Edgeweigh AI infrastructure

Keep model and agent work under one operating responsibility.

Arrange model access, GPU capacity, isolated execution, and the engineering that turns them into a running system.

Request access

Network context

Teams using the infrastructure behind Edgeweigh

These organizations are publicly identified as users of infrastructure in our partner network. They are not presented as Edgeweigh customers.

Hugging FaceTiDBKilo CodeQuoraOpenRouterFish AudioHygoGizmoSimularWiz

Four services. One team accountable for all of them.

Model access, GPU capacity, isolated execution, and the engineering that connects them. Start with the one that matches the work in front of you, and add another when the work calls for it.

Model APIs

Choose the model first. Keep the application path clear.

Review the latest supported models before deciding which application behavior must stay intact. One reviewed set replaces a scattered catalog, and the differences between families stay visible.

Learn more

Latest supported models

Anthropic
  • Claude Fable 5.1LanguageVisionReasoning
  • Claude Opus 5LanguageVisionReasoning
  • Claude Sonnet 5LanguageVisionReasoning
OpenAI
  • GPT-6 AstraLanguageVisionTools
  • GPT-5.6 SolLanguageVisionTools
  • GPT-Image-2ImageGenerationEditing
  • text-embedding-3-largeEmbeddingMultilingualRetrieval
  • text-embedding-3-smallEmbeddingRetrieval
Google
  • Gemini 3.8 FlashLanguageVisionTools
  • Nano Banana 2ImageGenerationEditing
  • Lyria 3.5AudioMusicGeneration
  • gemini-embedding-2EmbeddingMultimodalRetrieval
  • EmbeddingGemmaEmbeddingMultilingualRetrieval
Moonshot AI
  • Kimi K3LanguageVisionReasoning
  • Kimi K2.6LanguageVisionReasoning
  • Kimi K2.5LanguageVisionReasoning
Z.ai
  • GLM 5.3LanguageToolsReasoning
  • GLM 5.3 FlashLanguageVision
  • GLM 5.2Language
DeepSeek
  • V4 ProLanguageToolsReasoning
  • V4 FlashLanguageToolsReasoning
  • V4 Flash Vision ExpLanguageVisionTools
Qwen
  • Qwen3.8 MaxLanguageVisionTools
  • Qwen3 EmbeddingEmbeddingMultilingualRetrieval
  • Qwen3 RerankerRerankMultilingualRetrieval
  • Qwen-Image Text to ImageImageGeneration
  • Qwen-Image EditImageEditing
MiniMax
  • M3LanguageVisionReasoning
  • H3VideoGenerationEditing
  • Music 3.0AudioMusicGeneration
  • Hailuo 2.3VideoText-to-VideoImage-to-Video
Mistral
  • ShieldstralLanguageVisionSafety
  • Medium 3.5LanguageVisionTools
  • Small 4LanguageReasoningTools
ByteDance Seed
  • Seedream 5.0 ProImageGenerationReasoning
  • Seedance 2.5VideoGenerationEditing
  • Seedance 2.0VideoGenerationEditing
Kling
  • v3.0 4K Text-to-VideoVideoGeneration
  • v3.0 Pro Image-to-VideoVideoGeneration
  • V3.0 Motion ControlVideoEditing
Wan
  • 2.7VideoText-to-VideoImage-to-Video
  • 2.7 Video EditingVideoEditing
Cohere
  • rerank-v4.0-proRerankMultilingualRetrieval
  • rerank-v4.0-fastRerankMultilingualRetrieval
  • ProtocolTwo protocol modesCompatible mode for broad tooling, native mode where family-specific request and response semantics matter.
  • FamiliesReviewed model familiesLanguage, image, video, embedding and rerank families, each with its behavior fields stated.
  • ControlsAccount-level controlsCredentials, spending boundaries and network rules are set by the account owner after sign-in.
  • SupportA team on the accountThe people who scope the integration stay responsible once it is in production.

GPU Cloud

Start with the way the workload runs.

Duration, interruption tolerance, memory, and interconnect narrow the right capacity form. One capacity plan replaces stitched-together instances, and the workload fit stays explicit.

Learn more
B200 SXM 192 GB Large training · Reserved
H200 SXM 141 GB Memory-heavy inference · Reserved
H100 SXM 80 GB Training · Flexible forms
L40S 48 GB Inference · Batch
  • HardwareB200, H200, H100, A100, L40SAccelerator memory from 24 GB to 192 GB, with NVLink and RDMA on the SXM tiers.
  • FormsOn-demand, reserved, serverlessThe form follows the workload's tolerance for interruption, not the other way round.
  • WorkTraining, inference, batchDuration, continuity and memory are qualified before a hardware name enters the decision.
  • RegionsAsia Pacific, North America, EuropeRegional placement is confirmed during acceptance, alongside capacity and term.

Agent Sandboxes

Give every tool-using task a defined computer.

Code, browser, and desktop work runs inside an isolated boundary. Isolation arrives with the environment, and state, lifecycle, and network reach stay explicit.

Learn more
  • ExecuteCode, browser, desktopGenerated code, browser workflows and graphical work each run inside the same isolated boundary.
  • PreserveFilesystem, snapshots, templatesWorking state is retained across lifecycle changes and can be captured for reuse.
  • BoundNetwork policy, lifecycleOutbound access is declared, and an environment can be paused without losing approved state.
  • IntegrateClient languagesThe environment is driven from your own application through reviewed client libraries.

Agent Engineering

Design the work before selecting the agent.

Shape the process and decision boundaries before building. You own a running system rather than a prototype, and engineering responsibility stays with it as the process changes.

Learn more
Edgeweigh carriesYour team carries

Mapping the process into work an agent can carry, and setting the decision boundaries.

The process itself, the constraints around it, and the people who know how the work is done.

The agents and the shared knowledge layer, built on the three services above.

The judgement on which decisions an agent may take without review.

Evaluation against the process after launch, and the changes that follow.

The outcome the system is measured on inside the business.

  • ShapeDecompose the processThe work is split into tasks an agent can carry, and the knowledge each task needs is located.
  • BuildAgents and a shared layerAgents and the knowledge layer they share are built on the model, capacity and sandbox lines above.
  • OperateMeasure and adjustBehavior is evaluated against the process, and the system changes as the business changes.
  • OwnOne responsible teamThe team that designs the system is the team that runs it after launch.

Coordinated responsibility

You build the product.
We run the infrastructure behind it.

Use the whole system or select a single service. The same team keeps the working boundary coherent.

  • A person reviews every requestAccess is opened after a real conversation about the work, not by a signup form.
  • One team stays on the accountThe people who scope the workload are the people who operate it after launch.
  • Capacity matched to the workDuration, interruption tolerance and memory decide the form, before hardware is named.
  • One relationship across all fourModel access, capacity, execution and engineering stay under a single working boundary.

Ready to put model and agent work under one team?

Tell us what you plan to run and which services you need. A member of the team replies within one business day.

Request access