Model Launches

MiniMax ships M3, an open-weight frontier model built on its new MSA (MiniMax Sparse Attention) architecture — a KV-block selection scheme that cuts per-token compute at long context — carrying a 1M-token context window, native multimodality, and built-in agentic coding. JD Cloud was first to host it via an OpenAI-compatible API.

Alibaba Cloud ships Qwen 3.7 Plus, the multimodal sibling to the text-only Qwen 3.7 Max. It adds image and video understanding and is tuned to balance power and speed for everyday agentic tasks. Closed-weights and API-only, the preview pushed Alibaba to a top-5 lab on LM Arena's Vision leaderboard.

Anthropic officially launches Claude 4.8 Opus, its ultimate frontier model optimized for long-horizon autonomous system execution, repository-level software engineering, and multi-step complex logic.

Alibaba Cloud unveils Qwen 3.7 Max, a proprietary, API-only 'reasoning agent' model built for long-horizon autonomous work rather than chat. With 1T+ parameters and a 1M-token context window, it ranks 5th globally on Artificial Analysis's Intelligence Index (56.6) — the highest-ranked Chinese model — and is served only through Qwen Studio and Alibaba Cloud Model Studio.

Google releases Gemini 3.5 Flash in general availability, showcasing significant speedups in lightweight tool-calling capabilities and ultra-low-latency workflows.

xAI launches Grok 4, a major foundation upgrade incorporating advanced real-world visual logic and deep real-time tool execution, trained on the expanded Memphis 200k B200/H100 cluster.

DeepSeek launches V4, a native multi-modal reasoning Mixture-of-Experts (MoE) model matching frontier commercial performance at an unprecedented low cost.

OpenAI ships GPT-5.5, introducing key updates targeting multi-agent orchestration, complex coding pipelines, and long-horizon tool execution.

Moonshot AI launches Kimi K2.6, featuring advanced long-context coding and an elevated 'Agent Swarm' system coordinating up to 300 sub-agents.

Anthropic releases Claude Opus 4.7 as its flagship model, delivering advanced multi-step tool use, elite coding, and top-tier logical reasoning.

Meta launches Muse Spark, a proprietary 4D world model trained on first-person video and spatial data for advanced situational reasoning.

Anthropic announces Claude Mythos under a gated private preview, focusing on defensive cybersecurity, automated exploit analysis, and mechanistic interpretability.

Alibaba Cloud launches the Qwen 3.6 family, bringing deep structural optimizations for agentic programming and reasoning workloads.

Zhipu AI releases GLM-5.1, bringing logical refinement and expanded coding context for enterprise Chinese API users.

Zhipu AI rolls out GLM-5-Turbo, an optimized, fast-inference model tailored for 'OpenClaw' autonomous agent platforms.

MiniMax releases M2.7, offering expanded context windows and native multi-modal image-to-text generation capabilities.

OpenAI releases GPT-5.4, a major unified flagship update incorporating critical logical execution and context pipeline improvements from the 5.3 branch.

OpenAI rolls out GPT-5.3 Instant, a high-speed, cost-optimized conversational and logical execution model.

MiniMax launches M2.6, bridging agentic text capability and early multi-modal context features, seeing rapid pick-up on OpenRouter developer platforms.

Alibaba Cloud releases Qwen 3.5, engineered for high-performance agentic MoE configurations (397B-A17B flagship MoE).

ByteDance officially launches Seedance 2.0, a unified multi-modal video/audio generation model generating highly realistic, synchronized cinema clips.

MiniMax launches M2.5, which rapidly climbs OpenRouter popularity charts, demonstrating high adoption for agentic text scenarios.

Zhipu AI releases GLM-5, a flagship 744B-parameter Mixture-of-Experts (MoE) model trained entirely on sovereign Chinese hardware.

Google launches Gemini 3.1 Pro, a major agentic reasoning update to its flagship model series. Delivering native deep search execution, reinforcement-learning based planning, and substantial code-generation gains.

OpenAI launches GPT-5.3-Codex, establishing elite new milestones in multi-agent logical coordination and deep code architecture mapping.

Anthropic launches Claude Opus 4.6, delivering optimization bumps for agentic reasoning and complex coding workloads.

Moonshot AI launches Kimi K2.5, a 1-trillion parameter multimodal Mixture-of-Experts model powered by an advanced 'Agent Swarm' system.

Zhipu AI rolls out GLM-4.7, delivering major logic optimization, expanded context limits, and complex multi-agent orchestration tools.

Anthropic rolls out Claude Opus 4.5, a mid-generation bump focused on complex long-context reasoning and reduced hallucination rates.

Google introduces Gemini 3 Pro, focusing on advanced multi-step planning, long-horizon tool execution, and deeper Google Workspace integrations.

MiniMax shifts to its next-generation foundation architecture with the release of the M2 base model.

Zhipu AI launches GLM-4.6 alongside the multimodal 4.6V, establishing massive upgrades in multi-modal understanding, coding context, and reasoning capabilities.

Zhipu AI launches GLM-4.5V, a highly competitive open-weight vision-language model.

OpenAI launches GPT-5, a next-generation frontier multimodal model establishing new milestones in complex logic, native video processing, and abstract reasoning.

Zhipu AI releases GLM-4.5 alongside the high-speed GLM-4.5-Air model, optimizing API cost and latency.

Moonshot AI launches the foundational Kimi K2 model, establishing a massive baseline context size for long-form document processing.

Google ships Gemini 2.5 Flash, bringing major architectural speed and multi-lingual reasoning improvements to its widely used efficiency model.

ByteDance introduces Seedance, its first-generation unified video foundation model, focusing on realistic character physics.

Anthropic unveils the Claude 4 generation, launching both the flagship Claude Opus 4 and the balanced Claude Sonnet 4.

Alibaba Cloud releases the Qwen 3 model family, showing dramatic logical gains and native multi-lingual reasoning benchmarks.

OpenAI releases the full flagship o3 reasoning model, breaking previous records in complex physics, chemistry, coding, and mathematical benchmarks.

Meta officially launches Llama 4 Scout (109B parameter MoE) and Llama 4 Maverick (400B parameter MoE), featuring advanced open-weight architectures.

Google launches Gemini 2.5 Pro Experimental, focusing on advanced coding reasoning and complex multi-modal analysis.

Anthropic releases Claude 3.7 Sonnet, introducing native hybrid-reasoning capabilities, allowing users to toggle between fast-inference and deep-thinking modes.

xAI officially launches Grok 3, demonstrating breakthrough mathematical logic and long-horizon planning performance, trained on the 100k liquid-cooled Colossus cluster in Memphis.

OpenAI releases o3-mini, providing low-latency reasoning speeds and competitive coding capabilities for developers.

Moonshot AI releases Kimi K1.5, claiming parity with OpenAI o1 across mathematical, logical, and coding benchmarks.

DeepSeek launches R1, an open-source reasoning model that matches top Western proprietary APIs at an extreme fraction of the cost, disrupting global tech markets.

DeepSeek launches DeepSeek-V3, a massive 671B parameter Mixture-of-Experts (MoE) flagship model. Delivering state-of-the-art open-source logic at commodity pricing with Multi-head Latent Attention (MLA) and FP8 training.

Meta releases Llama 3.3, upgrading the 70B model size to Llama 3.1 405B intelligence levels at significantly reduced inference cost.

OpenAI releases the complete o1 model family, integrating deep reasoning capabilities directly into its flagship API and ChatGPT platforms.

Anthropic releases Claude 3.5 Haiku globally, replacing the older Claude 3 Haiku model with significantly sharper coding skills.

Anthropic releases an upgraded version of Claude 3.5 Sonnet ('v2') alongside a public beta of 'Computer Use' agentic functionality.

Meta releases Llama 3.2, introducing multimodal vision capabilities (11B/90B) and ultra-lightweight models (1B/3B) for edge devices.

Alibaba Cloud releases Qwen 2.5, establishing a highly popular open-source coding, mathematics, and multilingual benchmark family.

OpenAI unveils its first reinforcement-learning reasoning models, releasing o1-preview and o1-mini to ChatGPT Plus and API tiers.

Meta releases Llama 3.1, introducing the massive 405B parameter flagship alongside 8B/70B upgrades, expanding the context window to 128k.

OpenAI launches GPT-4o-mini as a high-speed, cost-effective multimodal model to completely replace GPT-3.5 Turbo.

Anthropic releases Claude 3.5 Sonnet, establishing new benchmarks in software engineering, complex logical reasoning, and chart parsing.

Alibaba Cloud releases Qwen 2, providing multilingual open-weights models in five size variations.

DeepSeek releases DeepSeek-V2, introducing Multi-head Latent Attention (MLA) and slashing API pricing down to commodity levels.

Google rolls out Gemini 1.5 Flash at Google I/O, optimized for speed and cost while retaining the signature 1-million token context window.

OpenAI launches GPT-4o, natively integrating real-time audio, vision, and text processing in a single fast-inference model.

Meta launches Llama 3 in 8B and 70B sizes, showing huge logical gains over Llama 2 via high-quality dataset curation.

MiniMax releases abab 6.5, bringing reasoning improvements to its proprietary MoE API line.

OpenAI releases the GA version of GPT-4 Turbo with vision (`gpt-4-turbo-2024-04-09`), increasing context logic and performance.

Anthropic makes Claude 3 Haiku generally available, completing its initial Claude 3 roll out with a highly efficient model.

Anthropic launches the Claude 3 family, introducing the flagship Claude 3 Opus alongside the balanced Claude 3 Sonnet.

Google launches Gemini 1.5 Pro in limited preview, introducing the first native 1-million token context window in LLM history.

Alibaba Cloud releases Qwen 1.5, offering highly competitive open models in diverse parameter sizes.

Zhipu AI rolls out GLM-4, offering a competitive commercial API matching early GPT-4 benchmarks.

MiniMax releases abab 6, establishing itself as the first Chinese lab to deploy a Mixture-of-Experts architecture.

Google introduces Gemini 1.0 in Ultra, Pro, and Nano tiers, its first native multimodal architecture designed from scratch.

OpenAI announces GPT-4 Turbo in preview (`gpt-4-1106-preview`), raising context to 128k, dropping prices, and adding JSON mode.

Meta releases Llama 2 in partnership with Microsoft, offering open weights for commercial and research use across 7B, 13B, and 70B sizes.

Anthropic launches Claude 2, featuring a public web interface and a massive 100k token context window.

OpenAI releases GPT-4, setting the ultimate gold standard for reasoning, coding, professional exams, and complex instructions.

Anthropic officially launches its first commercial model, Claude 1, emphasizing safety and helpfulness via Constitutional AI.

Meta releases its first series of open models (7B to 65B) to researchers, whose weights are quickly leaked to the public.

OpenAI launches ChatGPT, a free conversational interface built on the InstructGPT (GPT-3.5) series.

OpenAI rolls out InstructGPT models, utilizing RLHF to align model outputs with human instructions.

OpenAI releases the commercial API private beta for GPT-3 (175B parameters), showcasing early in-context learning.