Claude Fable 5.1
Anthropic's most capable generally available model, built for demanding reasoning and long-running agent work.
- Context
- 1M
- Input
- Text, Image
Claude Opus 5.5
Anthropic's recommended default for demanding work: long-running agentic coding and knowledge work, with adaptive thinking always on.
- Context
- 1M
- Input
- Text, Image
Claude Sonnet 5.5
Anthropic's balance of speed and intelligence for everyday coding, agent and enterprise work, with adaptive thinking on by default.
- Context
- 1M
- Input
- Text, Image
Claude Haiku 4.5
Anthropic's fastest and lowest-cost current model, with a 200K context window and optional extended thinking.
- Context
- 200K
- Input
- Text, Image
GPT-6 (Astra, Sol, Luna)
OpenAI's current reasoning family in three tiers: Astra for the hardest work, Sol for near-Astra results at lower cost, Luna for high volume.
- Context
- 1.05M
- Input
- Text, Image
GPT-5.6 (Sol, Terra, Luna)
OpenAI's previous generation, which introduced the Sol, Terra and Luna tiers: flagship, balanced and lowest cost.
- Context
- 1.05M
- Input
- Text, Image
gpt-oss (120b, 20b)
OpenAI's open-weight mixture-of-experts reasoning models under Apache 2.0: 120b fits one 80 GB GPU, 20b runs on 16 GB devices.
- Context
- 131K
- Input
- Text
Gemini 3.8 Flash
Google's current generally available Flash model, aimed at long-horizon engineering, autonomous agents and enterprise workflows at Flash speed and cost.
- Context
- 1.05M
- Input
- Text, Image, Video, Audio, PDF
Gemini 3.1 Pro
Google's current Pro model, in preview on the Gemini API, tuned for complex reasoning, software engineering and precise multi-step tool use.
- Context
- 1.05M
- Input
- Text, Image, Video, Audio, PDF
Gemini 3.5 Flash-Lite
Google's lowest-cost, lowest-latency 3.5 model, built for high-throughput sub-agent work, document parsing and simple extraction.
- Context
- 1.05M
- Input
- Text, Image, Video, Audio, PDF
Gemma (Gemma 4, Gemma 3)
Google's open-weight family, from edge-sized models to a 31B dense and a 26B mixture-of-experts model in Gemma 4.
- Context
- –
- Input
- Text, Image
Grok 4.7
xAI's frontier model for coding, agentic tasks and knowledge work, with text and image input and selectable reasoning effort.
- Context
- 500K
- Input
- Text, Image
DeepSeek Flash (V4.1)
DeepSeek's current Flash model with native image understanding, a 1M-token context and thinking on by default; the weights are MIT-licensed.
- Context
- 1M
- Input
- Text, Image
DeepSeek V4 Pro
DeepSeek's large V4 Pro model: text-only through the API, with a 1M-token context, tool calls and selectable thinking effort; the weights are MIT-licensed.
- Context
- 1M
- Input
- Text
DeepSeek-R1 Distill (8B)
A small open reasoning model: Qwen3-8B trained on DeepSeek-R1's chain of thought, for local use.
- Context
- 128K
- Input
- Text
Kimi K3
Moonshot AI's flagship with a 1M-token context, native image and video understanding and always-on thinking; the weights are published.
- Context
- 1.05M
- Input
- Text, Image, Video
Kimi K2.7 Code
Moonshot AI's dedicated coding model with a 256K context, always-on thinking, tool calling and a separate high-speed variant.
- Context
- 256K
- Input
- Text, Image, Video
Qwen 3.8 (Max, Flash)
Alibaba's Qwen 3.8 line: Max is the flagship, and Max and Flash both offer a 1M-token context, image and video input and hybrid thinking.
- Context
- 1M
- Input
- Text, Image, Video
Qwen 3.7 Plus
Qwen 3.7 Plus and Flash offer a 1M-token context, image and video input, hybrid thinking and function calling for multimodal agent work.
- Context
- 1M
- Input
- Text, Image, Video
GLM-5.3
Z.ai's flagship for software engineering and agent work, with a 1M-token context, up to 128K output and published weights; the Flash variants add image and video input.
- Context
- 1M
- Input
- Text
MiniMax M3
MiniMax's natively multimodal coding and agent model, with up to 1M tokens of context and thinking interleaved between tool calls; the weights are published.
- Context
- 1M
- Input
- Text, Image, Video
Mistral Large 3
Mistral's open-weight mixture-of-experts flagship (41B active, 675B total parameters) with image understanding, under Apache 2.0.
- Context
- 256K
- Input
- Text, Image
Mistral Medium 3.5
A dense 128B multimodal model that combines instruction following, reasoning and coding, with reasoning effort set per request.
- Context
- 256K
- Input
- Text, Image
Mistral Small 4
An open-weight mixture-of-experts model (about 6B active parameters) that combines instruct, reasoning and coding, with image input.
- Context
- 256K
- Input
- Text, Image
Ministral 3 (3B, 8B, 14B)
Small open-weight models in three sizes, each with base, instruct and reasoning variants and image understanding, for local and edge use.
- Context
- 256K
- Input
- Text, Image
Codestral
Mistral's coding model for low-latency, high-frequency tasks such as fill-in-the-middle and code generation.
- Context
- 128K
- Input
- Text
Llama 4 (Scout, Maverick)
Meta's natively multimodal mixture-of-experts models with 17B active parameters: Scout (109B total) and Maverick (400B total).
- Context
- 1M
- Input
- Text, Image
Llama 3.3 70B
Meta's 70B text model for multilingual dialogue in eight languages, with tool use.
- Context
- 128K
- Input
- Text
Llama 3.2 1B
A 1.2B-parameter multilingual model for constrained and on-device environments, small enough to run inside the EDDI JVM.
- Context
- 128K
- Input
- Text
Amazon Nova (Pro, Lite)
Amazon's multimodal models on Bedrock that accept text, image and video: Pro is the balanced tier, Lite the low-cost one.
- Context
- 300K
- Input
- Text, Image, Video
Cohere Command A
Cohere's 111B enterprise model focused on tool use, retrieval and 23 languages; Oracle also offers reasoning and vision variants.
- Context
- 256K
- Input
- Text
Phi-4-mini
Microsoft's 3.8B open model for memory-constrained, latency-sensitive work and maths and logic reasoning.
- Context
- 128K
- Input
- Text
No model matches this filter.
Where Models Run
The same model is often available from more than one place. Choose by where your data may go, what your team already runs, and how you authenticate.
Amazon Bedrock
Managed AWS access to models from Anthropic, Meta, Amazon, OpenAI, Mistral and others.
- Providers see no customer prompts or completions
- Private networking through VPC and PrivateLink
- In-region and cross-region inference for data residency
Google Vertex AI
Google Cloud's AI platform, now also called Gemini Enterprise Agent Platform, serving Gemini and Model Garden models.
- Regional endpoints keep processing in one jurisdiction
- VPC Service Controls for network isolation
- Authentication through Google Cloud credentials
Azure OpenAI
OpenAI models hosted in your Azure subscription, now part of Microsoft Foundry, called by deployment name.
- Data Zone deployments keep processing in the EU, US or APAC
- Private endpoints on the Azure backbone
- Provisioned throughput for reserved capacity
Oracle OCI Generative AI
Managed Oracle Cloud service for Cohere, Meta Llama and other models, on demand or on dedicated clusters.
- Dedicated AI clusters
- Private endpoints and data isolation
- Governed by OCI IAM policies
Groq
An inference cloud that runs open models on Groq's own LPU chips, through an OpenAI-compatible API.
- Very low latency
- Open models such as gpt-oss and Qwen
- One key, no infrastructure
OpenRouter
One API to hundreds of models from many providers, with an auto router that picks a model per request.
- One key for many vendors
- Automatic fallback between providers
- openrouter/auto chooses a model for each task
Ollama
Runs open models on your own hardware behind a local API, on the same Docker network as EDDI if you like.
- No cloud connection needed
- Data never leaves your machines
- One command to pull a model
Hugging Face
Hosted inference for open models by repository id, routed to partner providers under one token.
- Open models by repository id
- One token across providers
- No infrastructure to run
Jlama
A pure-Java inference engine that runs small models inside the EDDI JVM, with no model server at all.
- Nothing else to deploy
- Quantized models for a small footprint
- Works air-gapped once the weights are cached
Not in the catalog?
Any OpenAI-compatible endpoint works through type: openai and a baseUrl, including Cohere, Together, vLLM, llama-server and corporate gateways. OpenRouter reaches hundreds more models with one key.
Use More Than One
EDDI can chain models in a cascade: a fast, inexpensive model answers first, and a stronger one is called only when confidence is low. See Model Cascading.