模型

Inkling

Thinking Machines Lab's open-weights general-purpose multimodal Mixture-of-Experts model with 975B total parameters and 41B active parameters. Inkling accepts text, image, and audio inputs, produces text, and is designed for agentic and tool-use systems, coding assistants, chatbots, and retrieval-augmented generation.

Kimi K3

Moonshot AI's flagship Kimi model for frontier intelligence, agentic coding, knowledge work, and deep reasoning. Kimi K3 supports a 1-million-token context window for long-running software engineering and research workflows.

GPT-5.6 Luna

The fast, low-cost tier of OpenAI's GPT-5.6 series, optimized for high-volume, latency-sensitive tasks such as classification, extraction, routing, and lightweight agentic steps. Approaches the larger GPT-5.6 tiers on many benchmarks while running several times faster at a fraction of the price.

GPT-5.6 Sol

OpenAI GPT-5.6 系列旗舰模型，在提升可靠性和效率的同时，推进了编程、科学推理、长周期规划和智能体工作流能力。新增最高推理强度设置，以及可为复杂多步骤任务启动子智能体的 ultra 模式。

Muse Spark 1.1

Meta Superintelligence Labs' updated flagship, building on Muse Spark with stronger agentic reasoning, more reliable multi-agent orchestration, and improved multimodal understanding across voice, text, and image. Extends the context window and reduces latency and reasoning token usage while raising coding and tool-use accuracy. Powers Meta AI across its product ecosystem.

GPT-5.6 Terra

OpenAI GPT-5.6 系列的均衡层级，以少量峰值质量换取显著更低的延迟和成本。它保留了强大的推理、编程和智能体工具使用能力，并支持可配置的推理强度，适合作为需要大规模前沿能力的生产工作负载的默认选择。

Grok 4.5

xAI 迄今最强的模型，面向编程、智能体任务和知识工作，并与真实软件工程中的编程工具协同开发。提供实时信息访问、扩展推理和大上下文工具调用，并兼容 OpenAI API。

Claude Sonnet 5

Anthropic 能力最强的 Sonnet 级模型，以更低价格将前沿编程、智能体和专业工作能力带到中型层级，并缩小了与 Opus 4.8 的差距。支持可选推理强度的自适应思考、100 万 token 上下文窗口，以及文本、图像和文件输入。代号 Fennec。

Sakana Fugu Ultra

The higher-quality tier of Sakana AI's Fugu multi-agent orchestration system, tuned for the hardest coding, reasoning, science, and agentic tasks. Coordinates a swappable pool of frontier LLMs through one OpenAI-compatible endpoint, delegating sub-tasks, verifying intermediate work, and synthesizing a single answer. Sakana reports strong vendor benchmarks including 93.2 on LiveCodeBench, 73.7 on SWE-Bench Pro, and 82.1 on TerminalBench.

Sakana Fugu

Sakana AI's multi-agent orchestration model from Tokyo, delivered as a single OpenAI-compatible API. Fugu is itself a language model trained to call a pool of specialist LLMs (and recursive instances of itself), handling model selection, delegation, verification, and synthesis behind one endpoint. Built on Sakana AI's TRINITY and Conductor research, its routing intelligence is learned in model weights rather than hand-configured.

Sarvam-105B

Sarvam AI's sovereign 105B-parameter Mixture-of-Experts model activating ~9B parameters per token, with a 128K-token context window. Trained on 12 trillion tokens across 22 Indian languages using 128 sparse experts with Multi-head Latent Attention and a custom low-fertility Indic tokenizer. Wins the majority of pairwise comparisons on Indian-language and STEM benchmarks.

GLM-5.2

Z.ai's (formerly Zhipu AI) flagship open-weight coding model with a 1M-token context window. Mixture-of-Experts architecture with 753B total parameters and ~40B active per request, featuring two cost-balancing reasoning modes. Tops several coding benchmarks while remaining a fraction of the cost of comparable proprietary frontier models. MIT-licensed weights.

Sarvam-30B

Sarvam AI's 30B-parameter Mixture-of-Experts reasoning model trained from scratch with only 2.4B active parameters per token. Optimized for real-time deployment and Indian languages, delivering strong reasoning, coding, and conversational performance while remaining efficient to serve. Open-weights.

Sarvam-M

Sarvam AI's 24B-parameter instruction-tuned model derived from Mistral-Small-3.1-24B, post-trained on English plus eleven major Indic languages (bn, hi, kn, gu, mr, ml, or, pa, ta, te). Delivers large relative gains on Indian-language, math, and programming benchmarks over its base model, with a hybrid reasoning mode for complex tasks.

Command A+

Cohere's enterprise flagship model building on Command A with stronger reasoning, agentic tool use, and multilingual performance across 23 languages. Optimized for secure, high-throughput RAG, retrieval, and long-horizon agent workflows in regulated environments, with private and on-premise deployment options.

Kimi K2.7 Code

Moonshot AI's latest open-source, coding-focused model in the Kimi K2 family, built to complete end-to-end programming tasks reliably over long contexts. A 1-trillion-parameter model that cuts reasoning token usage by roughly 30% versus K2.6 while improving coding and agent performance — +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, and +31.5% on MLS Bench Lite for multi-language support. Released under a Modified MIT License and available via Kimi APIs and Hugging Face.

DiffusionGemma

Google DeepMind's experimental diffusion-based member of the Gemma 4 open model family. Unlike autoregressive models that generate text one token at a time, DiffusionGemma denoises a canvas of placeholder tokens to produce up to 256 tokens in parallel, finalizing output in one block. A Mixture-of-Experts model with 26B total parameters and 3.8B active per inference, delivering roughly 4x the throughput of similarly sized autoregressive Gemma models on local hardware. Excels at non-linear tasks like in-line editing, molecular sequencing, mathematical graphing, and self-correcting puzzles.

Gemma 4 12B

Google's medium-size open-weight model with 12 billion parameters from the Gemma 4 family. Encoder-free unified multimodal architecture that natively processes text, image, audio, and video inputs without dedicated encoders. Features a 256K context window and supports 140+ languages. First medium-sized model capable of natively ingesting audio. Suitable for local deployment on GPUs with 16GB VRAM.

Nemotron 3 Ultra

NVIDIA's flagship open 550B-parameter Mixture-of-Experts model with 55B active parameters, built for frontier reasoning and orchestration in long-running agentic systems. Features hybrid Mamba-Transformer architecture, LatentMoE routing, multi-token prediction, and NVFP4 precision for 5x higher throughput. Achieves 30% lower cost-to-task-completion on agentic benchmarks. Supports 1M+ token context window with 95% accuracy on Ruler@1M.

MiniMax M3

MiniMax's frontier open-weight model with 1M-token context window, native multimodality (text, image, video), and strong coding capabilities. Built on MiniMax Sparse Attention (MSA) architecture, achieving 59% on SWE-Bench Pro with significantly improved efficiency at long context.

Yi-Lightning

01.AI's flagship large language model with enhanced Mixture-of-Experts architecture. Ranked 6th on Chatbot Arena with particularly strong results in Chinese, Math, Coding, and Hard Prompts categories. Features advanced expert segmentation and optimized KV-caching.

Falcon-H1

TII's hybrid Mamba-Transformer model that outperforms comparable offerings from Meta's Llama and Alibaba's Qwen in the 30-70B parameter range. Designed for real-world AI on everyday devices and resource-limited settings with state-of-the-art efficiency.

Jamba Large 1.7

AI21's latest hybrid SSM-Transformer model with Mixture-of-Experts architecture. Features a 256K context window, improved grounding and instruction-following. 94B total parameters with 398B active, optimized for enterprise long-context tasks.

Ring-2.6-1T

InclusionAI's (Ant Group) trillion-parameter open-weights reasoning model with 63B active parameters per token. Built for real-world agent workflows with adaptive reasoning-effort modes. Features hybrid linear and MLA attention architecture with MIT license.