Open-weight models, on infrastructure you trust.

Qwen 3.8 Flash Next

Alibaba qwen3.8-flash-next

Qwen 3.8 Flash Next is Alibaba's next-generation flagship model, a hybrid mixture-of-experts design that activates only a fraction of its parameters per token for efficient long-context inference. It reads images as well as text and reasons step by step by default, with its thinking mode switchable off for direct answers. It supports native tool use across a 262K-token context, targeting demanding agentic and multimodal work.

Chat · Vision · Tools · Thinking

DeepSeek V4 Flash

DeepSeek deepseek-v4-flash

Frontier reasoning, open weights

Chat · Tools · Thinking

Qwen 3.8

Alibaba qwen3.8

Qwen 3.8 is Alibaba's latest open-weight multimodal model, built for agentic work with strong multilingual reasoning, native tool use, and image understanding. It offers a hybrid thinking mode that can reason at length or answer directly, and handles 262K-token contexts, making it a capable all-rounder for chat, coding, and document tasks.

Chat · Vision · Tools · Thinking

Qwen 3.6

Alibaba qwen3.6

Qwen 3.6 is Alibaba's flagship open-weight model, built for agentic work with strong multilingual reasoning, native tool use, and image understanding. It offers a hybrid thinking mode that can reason at length or answer directly, and handles 262K-token contexts, making it a capable all-rounder for chat, coding, and document tasks.

Chat · Vision · Tools · Thinking

Qwen 3.5

Alibaba qwen3.5

Alibaba’s flagship multimodal family

Chat · Vision · Tools · Thinking

gpt-oss

OpenAI gpt-oss

OpenAI’s open-weight release

Chat · Tools · Thinking

Gemma 4

Google DeepMind gemma4

Google DeepMind’s efficient multimodal family

Chat · Vision · Tools · Thinking

Qwen 3 Coder

Alibaba qwen3-coder

Purpose-built for code

Chat · Tools

Llama 3.3

Meta llama3.3

Meta’s 70B workhorse

Chat · Tools

Qwen 3 VL

Alibaba qwen3-vl

Qwen 3 VL is Alibaba's vision-language model, pairing Qwen 3's reasoning and tool use with detailed image understanding. It reads charts, documents, and photos alongside text and offers an explicit thinking mode, making it well suited to multimodal analysis and agentic tasks.

Chat · Vision · Tools · Thinking

Llama 4 Scout

Meta llama4-scout

Llama 4 Scout is Meta's natively-multimodal mixture-of-experts model, activating 17B parameters per token from a larger expert pool for efficient inference. It reads images as well as text and supports an exceptionally long, million-token context, aimed at long-document and multimodal workloads.

Chat · Vision · Tools

Phi-4

Microsoft phi-4

Phi-4 is Microsoft's small-but-capable model, trained with a heavy emphasis on high-quality and synthetic data to punch above its size on reasoning tasks. Its compact footprint makes it a practical choice for cost-sensitive and on-device deployments.

Chat · Thinking

Qwen 3 Coder Next

Alibaba qwen3-coder-next

Qwen 3 Coder Next is Alibaba's next-generation coding model, built on a hybrid architecture for efficient long-context inference. It targets agentic, repository-scale software tasks and tool use across a 262K-token context.

Chat · Tools

GLM 4.7 Flash

Z.ai glm-4.7-flash

GLM 4.7 Flash is Z.ai's fast, lightweight model in the GLM 4 line, offering a hybrid reasoning mode that trades depth for speed on demand. It's tuned for tool use and responsive agentic workflows across a roughly 200K-token context.

Chat · Tools · Thinking

Ministral 3

Mistral AI ministral-3

Ministral 3 is Mistral AI's compact instruction model family, offered in several sizes with tool use and image understanding. It's engineered for efficient, low-latency inference on modest hardware while retaining strong general capability, making it a good edge and on-device option.

Chat · Vision · Tools · Thinking

Devstral Small 2

Mistral AI devstral-small-2

Devstral Small 2 is Mistral AI's coding-focused model, tuned for software-engineering agents that navigate and edit real codebases. It combines tool use with image understanding in a compact 24B size, targeting practical developer automation.

Chat · Vision · Tools

GLM OCR

Z.ai glm-ocr

GLM OCR is Z.ai's document-understanding model, specialised for optical character recognition and structured extraction from images and scanned pages. It reads dense layouts and returns clean, machine-usable text across a 131K-token context.

Vision · OCR

Mistral Small 4

Mistral AI mistral-small-4

Mistral Small 4 is Mistral AI's mid-size flagship, a mixture-of-experts model that activates only a small fraction of its parameters per token for efficient inference. It reads images as well as text, supports tool use, and handles a 256K-token context, making it a capable all-rounder for chat, coding, and document work.

Chat · Vision · Tools · Thinking

LFM 2.5

Liquid AI lfm2-5

LFM 2.5 is Liquid AI's efficient model family built on a liquid-foundation architecture designed for fast on-device inference. It includes a mixture-of-experts variant and supports tool use, targeting responsive assistants that run well on constrained hardware.

Chat · Tools · Thinking

Devstral 2

Mistral AI devstral-2

Devstral 2 is Mistral AI's large coding model, built for software-engineering agents that explore and modify real repositories rather than answer isolated questions. At 123B dense parameters with a 256K-token context and native tool use, it targets long-running agentic development work. It ships under a modified MIT licence that permits commercial use.

Chat · Tools

LFM 2.5 VL

Liquid AI lfm2-5-vl

LFM 2.5 VL is Liquid AI's compact vision-language model, adding image understanding to the efficient LFM 2.5 line. Its small footprint suits multimodal assistants that need to run on-device or at low latency.

Chat · Vision

Nemotron 3 Nano

NVIDIA nemotron-3-nano

Nemotron 3 Nano is NVIDIA's small, efficient model tuned for reasoning and tool use, offered in dense and mixture-of-experts sizes. It's built for cost-effective agentic deployments and supports a 131K-token context.

Chat · Tools · Thinking

Granite 4.0 Micro

IBM granite-4-micro

Granite 4.0 Micro is IBM's smallest enterprise model, a hybrid design tuned for tool use and reliable instruction following. Its compact size and permissive licensing make it a practical fit for on-device and cost-sensitive business applications.

Chat · Tools

Magistral Small

Mistral AI magistral

Magistral Small is Mistral AI's reasoning model, trained to work through problems step by step before answering. It now reads images as well as text, extending that deliberate reasoning to visual input. At a compact 24B size it brings transparent reasoning and tool use within reach of single-GPU deployments.

Chat · Vision · Tools · Thinking

GLM 4.6V Flash

Z.ai glm-4.6v-flash

GLM 4.6V Flash is Z.ai's fast vision-language model, pairing image understanding with responsive, lightweight inference. It's suited to multimodal chat and analysis where latency matters, across a 131K-token context.

Chat · Vision · Tools · Thinking

Mistral Large 3

Mistral AI mistral-large-3

Mistral Large 3 is Mistral AI's frontier open-weight model, a mixture-of-experts design that activates a fraction of its parameters per token so inference cost tracks the active size rather than the full one. It reads images as well as text, uses tools natively, and handles a 256K-token context for demanding reasoning and long-document work.

Chat · Vision · Tools

MiMo V2.5

Xiaomi mimo-v2.5

MiMo V2.5 is Xiaomi's open model tuned for reasoning, tool use, and image understanding. It offers an explicit thinking mode and targets capable multimodal assistance in an efficient package.

Chat · Vision · Tools · Thinking

Ornith 1.5

DeepReinforce ornith-1.5

Ornith 1.5 is DeepReinforce's second-generation reasoning model, trained through a self-improvement loop in which the model proposes its own tasks, builds the scaffolding to solve them, and learns from the resulting rollouts. It reasons step by step before answering, calls tools natively, and reads images as well as text. A 262K-token context and a choice of dense 9B or sparse 35B and 122B sizes suit it to long agentic coding and research work.

Chat · Vision · Tools · Thinking

Voxtral Small

Mistral AI voxtral-small

Voxtral Small is Mistral AI's speech-understanding model: it takes spoken audio directly as input alongside text, so you can ask questions about a recording rather than transcribing it first. Built on the Mistral Small 24B backbone, it handles a 131K-token context and targets voice assistants and audio analysis.

Chat

Ornith 1.0

DeepReinforce ornith

Ornith 1.0 is DeepReinforce's hybrid-reasoning model, designed to combine deliberate step-by-step thinking with efficient long-context inference. It supports tool use and is offered in two sizes for different quality-versus-cost trade-offs.

Chat · Tools · Thinking

DeepSeek OCR

DeepSeek deepseek-ocr

DeepSeek OCR is DeepSeek's document-understanding model, specialised for optical character recognition and extracting text from images and scans. It focuses on accurate, structured reading of visual documents.

Vision · OCR

DeepSeek R1 Distill

DeepSeek deepseek-r1-distill

DeepSeek R1 Distill is a family of smaller models distilled from DeepSeek R1's reasoning traces onto Qwen and Llama backbones. They bring much of R1's step-by-step reasoning to sizes that run on modest hardware.

Chat · Thinking

dots.ocr

Rednote dots-ocr

dots.ocr is Rednote's document-parsing model, built to read complex layouts — tables, columns, and mixed text — and return structured output. It targets accurate OCR and document understanding across a 131K-token context.

Vision · OCR

BGE Reranker v2 M3

BAAI bge-reranker-v2-m3

BGE Reranker v2 m3 is BAAI's multilingual reranking model, scoring how well each candidate document answers a query to sharpen retrieval results. It's a lightweight, effective second stage for RAG pipelines across many languages.

Whisper

OpenAI whisper

Whisper is OpenAI's speech-to-text model, transcribing audio across many languages with strong robustness to accents and background noise. The large-v3 and faster large-v3-turbo variants suit everything from meeting notes to media captioning.

Transcription

FLUX.1 [schnell]

Black Forest Labs flux.1-schnell

FLUX.1 [schnell] is Black Forest Labs' fast text-to-image model, distilled to generate high-quality images in just a few diffusion steps. Its speed and permissive Apache-2.0 licence make it a strong default for quick drafts and interactive generation.

Image

Z-Image Turbo

Alibaba z-image-turbo

Z-Image Turbo is Alibaba's distilled text-to-image model built for speed, producing images in only a few diffusion steps. It trades some peak fidelity for near-interactive latency, making it a good default for fast drafts and previews.

Image

SDXL Turbo

Stability AI sdxl-turbo

SDXL Turbo is Stability AI's real-time text-to-image model, distilled from SDXL to generate images in as little as one to four steps. It's ideal for interactive and preview use where speed matters most.

Image

FLUX.2 [klein]

Black Forest Labs flux.2-klein

FLUX.2 [klein] is Black Forest Labs' efficient text-to-image model, offered in 4B and 9B sizes for a choice of speed versus fidelity. It brings the FLUX.2 line's image quality to hardware that can't run the largest checkpoints. Note that the two sizes ship under different terms: the 4B is Apache-2.0, while the 9B is released under the FLUX Non-Commercial License and is not licensed for commercial use.

Image

Qwen-Image 2512

Alibaba qwen-image-2512

Qwen-Image 2512 is Alibaba's high-fidelity text-to-image model, notable for rendering legible text within images. It targets detailed, prompt-faithful generation for design and content work.

Image

Stable Diffusion 3.5 Large Turbo

Stability AI stable-diffusion-3.5-large-turbo

Stable Diffusion 3.5 Large Turbo is Stability AI's distilled flagship text-to-image model, generating high-quality images in a handful of steps. It pairs the detail of SD 3.5 Large with fast, few-step inference.

Image

North Mini Code 1.0

Cohere north-mini-code-1.0

North-Mini-Code-1.0 is an open-weights sparse Mixture-of-Experts (MoE) developer model by Cohere optimized for agentic software engineering, featuring 30 billion total parameters with 3 billion active parameters per token, and supporting a 256K context window.

Chat · Tools · Thinking

Krea 2 Turbo

Krea krea-2-turbo

Krea 2 Turbo is Krea's distilled text-to-image model, a ~13B diffusion transformer tuned to render high-quality images in only a few steps. It pairs a Qwen3-VL text encoder with a Wan 2.1 VAE and targets fast, near-interactive generation rather than long multi-step sampling. Turbo trades a little peak fidelity for speed, making it a strong default for quick drafts and iteration.

Image

Qwen 3 TTS

Alibaba qwen3-tts

Qwen 3 TTS is Alibaba's open-weight text-to-speech model, generating natural, expressive speech across ten languages including English, Chinese, Japanese and Korean. This CustomVoice edition ships nine built-in speakers, male and female, and each can voice any of the supported languages. Built on a compact 1.7B talker, it targets fast, low-footprint speech synthesis that runs comfortably on modest GPUs.

Speech

Muse Glimmer

Meta muse-glimmer

Muse Glimmer is Meta Superintelligence Labs' compact multimodal model — a dense vision-language transformer distilled for agentic work on consumer hardware. It pairs a text decoder with a dedicated perception encoder to read images alongside text, and targets fast local inference across a 128K-token context.

Chat · Vision · Tools · Thinking

Nemotron 3.5 Lightning

NVIDIA nemotron-3.5-lightning

Nemotron 3.5 Lightning is NVIDIA's efficient mixture-of-experts reasoning model, activating roughly 3B of its 30B parameters per token through a hybrid Mamba–attention architecture for fast, low-cost inference. Tuned for agentic use, it supports tool calling and step-by-step reasoning across a 1M-token context.

Chat · Tools · Thinking

Llama Guard 3

Meta llama-guard3

Llama Guard 3 is Meta's safety classifier, fine-tuned to judge whether a prompt or a reply breaks a fixed hazard taxonomy covering violent and non-violent crime, sexual content, child safety, hate, self-harm, privacy and elections. The 1B size is built for fast text checks and names the categories it finds, making it a light first line of content filtering.

Moderation

SingGuard

Ant Group singguard

SingGuard is Ant Group's open safety model for screening both text and images, built on Qwen vision-language models. It judges content against a risk policy, so one small model can check prompts, replies and generated images. The 0.8B size suits small GPUs and fast checks, while the 4B size catches more.

Vision · Moderation

Qwen3Guard

Alibaba qwen3guard

Qwen3Guard is Alibaba's open safety classifier for prompts and replies, trained across 119 languages. It rates content safe, unsafe or controversial and names the category, from violence, illegal acts and sexual content to self-harm, personal data, politically sensitive topics and jailbreak attempts. The 0.6B size is the fastest and the 8B size the most accurate.

Moderation

nomic-embed-text-v1.5

Nomic AI nomic-embed-text

nomic-embed-text v1.5 is Nomic AI's open text-embedding model, producing high-quality vectors for search, retrieval, and RAG. It supports adjustable embedding dimensions via Matryoshka representation, letting you trade vector size for cost.

Embeddings

Qwen 3 Embedding

Alibaba qwen3-embedding

Qwen 3 Embedding is Alibaba's text-embedding family built on Qwen 3, offered in several sizes for search and retrieval. It delivers strong multilingual embeddings for RAG and semantic-search pipelines across a 32K-token context.

Embeddings

BGE M3

BAAI bge-m3

BGE-M3 is BAAI's versatile embedding model supporting dense, sparse, and multi-vector retrieval in a single model. It handles more than 100 languages and long inputs, making it a strong multilingual default for search and RAG.

Embeddings

EmbeddingGemma

Google DeepMind embeddinggemma

EmbeddingGemma is Google DeepMind's compact text-embedding model, built from Gemma 3 and sized to run comfortably on a laptop or a small GPU. It was trained on over 100 languages and produces 768-dimension vectors that can be truncated to 512, 256, or 128 dimensions via Matryoshka representation learning, letting you trade a little accuracy for a smaller, faster index. A practical default for local RAG, semantic search, and clustering.

Embeddings