NVIDIA

DeepSeek V4.1 Flash is the open-weight, cost-efficient sparse mixture-of-experts tier of the V4.1 family. It has 552B total parameters with 8B active on input and 16B on output, 1M context, and native vision in with text out. This is the original September 2026 checkpoint.

Gemma 3 4B IT is Google's instruction-tuned 4B open model with text and image input. It has a 128K context window and multilingual support in over 140 languages. It supports tool calling.

Muse Glimmer 30B is Meta's dense multimodal instruct model with about 30B parameters, distilled from Muse Spark. It takes text and image input and returns text, with reasoning and function calling. Weights are open under Apache 2.0.

MiniMax M3.1 Flash is MiniMax's open-weight multimodal model for complex agentic tasks, coding and cybersecurity. It accepts text, image and video input and has a 1M-token context window. It always thinks, with adjustable effort, and its weights are not yet released.

Mistral Large 4 is Mistral AI's open-weight Mixture-of-Experts model with about 1 trillion parameters and a 1M-token context. It takes text and image input and supports tool use, structured output and Low, Medium or High reasoning effort.

Kimi K2.6 is Moonshot AI's open-weight large mixture-of-experts model with agentic tool use, vision input and long-context memory. It supports reasoning and streaming, and takes text and image input.

Nemotron 3 Embed 1B is NVIDIA's 1.14B-parameter multilingual text embedding model covering 34 languages. It returns 2048-dimension vectors for semantic search, retrieval and RAG. It is built on Ministral-3-3B-Instruct and has a 32,768-token context.

Nemotron-3 Nano Omni 30B-A3B (Reasoning) is NVIDIA's multimodal mixture-of-experts model for video, audio, image and text understanding. It targets enterprise Q&A, summarization, transcription and OCR, and supports reasoning and function calling.

Nemotron 3.5 Bolt E4B is NVIDIA's open-weight 8B-parameter small language model, post-trained from Gemma 4 E4B. It improves multi-turn tool use, instruction following and input grounding for voice and device assistants. Context window is 65,536 tokens.

Nemotron 3.5 Content Safety is NVIDIA's open-weight 4B-parameter multimodal guardrail model, fine-tuned from Google Gemma-3-4B. It moderates inputs to and responses from LLMs and VLMs. It returns a safety classification, category labels and optional reasoning.

Nemotron 3.5 Embed 1B is NVIDIA's 1.55B-parameter multimodal and multilingual retrieval embedding model. It embeds text in 34 evaluated languages and document-page images into one 2048-dimension vector for semantic search and RAG. Context is 32,768 tokens.

Nemotron 3.5 Embed 8B is NVIDIA's 8.38B-parameter multimodal and multilingual retrieval embedding model. It embeds text in 34 evaluated languages and document-page images into one 4096-dimension vector for semantic search and RAG. Context is 32,768 tokens.

Nemotron 3.5 Lightning 30B A3B is NVIDIA's sparse mixture-of-experts chat model, optimized for low-latency agentic workloads. It supports function calling and coding. It allows up to 131,072 output tokens.

Nemotron 3.5 Super VL 120B A12B is NVIDIA's vision-language model for reasoning, coding and tool use. It is a hybrid Mamba-2 and Transformer mixture-of-experts with 120B total and 12B active parameters. It takes text, image and video input.

Nemotron Parse 2.0 is NVIDIA's document-understanding vision-language model. A page image goes in and structured text comes out as markdown, layout classes, bounding boxes and reading order. It does not generate images, and its context is 8,192 tokens.

Riva Translate 4B Instruct v2 is NVIDIA's compact 4B-parameter instruction-tuned model for neural machine translation. It covers 30+ languages. Context window is 4,096 tokens.

GPT-OSS 20B is OpenAI's open-weight reasoning model for agentic tasks and low-latency edge use, under Apache-2.0. It is a mixture-of-experts model with 20.9B total and 3.6B active parameters and a 131,072-token context. It runs within 16 GB of memory with MXFP4 weights.

GLM-5.3 is Z.AI's open-weight reasoning model for coding and agentic workflows. It targets complex software engineering, long-running agents, and vulnerability analysis. It uses the same base model as GLM-5.2 with scaled post-training for stronger coding and token efficiency.