FuriosaAI

aid: furiosa
name: FuriosaAI
description: 'FuriosaAI is a South Korean semiconductor company that designs data-center AI inference chips and the software
  stack that runs on them. Its second-generation NPU, RNGD, implements a Tensor Contraction Processor architecture on TSMC
  5nm and targets LLM, multi-modal and vision inference. The developer surface is software you run yourself rather than a
  hosted API: Furiosa-LLM ships an OpenAI-compatible server exposing chat, completions, responses, embeddings, score, rerank,
  models, tokenizer and Prometheus metrics endpoints, and the earlier Furiosa Model Server implements the KServe v2 Predict
  Protocol and the Triton Model Repository extension over REST and gRPC. Around that sit Python, Go and Rust packages, a furiosa-smi
  device CLI, container images, an APT/RPM package repository, and a Kubernetes cloud-native toolkit (device plugin, DRA driver,
  feature discovery, metrics exporter, NPU operator).'
url: https://raw.githubusercontent.com/api-evangelist/furiosa/refs/heads/main/apis.yml
x-type: company
x-source: harvest:secondary-market
specificationVersion: '0.20'
created: '2026-08-16'
modified: '2026-08-16'
image: https://furiosa.ai/favicon.ico
tags:
- Artificial Intelligence
- Machine Learning
- Inference
- Semiconductors
- NPU
- Hardware
- LLM
- Model Serving
- Kubernetes
- Developer Tools
- Compute
- Infrastructure
apis:
- aid: furiosa-llm-openai-server
  name: Furiosa-LLM OpenAI-Compatible Server
  description: The HTTP server started by `furiosa-llm serve <ARTIFACT_PATH>`. It hosts a single model on RNGD NPUs and exposes
    an OpenAI-compatible surface - /v1/completions, /v1/chat/completions, /v1/responses (OpenResponses), /v1/embeddings, /v1/models,
    /v1/models/{model_id} - plus the vLLM-originated /score and /rerank pooling endpoints, a tokenizer API (/tokenize, /detokenize,
    /tokenizer_info), GET /version and a Prometheus GET /metrics endpoint. It is customer-hosted software, so the base URL
    below is templated on the operator's own host; the documented default is http://localhost:8000/v1. FuriosaAI publishes
    no OpenAPI for this surface - the parameter tables in the serving docs are the contract.
  humanURL: https://developer.furiosa.ai/latest/en/furiosa_llm/furiosa-llm-serve.html
  baseURL: http://{furiosa_llm_host}:8000/v1
  tags:
  - LLM
  - Inference
  - OpenAI Compatible
  properties:
  - type: Documentation
    url: https://developer.furiosa.ai/latest/en/furiosa_llm/furiosa-llm-serve.html
  - type: APIReference
    url: https://developer.furiosa.ai/latest/en/furiosa_llm/reference.html
  - type: Authentication
    url: authentication/furiosa-authentication.yml
  - type: Conventions
    url: conventions/furiosa-conventions.yml
  - type: RateLimits
    url: rate-limits/furiosa-rate-limits.yml
- aid: furiosa-server-predict-v2
  name: Furiosa Model Server - Predict API (KServe v2)
  description: The REST and gRPC inference surface of Furiosa Model Server, serving TFLite/ONNX models on FuriosaAI NPUs.
    Implements the KServe/KFServing V2 Dataplane - server liveness and readiness, server and model metadata, model readiness,
    and inference - at /v2/*. Customer-hosted; the REST server defaults to port 8080 and the gRPC server to 8081.
  humanURL: https://github.com/furiosa-ai/furiosa-sdk/tree/main/python/furiosa-server
  baseURL: http://{furiosa_server_host}:8080
  tags:
  - Inference
  - Model Serving
  - KServe
  properties:
  - type: OpenAPI
    url: openapi/furiosa-predict-v2.yaml
  - type: Protobuf
    url: grpc/furiosa-predict.proto
  - type: Overlay
    url: overlays/furiosa-predict-v2-overlay.yaml
  - type: ErrorCatalog
    url: errors/furiosa-problem-types.yml
  - type: DataModel
    url: data-model/furiosa-data-model.yml
  - type: SourceCode
    url: https://github.com/furiosa-ai/furiosa-sdk
- aid: furiosa-model-repository-v2
  name: Furiosa Model Server - Model Repository API
  description: The model-management surface of Furiosa Model Server, implementing the Triton Inference Server Model Repository
    extension - POST /v2/repository/index to list the repository and each model's readiness state, and POST /v2/repository/models/{name}/load
    and /unload to bring models in and out of the server. Available over both REST and gRPC. No authentication is declared,
    so this surface must stay on a trusted network.
  humanURL: https://github.com/furiosa-ai/furiosa-sdk/tree/main/python/furiosa-server
  baseURL: http://{furiosa_server_host}:8080
  tags:
  - Model Serving
  - Model Management
  - Triton
  properties:
  - type: OpenAPI
    url: openapi/furiosa-model-repository-v2.yaml
  - type: Protobuf
    url: grpc/furiosa-model-repository.proto
  - type: Overlay
    url: overlays/furiosa-model-repository-v2-overlay.yaml
maintainers:
- FN: Kin Lane
  email: kin@apievangelist.com
- FN: APIs.json
  email: info@apis.io
common:
- type: Website
  url: https://furiosa.ai/
- type: DeveloperPortal
  url: https://developer.furiosa.ai/latest/en/
- type: Documentation
  url: https://developer.furiosa.ai/latest/en/
- type: APIReference
  url: https://developer.furiosa.ai/latest/en/furiosa_llm/reference.html
- type: GettingStarted
  url: https://developer.furiosa.ai/latest/en/get_started/furiosa_llm.html
- type: Support
  url: https://furiosa-ai.atlassian.net/servicedesk/customer/portals
- type: Community
  url: https://forums.furiosa.ai/
- type: Blog
  url: https://furiosa.ai/blog
- type: GitHubOrganization
  url: https://github.com/furiosa-ai
- type: Roadmap
  url: https://developer.furiosa.ai/latest/en/overview/roadmap.html
- type: ChangeLog
  url: https://developer.furiosa.ai/latest/en/whatsnew/index.html
- type: SignUp
  url: https://lp.furiosa.ai/furiosa-access-program
- type: PrivacyPolicy
  url: https://furiosa.ai/privacy-policy
- type: Contact
  url: https://furiosa.ai/contact
- type: Packages
  url: packages/furiosa-packages.yml
- type: SDKs
  url: packages/furiosa-packages.yml
- type: CLI
  url: cli/furiosa-cli.yml
- type: Sandbox
  url: sandbox/furiosa-sandbox.yml
- type: Authentication
  url: authentication/furiosa-authentication.yml
- type: Conventions
  url: conventions/furiosa-conventions.yml
- type: ErrorCatalog
  url: errors/furiosa-problem-types.yml
- type: Lifecycle
  url: lifecycle/furiosa-lifecycle.yml
- type: Deprecation
  url: lifecycle/furiosa-lifecycle.yml
- type: Conformance
  url: conformance/furiosa-conformance.yml
- type: DataModel
  url: data-model/furiosa-data-model.yml
- type: DomainSecurity
  url: security/furiosa-domain-security.yml
- type: LLMsTxt
  url: llms/furiosa-llms.txt
- type: AgentSkill
  url: skills/_index.yml
- type: Plans
  url: plans/furiosa-plans-pricing.yml
- type: RateLimits
  url: rate-limits/furiosa-rate-limits.yml
- type: ChangeLog
  url: changelog/furiosa-changelog.yml
x-enrichment:
  date: '2026-08-16'
  status: enriched
  artifacts_added: 26
  pass: local-v1
x-base-url-note: All three baseURL values are TEMPLATED on purpose, not missing. FuriosaAI operates no hosted API; both HTTP
  surfaces are servers the customer starts on their own machine. The published defaults are http://localhost:8000/v1 for `furiosa-llm
  serve` and ports 8080 (REST) / 8081 (gRPC) for furiosa-server. repair-api-bases.py reports `no-usable-base` for this repo,
  which is the correct and expected reading of a self-hosted product rather than a defect to repair.
x-contract-discovery:
  date: '2026-08-16'
  openapi_on_api_hosts: /openapi.json, /openapi.yaml, /swagger.json and /api-docs all returned 404 on furiosa.ai and developer.furiosa.ai;
    api.furiosa.ai and docs.furiosa.ai do not resolve. The two OpenAPI documents in this repo were found instead in FuriosaAI's
    own GitHub repository at furiosa-ai/furiosa-sdk/python/furiosa-server/openapi/.
  ownership_note: 'Ownership check on those two specs. Their info.title reads "Predict API" and "Model Repository" and info.description
    points at kubeflow/kfserving and triton-inference-server rather than at FuriosaAI, and servers[] is empty. They are nonetheless
    FuriosaAI''s published contract: the furiosa-server README in the same first-party repository states the server is "compliant
    with KFServing''s V2 Dataplane specification and Triton''s Model Repository specification" and lists "OpenAPI specification
    support" as a shipped feature, so the specs describe FuriosaAI''s own product implementing a borrowed protocol verbatim
    - the same reason the .proto files carry upstream provenance headers.'
  graphql: No GraphQL surface exists or is documented.
  mcp: No MCP server, hosted or stdio, in 158 public repos or any registry. The developer center's only MCP mention is negative
    - the Responses API page states mcp built-in tools are not supported.
  a2a: /.well-known/agent-card.json and /.well-known/agent.json returned 404 on all three hosts.
  llms_txt: /llms.txt returned 404 on furiosa.ai and developer.furiosa.ai.