Reasoning and generation
Frontier and open-weight language models for planning, tool calling, drafting and extraction, with vision where the model supports it.
Model Hub
Every model. One control plane.
Language, image, video, speech, music, transcription and embeddings, from frontier APIs and from open-weight models running on your own GPUs. Your agents ask for a model; the hub decides which preset and which deployment answers.
On-premise inferenceAny providerSwap without redeploying
In the catalog today
OpenAIGoogleAnthropicRunwareDeepgramAssemblyAIOllama and vLLM on your own GPUs
LanguageImageVideoSpeechMusicTranscriptionEmbedding
Why this layer exists
Models move. Your agents should not.
Nothing about a model is permanent. A provider posts a sunset date, renewal reprices you, a region runs out of capacity, procurement asks for a second source. Each is a configuration change here and a migration project everywhere else.
Retired
A sunset notice lands. You repoint one logical model at a live deployment, and the agents that name it never learn there was a notice.
Repriced
Your rate moves, or something cheaper ships. Point the Balanced variant at the new deployment, watch the metering for a week, and the old one is still a click away.
Throttled
A provider caps you at the worst hour of the week. The same logical model resolves to a second deployment and the queue drains.
How it resolves
The model your agents name is not where it runs
A logical model is what your agents ask for. A variant is how it should behave. A deployment is where it actually runs. Change the bottom layer and the top layer never notices.
Logical modelsWhat agents ask for
GPT-5gpt-5
Gemma 4gemma-4
Claude Sonnet 4.6claude-sonnet-4-6
VariantsHow it should behave
Fast
BalancedRecommended
High quality
DeploymentsWhere it actually runs
Google Vertex MaaSvertex-maas
vLLM · your GPUsvllm
Logical models
GPT-5gpt-5
Gemma 4gemma-4
Claude Sonnet 4.6claude-sonnet-4-6
Variants
Fast
BalancedRecommended
High quality
Deployments
Google Vertex MaaSvertex-maas
vLLM · your GPUsvllm
Your agents still ask for gemma-4. Only the deployment changed.
The catalog
One catalog, every modality
Language, image, video, speech, music, transcription and embeddings sit side by side in a single workspace catalog, each one carrying its provider, its identifier, its status and its rate.
31 models
GPT-5
gpt-5openai
GPT-5.2
gpt-5_2openai
Gemini 2.5
gemini-2_5google
Gemini 3
gemini-3google
Gemini 3 Flash
gemini-3-flashgoogle
Gemini 2.5 Flash Lite
gemini-2_5-flash-litegoogle
Gemini 3.1 Flash Lite
gemini-3_1-flash-litegoogle
Gemini 3.1 Pro
gemini-3_1-progoogle
Gemini 3.5 Flash
gemini-3_5-flashgoogle
Gemma 4
gemma-4google
Claude Haiku 4.5
claude-haiku-4-5anthropic
Claude Sonnet 4.6
claude-sonnet-4-6anthropic
Claude Opus 4.7
claude-opus-4-7anthropic
Qwen 3 14B
ollama-qwen3-14bollamaSelf-hosted
Gemma 3 4B Vision
ollama-gemma3-4b-visionollamaSelf-hosted
Gemma 4 (open weights)
gemma-4-vllmvllmSelf-hosted
GPT Image 1.5
gpt-image-1_5openai
Nano Banana
nano-bananagoogle
Nano Banana Pro
nano-banana-progoogle
Kling Video 2.6 Pro
kling-v2_6runware
Kling Video 3.0
kling-v3runware
Kling Video 3.0 Pro
kling-v3-prorunware
Kling Video O3 Pro
kling-o3-prorunware
Kling Avatar 2.0 Standard
kling-avatar-2_0-standardrunware
Kling Avatar 2.0 Pro
kling-avatar-2_0-prorunware
ElevenLabs v3
eleven-v3runware
ElevenLabs Turbo v2
eleven-turbo-v2runware
ElevenLabs Music v1
eleven-music-v1runware
Deepgram Nova-3
deepgram-nova-3deepgram
AssemblyAI Best
assemblyai-bestassemblyai
Gemini Embedding 001
gemini-embedding-001google
Relative cost bands, not quoted rates. Every model surfaces its own per-1K, per-second or per-minute pricing in the console.
Variants
Tune the trade-off, not the code
A variant is a saved preset over a base model. Point an agent at the preset instead of the raw model, and you can retune latency, cost and quality centrally, without touching a single agent.
Balanced quality and speedRecommended
The default for most agents. Enough headroom for tool calling and multi-step work without paying frontier rates on every turn.
Illustrative
Modalities
Add a modality, not a vendor
Speech, video and embeddings are governed, metered and deployed exactly like a language model. Giving an agent a voice is a line of configuration, not a new contract, a new key and a new line on the bill.
Image generation
Generate and edit imagery inside a workflow, with the same governance and metering as every other model in the catalog.
Video generation
Text- and image-to-video, including avatar presenters. Billed by the second rather than by the token, and surfaced that way.
Text to speech
Give an agent a voice. Low-latency turbo voices for interactive use, higher-fidelity voices for produced output.
Music generation
Score and sound-bed generation for produced media, available to agents as an ordinary tool call.
Speech to text
Transcribe audio and video across 120+ languages, with speaker diarization. A first-class modality, not an add-on.
Embeddings
Vector representations that power grounded retrieval over your indexed files and knowledge graph.
What it buys you
The next model is a decision, not a migration
A lab ships something better on a Tuesday. You add the deployment, point the Balanced variant at it, and every agent in the workspace is on it that afternoon. The old deployment stays active for as long as it takes to trust the new numbers. Model choice stops being a quarterly project.
Run the old model and the new one against the same work. Settle the argument on your numbers, not somebody else's benchmark.
Give one agent a voice before the business case exists. A week of metering says whether the deployment stays on.
Work you had ruled out because the data cannot leave the building becomes ordinary work, on the same catalog as everything else.
Sovereignty
Frontier in the cloud, or nothing leaves the building
The same catalog, the same agents, the same presets. What changes is where inference happens. Run open-weight models on your own GPUs through Ollama or vLLM and no prompt, document or embedding ever crosses your network boundary.
Hosted frontier models
- Managed capacity from OpenAI, Google Vertex and Anthropic
- Specialist providers for image, video, speech and transcription
- Per-token, per-second and per-minute rates surfaced per deployment
openaivertex-maasanthropicrunware
Your GPUs, air-gapped
- Open-weight language and embedding models served by Ollama or vLLM
- Prompts, documents and embeddings stay inside
- Offline, signed model updates for air-gapped sites
ollamavllm
Most deployments run both. Language and embeddings serve from your own GPUs, and the modalities that still need a provider are in the catalog because an administrator put them there. Anything that must not leave the building is not in it.
See how it lands in your data centreGovernance
Every model accounted for
The catalog is workspace-scoped and administered, not a free-for-all. Administrators decide which models exist, which are recommended, and which credentials they run under.
At a glance
- Catalog scope
- Models, variants and deployments are configured per workspace
- Status control
- Any model or deployment can be made active or inactive centrally
- Recommended presets
- Mark the variant teams should reach for by default
- Pricing visibility
- Per-1K token, per-second and per-minute rates surfaced per deployment
- Metering
- Usage attributed by model, agent and group across the billing period
- Credentials
- Organization keys override platform defaults, and stay masked
- Local runtimes
- Ollama and vLLM deployments served from your own hardware
- Failover
- Repoint a logical model at another deployment without redeploying agents
A catalog only administrators can change, running under credentials you control.
SOC 2 Type IIAICPA
HIPAACompliant
GDPRReady
ISO 27001Certified
See Exemplary AI in action
Book a demo and we'll show you purpose-built agents grounded in your own knowledge — deployed on your infrastructure.