Skip to content

AI models directory

Every major open-weight and API model family in one place: versions, download sizes, the VRAM you need, context windows, licences and retirement dates. Model data syncs daily from the Ollama library, Hugging Face and the vendors' docs.

  • 18Model families
  • 56Current models
  • 70Run on your own hardware
  • 7API retirements in 90 days
What can my GPU run? Compare models VRAM calculator

What can I run on my GPU?

Our picks for each amount of GPU memory, using Ollama's default 4-bit builds and an 8K context. On a CPU-only machine or a Mac, read the numbers as system RAM.

CPU only / 4–6 GB · Laptops, mini PCs, small VPS

Qwen3.5 2B

Best all-round tiny model, with image input

Top pick

~3.5 GB VRAM2.6 GB download256K context

Llama 3.2 3B

Fast and well supported everywhere

~3 GB VRAM2 GB download128K context

Phi-4-mini 3.8B

Strong maths and reasoning for its size

~3.5 GB VRAM2.5 GB download128K context

Granite 4.2 3B

Apache 2.0, good at RAG and JSON output

~3 GB VRAM2.2 GB download128K context

Latest model releases

Newest models across all families

API models retiring soon

Announced shutdowns in the next 12 months

All models

56 models

AbilitiesLicence / access
Gemini 4 HydrogenEveryday coding and agents; up to 1M output tokensGeminiNov 2026API1MToolsReasoningVisionAudioProprietary API
Gemini 4 CarbonStrongest reasoning: software engineering, research, securityGeminiNov 2026API1MToolsReasoningVisionAudioProprietary API
Claude Fable 5.2Most demanding reasoning and long-horizon agent workClaudeNov 2026API1MToolsReasoningVisionProprietary API
Claude Mythos 5.2ClaudeNov 2026API1MToolsReasoningVisionProprietary API
Kimi K3.1 ProMost capable Kimi: coding, research and agentsKimiNov 2026Cloud-size—ToolsReasoningVisionKimi K3.1 Pro License
GPT-6.1 TerraFast: visual reasoning, computer use, writingGPTOct 2026API1.05MToolsReasoningVisionProprietary API
Gemini 4 ArgonEfficient frontier reasoning, faster and cheaper than CarbonGeminiOct 2026API1MToolsReasoningVisionAudioProprietary API
Claude Haiku 5.5Fastest: classification, extraction, routingClaudeOct 2026API1MToolsReasoningVisionProprietary API
Nemotron 3.5 SuperVision-language model for coding, tools and documentsNemotronOct 2026~100 GB1MToolsReasoningVisionOpenMDW 1.1
Nemotron 3.5 BoltCompact local model built from Gemma 4 E4BNemotronOct 2026~9.5 GB128KToolsReasoningVisionAudioOpenMDW 1.1
GPT-6.1 SolNear-Astra quality at a lower priceGPTSep 2026API1.05MToolsReasoningVisionProprietary API
Claude Sonnet 5.5Best balance of speed and intelligenceClaudeSep 2026API1MToolsReasoningVisionProprietary API
Claude Opus 5.5Recommended default: agentic coding and knowledge workClaudeSep 2026API1MToolsReasoningVisionProprietary API
GPT-6 LunaCheapest, for focused high-volume tasksGPTSep 2026API1.05MToolsReasoningVisionProprietary API
GPT-6 AstraMost capable: hardest reasoning and codingGPTSep 2026API1.05MToolsReasoningVisionProprietary API
DeepSeek-V4.1 FlashMultimodal MoE with 1M context and a compressed KV cacheDeepSeekSep 2026Cloud-size—ToolsReasoningVisionMIT
EmbeddingGemma 2Multimodal: text, images, audio and video in one 768-dim spaceEmbedding modelsSep 2026~2 GB256KVisionEmbeddingsApache 2.0
Muse Spark OpenFrontier-size multimodal model for coding and long agent runsMeta MuseSep 2026Cloud-size—ToolsReasoningVisionMuse Model License
Mistral Large 4Largest Mistral model, multimodal, 1M contextMistralSep 2026Cloud-size—ToolsReasoningVisionMistral Ecosystem Open-Weight License
GLM-5.3Flagship open coding model with 1M contextGLMAug 2026Cloud-size—ToolsReasoningGLM-5.3 License
GLM-5.3-FlashFirst multimodal GLM-5 model; strong coding at low costGLMAug 2026Cloud-size—ToolsReasoningVisionMIT
Granite 4.2Enterprise RAG, tools and JSON with a thinking modeGraniteAug 2026~3 GB128KToolsReasoningApache 2.0
Muse GlimmerLocal agent model with tools, reasoning and image inputMeta MuseAug 2026~20.5 GB128KToolsReasoningVisionApache 2.0
Nemotron 3.5 LightningFast always-on agent model with 1M contextNemotronAug 2026~28.5 GB1MToolsReasoningOpenMDW 1.1
Qwen3.8Flagship local model for coding, agents and visionQwenAug 2026~20.5 GB256KToolsReasoningVisionApache 2.0
Qwen3.8 Flash-NextEarly preview of the Qwen4 architectureQwenAug 2026~132.5 GB256KToolsReasoningVisionQwen Community License
Gemini 3.5 Flash-LiteFastest, cheapest model for high throughputGeminiJul 2026API1MToolsReasoningVisionAudioProprietary API
Kimi K3First open 3T-class model, native visionKimiJun 2026Cloud-size—ToolsReasoningVisionKimi K3 License
Kimi K2.7 CodeCoding agent model, ~30% fewer thinking tokens than K2.6KimiJun 2026Cloud-size—ToolsReasoningVisionModified MIT
Nemotron 3 UltraLargest Nemotron, for long agent workflowsNemotronJun 2026Cloud-size1MToolsReasoningOpenMDW 1.1
DeepSeek-V4 ProFlagship with three reasoning levels and 1M contextDeepSeekApr 2026Cloud-size—ToolsReasoningMIT
Granite Guardian 4.1Checks prompts and answers for risk; LLM judgeGraniteApr 2026~6.5 GB128KToolsReasoningApache 2.0
Qwen3.6Agentic coding; the 35B MoE is very fastQwenApr 2026~20.5 GB256KToolsReasoningVisionApache 2.0
Gemma 4Multimodal (image + audio), reasoning, 256K contextGemmaMar 2026~6 GB256KToolsReasoningVisionAudioApache 2.0
Mistral Medium 3.5One model for chat, reasoning and codingMistralMar 2026~88.5 GB256KToolsReasoningVisionModified MIT
Nemotron Cascade 2Reasoning and agentic MoENemotronMar 2026~27 GB256KToolsReasoningNVIDIA Open Model License
Gemini 3.1 Pro (preview)GeminiFeb 2026API1MToolsReasoningVisionAudioProprietary API
Qwen3.5Multimodal family from 0.8B to 122B, best for small GPUsQwenFeb 2026~2 GB256KToolsReasoningVisionApache 2.0
TranslateGemmaTranslation across 55 languagesGemmaJan 2026~4.5 GB128KVisionGemma Terms of Use
GLM-4.7-FlashStrong 30B-class model for one GPUGLMJan 2026~21.5 GB198KToolsReasoningMIT
GLM-OCRDocument OCR and layout understandingGLMJan 2026~2.5 GB128KToolsVisionMIT
Qwen3-Coder-NextAgentic coding with 256K context on one large GPUQwen CoderJan 2026~58 GB256KToolsApache 2.0
Devstral 2Large agentic coding model, 256K contextDevstralDec 2025~83 GB256KToolsModified MIT
Devstral Small 2Coding agent on one GPU, image input, 384K contextDevstralDec 2025~17 GB384KToolsVisionApache 2.0
Ministral 3Edge models with vision and 256K contextMistralDec 2025~4 GB256KToolsVisionApache 2.0
gpt-oss-safeguardSafety classifier that follows your own written policygpt-ossOct 2025~16 GB128KToolsReasoningApache 2.0
gpt-ossReasoning and tool use; 20b fits 16 GB, 120b fits one 80 GB GPUgpt-ossAug 2025~16 GB128KToolsReasoningApache 2.0
Qwen3-EmbeddingTop multilingual quality; 8B for best resultsEmbedding modelsJun 2025~1.5 GB40KEmbeddingsApache 2.0
Mistral Small 3.2Reliable 24B model with vision and tool callingMistralJun 2025~17 GB128KToolsVisionApache 2.0
MedGemmaMedical text and image researchGemmaMay 2025~4.5 GB128KVisionHealth AI Developer Foundations
Llama 4Multimodal MoE; Scout has a 10M-token contextLlamaApr 2025~74.5 GB10MToolsVisionLlama 4 Community License
Phi-4-reasoningStep-by-step reasoning (plus a "plus" variant)PhiApr 2025~13 GB32KText onlyMIT
Phi-4-mini-reasoningMaths reasoning on CPUs and small GPUsPhiApr 2025~4.5 GB128KText onlyMIT
nomic-embed-text v2 MoEMultilingual MoE embedding modelEmbedding modelsFeb 2025~2 GB512EmbeddingsApache 2.0
Phi-4-miniSmall model with tool calling and 128K contextPhiFeb 2025~3.5 GB128KToolsMIT
Phi-414B model strong at maths and reasoningPhiDec 2024~11 GB16KText onlyMIT
Snowflake Arctic Embed 2Multilingual retrieval with strong English scoresEmbedding modelsNov 2024~2 GB8KEmbeddingsApache 2.0
bge-m3Multilingual, 8K input, dense + sparse retrievalEmbedding modelsJan 2024~2 GB8KEmbeddingsMIT

Open-weight models

DeepSeekDeepSeek DeepSeek's open MoE models: R1 made open reasoning models mainstream, and DeepSeek-V4 brings 1M-token context under the MIT licence. Newest: DeepSeek-V4.1 Flash From ~2 GB 160K context DevstralMistral AI Mistral's coding models: Devstral agents that explore codebases and edit many files, and Codestral for autocomplete. Newest: Devstral 2 From ~16 GB 384K context GemmaGoogle DeepMind Google DeepMind's open models, built from Gemini research. Gemma 4 runs from phones to one GPU, takes images and audio, and is… Newest: Gemma 4 From ~1 GB 256K context GLMZ.ai (Zhipu AI) Z.ai's GLM models: GLM-5.3 is a leading open coding model, GLM-5.3-Flash is multimodal, and GLM-4.7-Flash runs on a single 24 GB GPU. Newest: GLM-5.3 From ~2.5 GB 1M context gpt-ossOpenAI OpenAI's open-weight reasoning models: gpt-oss-20b runs on a 16 GB GPU and gpt-oss-120b fits on a single 80 GB card. Both are… Newest: gpt-oss-safeguard From ~16 GB 128K context GraniteIBM IBM's enterprise-focused open models: Apache 2.0, trained on cleared data, and good at RAG, tool calling and structured JSON output. Newest: Granite 4.2 From ~1.5 GB 128K context KimiMoonshot AI Moonshot AI's giant open MoE models. Kimi K3 is the first open 3T-class model, with native vision and a 1M-token context. Newest: Kimi K3.1 Pro LlamaMeta Meta's open-weight model family. Llama 3.1 is still the most-pulled model on Ollama; Llama 4 added mixture-of-experts and up to a 10M-token… Newest: Llama 4 From ~1.5 GB 10M context Meta MuseMeta Meta's newer open models: Muse Glimmer, a 30B Apache 2.0 model for local agents, and Muse Spark Open, an 816B multimodal MoE… Newest: Muse Spark Open From ~20.5 GB 128K context MistralMistral AI Mistral AI's open-weight models from France: Apache 2.0 small models, the 128B Mistral Medium 3.5 and the 1T-parameter Mistral Large 4. Newest: Mistral Large 4 From ~4 GB 256K context NemotronNVIDIA NVIDIA's open models with open weights, data and recipes: hybrid Mamba-Transformer MoE models with up to a 1M-token context. Newest: Nemotron 3.5 Super From ~4 GB 1M context PhiMicrosoft Microsoft's small language models: Phi-4 (14B) and Phi-4-mini (3.8B) punch above their size on maths and reasoning, under the MIT licence. Newest: Phi-4-reasoning From ~3.5 GB 128K context QwenAlibaba Cloud (Qwen team) Alibaba's open-weight model family, from 0.8B phone-sized models to 125B MoE, most of it under Apache 2.0. Qwen is the most popular… Newest: Qwen3.8 From ~1.5 GB 256K context Qwen CoderAlibaba Cloud (Qwen team) Qwen's code-specialised models: tiny autocomplete models, a fast 30B MoE and Qwen3-Coder-Next, an 80B MoE that activates only 3B parameters per token. Newest: Qwen3-Coder-Next From ~1 GB 256K context

API-only models

ClaudeAnthropic Anthropic's Claude models: Fable, Opus, Sonnet and Haiku, all with 1M-token context, tool use and vision, plus every API ID and retirement… Newest: Claude Fable 5.2 1M context GeminiGoogle DeepMind Google's Gemini models in the Gemini API: Gemini 4 Carbon, Argon and Hydrogen, the 3.x Flash models, and shutdown dates for 2.5. Newest: Gemini 4 Hydrogen 1M context GPTOpenAI OpenAI's GPT models in the API: GPT-6 Astra, GPT-6.1 Sol and Terra, GPT-6 Luna, older GPT-5.x and o-series, with every shutdown date. Newest: GPT-6.1 Terra 1.05M context

Embedding models

Embedding modelsGoogle, Alibaba, Nomic, BAAI, Snowflake, IBM and others Open embedding models for semantic search and RAG that run locally in Ollama, from 22M-parameter MiniLM to the multimodal EmbeddingGemma 2. Newest: EmbeddingGemma 2 From ~1 GB 256K context

How this directory works

Each family page lists the models we think matter for people who run or choose AI models: the current generation plus the older ones still in wide use. Download sizes, context windows, input types and abilities sync daily from the Ollama library and Hugging Face. API models come from the vendors' own model and deprecation pages, and we check those by hand.

"VRAM" is an estimate for Ollama's default 4-bit build with an 8K-token context: the download size plus about 10% overhead and 0.5 GB for the context. Longer contexts and higher-precision builds need more memory. Use the LLM VRAM calculator for an exact setup. We leave out benchmark scores and per-token prices on purpose: they change weekly and are better read at the source.

Back to top