body intelligence, advanced dexterity, and multi-robot team collaboration to humanoid robots.
Google DeepMind has introduced Gemini Robotics ER 2, a Gemini-based robot foundation model for embodied reasoning. It delivers step-change gains in video understanding, tool orchestration, and multi-robot collaboration, enabling robots to reason, cooperate, and solve real-world tasks.
OpenAI has introduced lower pricing for the Luna and Terra variants of GPT-5.6, using more efficient models to help enterprises deploy AI workflows at scale.
Google DeepMind has released Lyria 3.5, a new music generation model inside Google Flow Music, with improvements across musicality, lyric quality, vocal expressiveness, and creative control. It produces more natural and complex melodic structures, follows prompts with stronger structural awareness, renders more realistic and emotional vocals, and offers easier control over output rhythm and duration.
OpenAI introduced two transcription models in its API: GPT-Live-Transcribe for low-latency real-time transcription and GPT-Transcribe for asynchronous batch processing of finished audio. Both show stronger contextual understanding and improved accuracy across accents, languages, numbers, and noisy speech, sitting between Whisper and Gemini.
