DeepInfra

Kimi K3 is Moonshot AI's open-weight Mixture-of-Experts model with 2.8T total and 104B active parameters and a 1M-token context window. It accepts text, image and video input and produces text, always reasons with low, high or max effort, and is built for long-horizon coding, knowledge work and reasoning.

MiniMax-M3 is MiniMax's natively multimodal Mixture-of-Experts model with about 428B total and 23B active parameters. It takes text, image and video input and produces text, with a context window of up to 1M tokens, and can also operate a desktop computer.

Claude Fable 5 is Anthropic's Mythos-class model released for general use, built for demanding reasoning and long-horizon agentic work. It takes text and image input and produces text output, with a 1M-token context window, up to 128K output tokens and always-on adaptive thinking.

GLM-5.3 is a large-scale reasoning model from Z.ai, built for complex software engineering and long-horizon agent tasks. It supports text input and output with a 1M-token context window, and improves on GLM-5.2 in coding and in the balance between performance and token efficiency.

DeepSeek V4.1 Flash is a multimodal Mixture-of-Experts model with 552B backbone parameters (8B active on input, 16B on output) and a 1M-token context window. It natively processes images and text, with strong reasoning, coding, and agentic performance.

Gemma 4 31B is the dense instruction-tuned model in Google DeepMind's open-weight Gemma 4 family, with about 30.7B parameters. It takes text and image input and produces text output, with a 256K-token context window, configurable thinking, native function calling and pretraining on over 140 languages.

gpt-oss-120b is OpenAI's open-weight Mixture-of-Experts language model with 117B total and 5.1B active parameters, released under the Apache 2.0 license. It takes text input and produces text output with a 128K-token context window, configurable low, medium and high reasoning, full chain-of-thought access and native tool use including function calling and structured outputs.

DeepSeek-V3.2 uses DeepSeek Sparse Attention for efficient long contexts, with strong reasoning and tool use.

DeepSeek V4 Flash 0731 is the July 31, 2026 official release of DeepSeek V4 Flash, a Mixture-of-Experts model with 284B total and 13B active parameters and a 1M-token context window. It keeps the preview's architecture and size but is re-post-trained for much stronger agentic performance, with low, high and max reasoning effort.

Moonshot AI's open-weight multimodal agentic model has 1T total and 32B active parameters, built for long-horizon coding and tool use.

Qwen3.5-397B-A17B is a multimodal Mixture-of-Experts model from Alibaba's Qwen team, with 397B total parameters and 17B active per token. It takes text, image and video input and produces text, with a 262K-token native context window (extendable to about 1M), a default thinking mode, tool calling and support for 201 languages and dialects.

Qwen's dense 27B vision-language model supports reasoning, coding and tool-driven tasks.

Qwen 3.8 27B is an open-weight dense model from Alibaba's Qwen team with 27B parameters and a vision encoder, using hybrid Gated DeltaNet and gated attention layers. It takes text, image and video input with a native 262K-token context window and is built for coding, professional and research work, and long-horizon agentic tasks, with adjustable thinking effort.

GLM-5.2 is Z.ai's open-weight Mixture-of-Experts model for long-horizon tasks and software engineering. It takes text input and output with a 1M-token context window and up to 128K output tokens, and offers multiple thinking effort levels to balance performance and latency.

DeepSeek V4.1 Flash is a multimodal Mixture-of-Experts model with 552B backbone parameters (8B active on input, 16B on output) and a 1M-token context window. It natively processes images and text, with strong reasoning, coding, and agentic performance.

Gemma 3 27B Instruct is the 27-billion-parameter instruction-tuned model in Google DeepMind's Gemma 3 family of open-weight multimodal models. It takes text and image input and produces text output, with a 128K-token context window and support for over 140 languages, for tasks such as chat, summarization, question answering and image understanding.

Gemma 4 26B A4B Instruct is the Mixture-of-Experts model in Google DeepMind's open-weight Gemma 4 family, with 25.2B total and 3.8B active parameters per token. It takes text and image input and produces text output, with a 256K-token context window, configurable thinking and native function calling.

Mistral Small 3.2 is a 24B open-weight model from Mistral AI that accepts text and image input, released in June 2025 as a minor update to Mistral Small 3.1. It improves instruction following, produces fewer repetitive or infinite generations, and has a more robust function-calling template.

NVIDIA Nemotron 3 Ultra is built for frontier reasoning, orchestration, coding agents, deep research, and complex enterprise workflows. It delivers up to 5x faster inference and up to 30% lower cost for agentic workloads.

Qwen 3 235B A22B Instruct 2507 is the July 2025 update of Alibaba's Qwen3 Mixture-of-Experts model, with 235B total and 22B active parameters. This variant runs in non-thinking mode only, answering directly without reasoning blocks, and has a native 262K-token context window.

Qwen 3 Coder 480B Turbo is a faster serving variant of Qwen3 Coder 480B A35B Instruct, Alibaba's Qwen team's Mixture-of-Experts coding model with 480B total and 35B active parameters. The model has a native 262K-token context window and is built for agentic coding, tool calling and repository-scale work.

Qwen 3 Next 80B is a Mixture-of-Experts model from Alibaba's Qwen team with 80B total and 3B active parameters, using hybrid Gated DeltaNet and gated attention layers with multi-token prediction for efficient long-context work. It takes text input and output with a native 262K-token context window.

Qwen 3.6 35B A3B is a fast mixture-of-experts model with 35B total parameters and ~3B active per token. Strong at agentic coding, STEM reasoning, and tool use, with a native 256K context window.

Qwen3-VL 235B A22B Instruct is Alibaba's Qwen vision-language Mixture-of-Experts model with about 236B total and 22B active parameters. It takes image, video and text input and produces text, with a native 256K-token context, OCR in 32 languages, GUI agent abilities and long-video understanding.

MiMo-V2.6-Flash is Xiaomi's efficiency-balanced omnimodal model: a sparse Mixture-of-Experts with 309B total and 15B activated parameters, natively understanding text, image, video, and audio with a 1M-token context window. Trained with mixed reinforcement learning across coding, general agents, visual, and cybersecurity tasks, it is built for long-horizon coding, tool use, and computer use.

GLM-4.6 is Z.ai's open-weight Mixture-of-Experts model with 355B total and 32B active parameters. It takes text input and output with a 200K-token context window and up to 128K output tokens, and is built for agentic coding, front-end development, reasoning, writing and role-play.

GLM-4.7 is Z.ai's open-weight Mixture-of-Experts model with 355B total and 32B active parameters. It takes text input and output with a 200K-token context window and up to 128K output tokens, and is built for agentic coding, UI generation, tool use, reasoning and writing.

The DeepSeek R1 model has undergone a minor version upgrade, with the current version being DeepSeek-R1-0528.

DeepSeek-V3, a strong Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token. To achieve efficient inference and cost-effective training, DeepSeek-V3 adopts Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, which were thoroughly validated in DeepSeek-V2.

DeepSeek V3.1 is DeepSeek's hybrid Mixture-of-Experts model with 671B total and 37B active parameters, supporting both thinking and non-thinking modes in a single model. It has a 128K-token context window and improves tool calling and agent tasks over DeepSeek V3.

DeepSeek V4 Flash (Preview) is the April 2026 preview release of DeepSeek's efficiency-focused Mixture-of-Experts model, with 284B total and 13B active parameters and a 1M-token context window. It offers a non-thinking mode and two thinking levels (high and max) for fast reasoning and coding workloads.

DeepSeek-V4-Flash-Vision-Exp is DeepSeek's experimental multimodal model in the V4-Flash family, adding visual understanding to the V4-Flash architecture. It serves a 1M-token (1,048,576) context window and supports image input with visual grounding, tool calling, structured/JSON output, and configurable reasoning effort (low/high/max, or disabled).

Gemma 3 12B Instruct is the 12-billion-parameter instruction-tuned model in Google DeepMind's Gemma 3 family of open-weight multimodal models. It takes text and image input and produces text output, with a 128K-token context window and support for over 140 languages, for tasks such as chat, summarization, question answering and image understanding.

Gemma 3 4B Instruct is the 4-billion-parameter instruction-tuned model in Google DeepMind's Gemma 3 family of open-weight multimodal models. It takes text and image input and produces text output, with a 128K-token context window and support for over 140 languages, for tasks such as chat, summarization, question answering and image understanding.

Gemma 4 31B Instruct Turbo is a faster serving variant of Gemma 4 31B, the dense instruction-tuned model in Google DeepMind's open-weight Gemma 4 family. The base model takes text and image input and produces text output, with a 256K-token context window, configurable thinking and native function calling.

Granite 4.2 8B is IBM's 8B-parameter dense reasoning model in the open-weight Granite 4.2 family, for reasoning, code generation, tool calling and agentic workflows. It takes text input and produces text output with a 128K-token native context, and offers full thinking, non-thinking and low-effort thinking modes per request.

Ling 3.0 Flash is inclusionAI's hybrid reasoning model, a hybrid-linear Mixture-of-Experts with 124B total and 5.1B active parameters built for efficient agentic workflows such as coding and deep research. It takes text input and output with a 256K-token context window, and its thinking mode can be turned off per request.

Ling 3.0 Flash VL is inclusionAI's multimodal version of Ling 3.0 Flash, a Mixture-of-Experts model with 124B total and 5.5B active parameters. It takes text, image and video input and produces text with a context of up to 256K tokens, and is built for document and chart understanding, GUI agents and long-video question answering.

Ling 3.1 Flash is inclusionAI's native hybrid reasoning model, a hybrid-linear Mixture-of-Experts with 560B total and 25B active parameters that builds on Ling 3.0 Flash. It takes text input and output with a 1M-token context window, is designed for complex real-world tasks, and lets thinking be turned on or off.

Meta Llama 3.1 8B Instruct Turbo is a faster serving variant of Llama 3.1 8B Instruct, Meta's small multilingual instruction-tuned text model. The model has a 128K-token context window, officially supports eight languages and supports tool use.

Llama 3.3 70B Instruct is Meta's 70B-parameter multilingual instruction-tuned model, tuned for assistant-style dialogue in English, German, French, Italian, Portuguese, Hindi, Spanish and Thai. It takes text input and produces text and code with a 128K-token context window. This Turbo variant is a faster serving variant of Llama 3.3 70B Instruct.

Llama 4 Scout is Meta's natively multimodal Mixture-of-Experts model with 17B active parameters, 16 experts and 109B total parameters. It takes text and image input and produces text and code with a context window of up to 10M tokens, and is built for assistant chat and visual reasoning such as captioning and image question answering.

Llama Guard 4 is a natively multimodal safety classifier with 12 billion parameters trained jointly on text and multiple images. Llama Guard 4 is a dense architecture pruned from the Llama 4 Scout pre-trained model and fine-tuned for content safety classification. Similar to previous versions, it can be used to classify content in both LLM inputs (prompt classification) and in LLM responses (response classification). It itself acts as an LLM: it generates text in its output that indicates whether a given prompt or response is safe or unsafe, and if unsafe, it also lists the content categories violated.

Muse Glimmer 30B is an open-weight multimodal model from Meta Superintelligence Lab, distilled from Muse Spark. It is a dense 30B transformer that takes text and image input and produces text with a context window of at least 131K tokens, built for local and coding agents, tool calling and multimodal reasoning on a single consumer GPU or Mac.

Phi-4 is Microsoft's 14B-parameter dense decoder-only language model, trained on a blend of synthetic data, filtered public web content and academic books and Q&A data with a focus on reasoning. It takes text input and produces text output with a 16K-token context window and is intended primarily for English.

12B model trained jointly by Mistral AI and NVIDIA, it significantly outperforms existing models smaller or similar in size.

Mistral Small 3 is a 24B-parameter language model optimized for low-latency performance across common AI tasks. Released under the Apache 2.0 license, it features both pre-trained and instruction-tuned versions designed for efficient local deployment. The model achieves 81% accuracy on the MMLU benchmark and performs competitively with larger models like Llama 3.3 70B and Qwen 32B, while operating at three times the speed on equivalent hardware.

Nemotron Content Safety 3.5 is a multimodal safety classifier developed by NVIDIA. A compact safety model that handles text, images, and custom policies. It outputs a safe/unsafe classification plus a reasoning trace, and can be used as an inference-time guardrail, as a judge for LLM safety testing and evaluation, or with the accompanying training dataset to post-train models for safer behavior.

NVIDIA Nemotron 3.5 Lightning is NVIDIA's fastest open model for always-on agents and high-volume specialized tasks. It delivers a substantial leap in agentic capability over its predecessor Nemotron 3 Nano, with up to 4x higher throughput on a 1M-token context.

NVIDIA Nemotron 3.5 Super VL is an open multimodal model for agents that see, reason, and act. Its 120B-parameter Mixture-of-Experts design (12B active) understands text, images, documents, GUIs, and video, and adds reasoning and tool use on top.

gpt-oss-20b is an open-weight 21B parameter model released by OpenAI under the Apache 2.0 license. It uses a Mixture-of-Experts (MoE) architecture with 3.6B active parameters per forward pass, optimized for lower-latency inference. The model is trained in OpenAI’s Harmony response format and supports reasoning level configuration, fine-tuning, and agentic capabilities including function calling, tool use, and structured outputs.

Qwen3.5-35B-A3B is an efficient Mixture-of-Experts model from Alibaba's Qwen3.5 series with 35B total parameters and only 3B activated per token. It features a 262K token context window (extensible to 1M with YaRN), thinking/reasoning mode, tool calling, and support for 201 languages. Delivers strong performance on reasoning, coding, and vision-language tasks at a fraction of the compute cost.

Qwen3.5-9B is a high-performance model from Alibaba's Qwen3.5 series with a hybrid Gated Delta Networks and sparse MoE architecture. It features a 262K token context window, thinking/reasoning mode, tool calling, multi-token prediction, and support for 201 languages. Excels at reasoning, coding, instruction following, and long-context tasks.

Qwen2.5 is a model pretrained on a large-scale dataset of up to 18 trillion tokens, offering significant improvements in knowledge, coding, mathematics, and instruction following compared to its predecessor Qwen2. The model also features enhanced capabilities in generating long texts, understanding structured data, and generating structured outputs, while supporting multilingual capabilities for over 29 languages.

Qwen3 14B is a dense 14.8B-parameter model in the Qwen3 family from Alibaba's Qwen team, released in April 2025. It switches between a thinking mode for complex reasoning and a fast non-thinking mode, supports 119 languages, and has a native 32K-token context window (131K with YaRN).

Qwen3 30B A3B is a Mixture-of-Experts model in the Qwen3 family with 30.5B total and 3.3B active parameters from Alibaba's Qwen team, released in April 2025. It switches between a thinking mode for complex reasoning and a fast non-thinking mode, supports 119 languages, and has a native 32K-token context window (131K with YaRN).

Qwen3 32B is a dense 32.8B-parameter model in the Qwen3 family from Alibaba's Qwen team, released in April 2025. It switches between a thinking mode for complex reasoning and a fast non-thinking mode, supports 119 languages, and has a native 32K-token context window (131K with YaRN).

Qwen3 VL 30B A3B Instruct is a Mixture-of-Experts vision-language model from Alibaba's Qwen team. It takes text, image and video input and produces text with a native 256K-token context window, and handles visual agent tasks, visual coding, 2D and 3D grounding, long video understanding and OCR in 32 languages.

Qwen3.5-27B is Alibaba's largest dense Qwen3.5 model, delivering near-frontier quality across reasoning, coding, and instruction following. It features a 262K token context window (extensible to 1M), thinking/reasoning mode, tool calling, multi-token prediction, and support for 201 languages. Best suited for production deployments and complex enterprise tasks requiring top-tier performance.

L3 8B Lunaris v1 Turbo is a faster serving variant of Lunaris v1, Sao10K's 8B generalist and roleplay model. Lunaris is a merge of four Llama 3 8B fine-tunes on Meta Llama 3 8B Instruct, aimed at better creativity and logic than Sao10K's Stheno v3.2.

L3.1 70B Euryale v2.2 is Sao10K's fine-tune of Meta Llama 3.1 70B Instruct for creative writing and roleplay. Compared with v2.1 it was trained on more roleplay and creative writing data, with added data for system-prompt adherence and reasoning.

Hy4 Preview is a 770B-parameter Mixture-of-Experts model from the Tencent Hy Team, with 49B parameters active per token and a 1M-token context window. It uses Gated DeepSeek Sparse Attention with IndexCache and is post-trained for software engineering, office and data analysis, game development and scientific research, with adjustable reasoning effort.

Inkling Small is a smaller model in Thinking Machines Lab's Inkling family, a Mixture-of-Experts with 276B total and 12B active parameters. Like Inkling, it accepts text, image and audio input and generates text, with a context of up to 1M tokens and variable thinking effort.

MiMo-V2.6-Pro is the flagship of Xiaomi's MiMo-V2.6 series: a sparse Mixture-of-Experts model with 1.02T total and 42B activated parameters, natively omnimodal across text, image, video and audio, with a 1M-token context window. Trained with large-scale mixed reinforcement learning and groupwise agentic grading, it is built for demanding agentic work including long-horizon coding, multi-tool agents and cybersecurity tasks.

GLM-5.3-Flash is a native multimodal model from Z.ai. It is suited for efficient coding and long-horizon agent tasks. Its hybrid sparse and linear attention architecture maintains accurate long-context behavior while reducing compute overhead.

Claude Opus 4.7 is Anthropic's Opus model for advanced software engineering, especially complex, long-running coding tasks. It builds on Opus 4.6 with more precise instruction following, verification of its own outputs and higher-resolution vision, and offers a 1M-token context window, up to 128K output tokens and adaptive thinking.

Claude Opus 4.8 is Anthropic's Opus model for dependable, long-running agentic work and coding. It builds on Opus 4.7 with sharper judgment and more candid flagging of doubts about its own output, and takes text and image input with a 1M-token context window, up to 128K output tokens and adaptive thinking.

Claude Opus 5 is Anthropic's Opus model for daily professional work, with gains over Opus 4.8 in software engineering, knowledge work and scientific research. It checks its own work and iterates until a task succeeds, and takes text and image input with a 1M-token context window, up to 128K output tokens and adaptive thinking.

Claude Opus 5.5 is Anthropic's Opus model for long-running agentic coding and knowledge work. Compared with Opus 5 it uses fewer tokens per task and writes more clearly, and it takes text and image input with a 1M-token context window, up to 128K output tokens and always-on adaptive thinking.

Claude Sonnet 4.6 is Anthropic's best combination of speed and intelligence, offering strong performance on coding, reasoning, and general tasks with excellent speed and cost efficiency. It features a 1M token context window and 64K max output tokens.

Claude Sonnet 5 is Anthropic's Sonnet model built for agentic work: planning, using tools such as browsers and terminals, coding, reasoning and knowledge work. It takes text and image input and produces text, with a 1M-token context window, up to 128K output tokens and adaptive thinking, and improves on Sonnet 4.6.

Claude Sonnet 5.5 is Anthropic's Sonnet model, balancing speed and intelligence for well-scoped everyday work such as building features, fixing bugs and creating documents, slides and spreadsheets. It improves on Sonnet 5 in coding, token efficiency and clear writing, with a 1M-token context window, up to 128K output tokens and adaptive thinking.

Gemini 3.1 Pro Preview is the preview release of Google's Gemini 3.1 Pro, a reasoning model for software engineering and agentic workflows that need precise tool use. It refines the Gemini 3 Pro series with better thinking, improved token efficiency and more grounded, factually consistent answers. It takes text, image, video, audio and PDF input and produces text output, with a 1M-token context window and up to 65K output tokens.

Gemini 3.5 Flash is a Flash model in Google's Gemini 3 series, optimized for real-world tasks at higher speed and lower cost. It supports thinking. It takes text, image, video, audio and PDF input and produces text output, with a 1M-token context window and up to 65K output tokens.

Gemini 3.7 Flash is a Flash model in Google's Gemini 3 series of natively multimodal reasoning models, built as a workhorse for coding and agentic work. It improves on Gemini 3.6 Flash in coding and web development, supports low, medium and high thinking levels, and has a 1M-token context window with up to 65K output tokens.

Qwen 3.7 Max is the largest model in the Qwen 3.7 series, with deep thinking, function calling, prompt caching, and multimodal input support for images and video. It excels at programming, office and productivity tasks, and long-running autonomous agent workflows.

Qwen 3.8 Flash is an efficient multimodal model from Alibaba's Qwen team, built on the open-weight Qwen3.8 Flash Next Mixture-of-Experts model with 125B total and 6B active parameters. It accepts text, image and video input with a 1M-token context window, supports function calling and structured output, and has a thinking mode that can be turned off.

Qwen 3.8 Max is Alibaba's flagship 2.4-trillion-parameter MoE model, with major gains over Qwen 3.7 Max in software engineering and office-productivity workflows and strong long-horizon, multi-agent performance. It accepts both text and vision-language input (images and video), operates in thinking mode only, and supports a 1M-token context window.

Claude Haiku 4.5 is Anthropic's small, fast model for real-time, low-latency work such as chat assistants, customer support agents and pair programming. It takes text and image input and produces text output, with a 200K-token context window, up to 64K output tokens and extended thinking.

Optimized specifically for multimodal agent scenarios. It features enhanced agent capabilities, upgraded multimodal comprehension, and more flexible context management.

Seed 2.0 Code is the dedicated programming model in ByteDance Seed's Seed 2.0 family, built for software development and coding agents. It supports deep thinking and tool calling with a 256K-token context window and up to 128K output tokens.

Seed 2.0 Mini is the throughput-optimized model in ByteDance Seed's Seed 2.0 family, built for low-latency, high-concurrency and batch workloads. It supports multimodal understanding, tool calling and four thinking levels from minimal to high, with a 256K-token context window.

Seed 2.0 Pro is the general agent model in ByteDance Seed's Seed 2.0 family that focuses on long-chain reasoning and robustness in complex, multi-step workflows. It supports adjustable deep thinking, multimodal understanding and tool calling, with a 256K-token context window and up to 128K output tokens.

Gemini 2.5 Flash is Google's thinking model built for price-performance, suited to large-scale processing, low-latency, high-volume tasks and agentic use cases. It takes text, image, video and audio input and produces text output, with a 1M-token context window and up to 65K output tokens.

Gemini 3.1 Flash-Lite is Google's low-latency, cost-effective multimodal model for high-frequency, lightweight tasks such as high-volume agentic workflows and simple data extraction. It takes text, image, video, audio and PDF input and produces text output, with a 1M-token context window, up to 65K output tokens and thinking support.

Gemma 4 31B Instruct Ultra is a speed-optimized serving variant of Gemma 4 31B, the dense instruction-tuned model in Google DeepMind's open-weight Gemma 4 family. The base model takes text and image input and produces text output, with a 256K-token context window, configurable thinking and native function calling.

GPT OSS 120B Ultra is a speed-optimized serving variant of gpt-oss-120b, OpenAI's open-weight Mixture-of-Experts language model with 117B total and 5.1B active parameters. The base model takes text input and produces text output with a 128K-token context window, configurable reasoning effort and native tool use.

DeepSeek V4.1 Flash is a multimodal Mixture-of-Experts model with 552B backbone parameters (8B active on input, 16B on output) and a 1M-token context window. It natively processes images and text, with strong reasoning, coding, and agentic performance.

Kimi K3 is Moonshot AI's open-weight Mixture-of-Experts model with 2.8T total and 104B active parameters and a 1M-token context window. It accepts text, image and video input and produces text, always reasons with low, high or max effort, and is built for long-horizon coding, knowledge work and reasoning.

GLM-5.3 is a large-scale reasoning model from Z.ai, built for complex software engineering and long-horizon agent tasks. It supports text input and output with a 1M-token context window, and improves on GLM-5.2 in coding and in the balance between performance and token efficiency.

GLM-5.3-Flash is a native multimodal model from Z.ai. It is suited for efficient coding and long-horizon agent tasks. Its hybrid sparse and linear attention architecture maintains accurate long-context behavior while reducing compute overhead.

MiniMax-M3 is MiniMax's natively multimodal Mixture-of-Experts model with about 428B total and 23B active parameters. It takes text, image and video input and produces text, with a context window of up to 1M tokens, and can also operate a desktop computer.

Qwen 3 Coder 480B Turbo is a faster serving variant of Qwen3 Coder 480B A35B Instruct, Alibaba's Qwen team's Mixture-of-Experts coding model with 480B total and 35B active parameters. The model has a native 262K-token context window and is built for agentic coding, tool calling and repository-scale work.