Kimi K3.1 Pro is Moonshot AI's flagship open-weight multimodal reasoning model and a post-trained successor to Kimi K3, with a 1M-token context window. It is built for long-horizon coding, 3D creation, financial research, and end-to-end agentic knowledge work, with native image and video understanding.
DeepSeek V4.1 Flash 1017 is the October 17, 2026 snapshot of DeepSeek V4.1 Flash, served by DeepSeek under the deepseek-flash API model name. V4.1 Flash is a sparse mixture-of-experts model from DeepSeek, and the first built on the company's Causal Encoder-Decoder (CED) architecture. It activates 8B parameters on input and 16B on output from a 552B-parameter backbone, an asymmetric split that keeps per-token compute low relative to the model's total size. Image understanding is native to the architecture, with visual and text embeddings trained jointly from the start of pre-training rather than added afterward as in the earlier experimental V4 Flash Vision Exp.
GLM-5.3-Flash is the first native multimodal model in the GLM-5 series, delivering stronger intelligence than GLM-5.2 while maintaining an exceptionally cost-efficient architecture
GLM-5.3-FlashX is a high-speed native multimodal LLM by Z.AI. Using a sparse-linear attention hybrid architecture with 320B parameters (18B active), it supports a 1M context window and speeds up to 200 Tokens/s. It integrates multimodal perception and a visual-coding feedback loop for autonomous agents and engineering.
DeepSeek V4.1 Flash is a next-gen native multimodal MoE model released in September 2026. Featuring an innovative CED asymmetric architecture and CSA2 mechanism, it slashes KV cache storage to 1/4. Offering Flash-level low pricing and speed, it surpasses the flagship V4 Pro in coding and Agent benchmarks, democratizing ultra-long context computing.
Qwen3.8 Max is the flagship model in Alibaba's Qwen3.8 series, the general-availability successor to the Qwen3.8 Max Preview. It is a multimodal reasoning model intended for complex reasoning, visual understanding, coding, and agentic workflows.
Wan3.0-Video-Prime is Alibaba's high-speed video generation model, combining the full multimodal capabilities of the Wan 3.0 series with significantly accelerated generation speeds. It is engineered for efficient enterprise production, delivering long-duration, high-fidelity video and native audio.
Qwen3.8 Flash is a multimodal reasoning model from Alibaba. It is suited for coding assistance, agentic workflows, visual understanding, document and codebase analysis, desktop interaction, chart analysis, and long-video analysis.
Wan 3.0 is Alibaba’s state-of-the-art video generation model, featuring 30-second single-pass rendering, native audio generation, and multi-format document inputs (PDF, DOCX, etc.).
Released in August 2026, GLM-5.3 is Z.ai's 743B-parameter open-weights flagship model. Relying purely on post-training scaling with a 1M context window, it delivers open-source SOTA in long-horizon agentic tasks and complex software engineering. It boasts breakthrough emergent cybersecurity capabilities, closely rivaling closed frontier systems like Claude Fable 5.
DeepSeek V4 Pro 0813 is a large-scale mixture-of-experts model from DeepSeek. This is the GA release of DeepSeek V4 Pro.
Qwen3.8 Max is the flagship model in Alibaba's Qwen3.8 series, the general-availability successor to the Qwen3.8 Max Preview. It is a multimodal reasoning model intended for complex reasoning, visual understanding, coding, and agentic workflows.
MiniMax H3 is an omni-modal generative foundation model featuring a unified context architecture. It natively processes text, image, video, and audio as a single cohesive stream, allowing a single forward pass to jointly generate high-fidelity video synchronized with native stereo audio (encompassing speech, sound effects, and music). This breaks the siloed boundaries of traditional generation pipelines, delivering a disruptive breakthrough for commercial visual production, advertisement asset creation, and gaming LiveOps.
DeepSeek V4 Flash 0731 is a sparse mixture-of-experts model from DeepSeek, with 13B active parameters out of 284B total. This re-post-trained revision is suited for coding, reasoning, and agent workflows. This is the GA release of DeepSeek V4 Flash.
Seedance 2.5 is ByteDance’s flagship AI video model launched in July 2026. It doubles single-pass runtime to a 30-second native long take in 4K without stitching. Supporting up to 50 multimodal references (30 images, 10 videos, 10 audio) with a 20% boost in prompt adherence, it introduces region-level editing and 3D white-model control to lock character and lighting consistency for professional video production.
Kimi K3 is a powerful 2.8-trillion-parameter open-weight multimodal reasoning model developed by Moonshot AI. Featuring a massive 1-million-token context window and native vision capabilities, it is specifically engineered for long-horizon coding, complex software engineering, and advanced agentic knowledge work. By utilizing innovative architectures like Kimi Delta Attention, it delivers exceptional scaling efficiency and autonomous problem-solving performance.
Qwen3.7-Flash is a lightweight, ultra-fast large language model open-sourced by Alibaba. It delivers high speed and low latency while maintaining excellent performance in multilingual understanding, coding, and logical reasoning, making it ideal for real-time chat and high-concurrency enterprise applications.
Seedream 5.0 Pro is ByteDance’s flagship multimodal image model built for professional design. Moving beyond basic generation, it excels in complex information visualization, high-density infographic layout, and photorealistic portrait texture. Its standout native layer-based architecture supports pixel-level interactive precision editing and locale-aware multilingual text rendering.
GLM-5.2 is Z.ai's open-weight, flagship 744B-parameter Mixture-of-Experts (MoE) model engineered for long-horizon software engineering and complex agentic workflows. It features a robust 1,000,000-token context window alongside adjustable "Thinking Effort" modes to optimally balance reasoning performance against execution speed. Boasting SOTA benchmarks on Terminal-Bench and SWE-bench Pro, it delivers near-proprietary, frontier-level coding intelligence under an open-source MIT license.
Seedance 2.0 Mini is a lightweight, cost-effective AI video generation model launched by ByteDance on August 12, 2026. As an affordable variant of the flagship series, it is optimized for high-volume social media and marketing video production. Supporting native "four-modality" reference inputs (up to 12 concurrent assets across text, images, videos, and audio), the model drastically reduces inference costs and latency while delivering exceptional character consistency and cinematic camera controls.
Kimi-K2.7-Code is Moonshot AI's open-weight, 1-trillion parameter Mixture-of-Experts (MoE) model meticulously optimized for software engineering and agentic workflows. Supporting a 256K context window with native vision-text processing, it operates exclusively in long-thinking reasoning mode. It significantly minimizes overthinking—cutting thinking-token usage by 30%—while dramatically boosting multi-turn tool calling, instruction compliance, and success rates in complex, long-horizon coding tasks.
Qwen3.7-Plus is Alibaba's proprietary multi-modal Hybrid-Agent model powered by All-field Thinking architecture. Supporting a 1,000,000-token context window and 65,536-token output, it seamlessly processes text, image, and video inputs. By dynamically toggling between deep reasoning and standard conversation, it tackles complex logical tasks with high accuracy. It delivers near-flagship intelligence matching Qwen3.7-Max but operates at a fraction of the cost, making it ideal for deploying scalable, low-latency AI agents.
MiniMax M3 is the flagship open-source MoE large language model developed by MiniMax. Built on a proprietary Modular Sparse Attention (MSA) architecture, it holds 428 billion total parameters while activating only 23 billion. The model natively supports a 1-million-token context window and delivers top-tier performance in full-stack advanced coding, long-horizon agent collaboration, and cross-modal comprehension, offering global developers exceptional cost-efficiency.
Qwen3.7-Max is Alibaba's closed-source, flagship AI model engineered for the agentic era. Supporting a 1,000,000-token context window, it excels at complex coding, advanced office automation, and deep scientific reasoning. Designed as a versatile foundation for multi-agent workflows, it delivers elite long-horizon autonomy, sustaining multi-hour autonomous execution across thousands of consecutive steps. It sets SOTA benchmarks in complex logic without requiring framework-specific tuning. [1, 2, 3, 4, 5, 6, 7]
DeepSeek V4 Flash is an efficiency-optimized Mixture-of-Experts model from DeepSeek with 284B total parameters and 13B activated parameters, supporting a 1M-token context window. It is designed for fast inference and high-throughput workloads, while maintaining strong reasoning and coding performance. The model includes hybrid attention for efficient long-context processing. Reasoning efforts high and xhigh are supported; xhigh maps to max reasoning. It is well suited for applications such as coding assistants, chat systems, and agent workflows where responsiveness and cost efficiency are important.
DeepSeek V4 Pro is a large-scale Mixture-of-Experts model from DeepSeek with 1.6T total parameters and 49B activated parameters, supporting a 1M-token context window. It is designed for advanced reasoning, coding, and long-horizon agent workflows, with strong performance across knowledge, math, and software engineering benchmarks. Built on the same architecture as DeepSeek V4 Flash, it introduces a hybrid attention system for efficient long-context processing. Reasoning efforts high and xhigh are supported; xhigh maps to max reasoning. It is well suited for complex workloads such as full-codebase analysis, multi-step automation, and large-scale information synthesis, where both capability and efficiency are critical.
Kimi K2.6 is Moonshot AI's next-generation multimodal model, designed for long-horizon coding, coding-driven UI/UX generation, and multi-agent orchestration. It handles complex end-to-end coding tasks across Python, Rust, and Go, and can convert prompts and visual inputs into production-ready interfaces. Its agent swarm architecture scales to hundreds of parallel sub-agents for autonomous task decomposition - delivering documents, websites, and spreadsheets in a single run without human oversight.
Seedance 2.0 Fast is ByteDance’s speed-optimized multimodal AI video model designed for rapid turnaround times. It natively integrates text, image, and audio inputs in a single step, co-generating cinematic video alongside perfectly synchronized sound effects and dialogue.
Qwen 3.6 Plus builds on a hybrid architecture that combines efficient linear attention with sparse mixture-of-experts routing, enabling strong scalability and high-performance inference. Compared to the 3.5 series, it delivers major gains in agentic coding, front-end development, and overall reasoning, with a significantly improved “vibe coding” experience. The model excels at complex tasks such as 3D scenes, games, and repository-level problem solving, achieving a 78.8 score on SWE-bench Verified. It represents a substantial leap in both pure-text and multimodal capabilities, performing at the level of leading state-of-the-art models.
MiniMax-M2.7 is an open-weight, 230-billion parameter Mixture-of-Experts (MoE) model featuring a 205K context window, specifically engineered for autonomous productivity. Driven by a breakthrough recursive self-evolution pipeline, it achieved a 30% performance boost through autonomous optimization. It excels at multi-agent collaboration, multi-round complex office suite editing, and system-level software engineering, scoring a SOTA 1495 ELO on GDPval-AA and an impressive 56.22% on SWE-Pro.
Qwen3.5-Plus is Alibaba Cloud's flagship MoE large language model, accessible via Model Studio. With 397B total and 17B active parameters, it balances top-tier performance with low cost. It features a 1-million-token context window, excelling in complex reasoning, advanced coding, and agent workflows. Supporting 201 languages, it is built for global enterprise scale.
Seedream 5.0 Lite is a highly efficient, lightweight multi-modal model optimized for edge devices and fast inference. It delivers powerful text and vision capabilities with minimal compute requirements, making it ideal for responsive on-device AI applications.
MiniMax-M2.5 is an open-weight, 230-billion parameter Mixture-of-Experts (MoE) model that activates only 10 billion parameters per token. Designed natively for agentic workflows, it features a 200K context window and achieves an impressive 80.2% on SWE-bench Verified. Driven by large-scale reinforcement learning, it delivers elite performance in full-stack coding, multi-turn tool calling, and office document generation while maintaining superior cost-efficiency.
GLM-5.1 delivers a major leap in coding capability, with particularly significant gains in handling long-horizon tasks. Unlike previous models built around minute-level interactions, GLM-5.1 can work independently and continuously on a single task for more than 8 hours, autonomously planning, executing, and improving itself throughout the process, ultimately delivering complete, engineering-grade results.
Seedance 2.0 is a revolutionary audio-video joint generation model launched by ByteDance in February 2026. Pioneering quad-modal unified input, it seamlessly fuses text, images, video, and audio references. Delivering Hollywood-grade cinematic quality and precise camera control, it achieves industry-first native audio and accurate lip-sync output in a single pass, completely reshaping multi-scene narrative workflows.
Kimi K2.5 is Moonshot AI's native multimodal model, delivering state-of-the-art visual coding capability and a self-directed agent swarm paradigm. Built on Kimi K2 with continued pretraining over approximately 15T mixed visual and text tokens, it delivers strong performance in general reasoning, visual coding, and agentic tool-calling.