Tenstorrent

Tenstorrent console API — full model catalog v3

본 스킬의 SSoT. console.tenstorrent.com 는 OpenAI-compatible chat, async image/video jobs, and audio endpoints를 노출하거나 노출 예정인 inference surface로 다룬다.

Authentication

Use Authorization: Bearer .

export TENSTORRENT_KEY="your-api-key"

Do not commit real keys or .env files.

Model Catalog (15 models)

Chat (POST /v1/chat/completions)

| Model ID | Notes | |---|---| | deepseek-ai/DeepSeek-R1-0528 | reasoning model; final answer in content, reasoning trace may appear in reasoning. | | Qwen/Qwen3-32B | default chat; thinking model; trace may appear in reasoning and reasoning_content. |

curl -sS -X POST "https://console.tenstorrent.com/v1/chat/completions" \
  -H "Authorization: Bearer $TENSTORRENT_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen/Qwen3-32B",
    "messages": [{"role":"user","content":"한국어로 한 문장 인사해줘."}],
    "max_tokens": 256
  }'

Parsing priority: content > reasoning > reasoning_content.

Image (POST /v1/image/jobs)

| Model ID | Notes | |---|---| | sdxl | default image model; 1024x1024; observed target cost around $0.02. | | tt-sd3.5 | Stable Diffusion 3.5. | | tt-z-image-turbo | Z-Image Turbo. |

curl -sS -X POST "https://console.tenstorrent.com/v1/image/jobs" \
  -H "Authorization: Bearer $TENSTORRENT_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "sdxl",
    "prompt": "a small robot smiling in a clean studio, soft light",
    "width": 1024,
    "height": 1024,
    "steps": 20
  }'

Poll GET /v1/image/jobs/{id} until status: completed, then download the returned presigned image URL immediately. Presigned artifact URLs expire.

Video (POST /v1/video/jobs)

| Model ID | Notes | |---|---| | Wan2.2-T2V-A14B-Diffusers | default video; full model; ~5s 1280x720 output; observed target cost around $0.05. | | Prodia Wan 2.2 Lighting Text to Video | distilled/Lightning display name from usage page. | | prodia/Wan2.2-T2V-A14B-Lightning | Prodia Lightning literal. | | Wan2.2-T2V-A14B-Lighting-Diffusers | usage-page spelling keeps Lighting. | | Wan2.2-T2V-A14B-Lightning | Lightning literal. | | Wan2.2-T2V-A14B-Lightning-Diffusers | Lightning Diffusers literal. | | Wan2.2-T2V-A14B-Lightning-Diffusers-FP8 | FP8 Lightning Diffusers literal. | | Wan2.2-T2V-Lightning | short Lightning literal. |

curl -sS -X POST "https://console.tenstorrent.com/v1/video/jobs" \
  -H "Authorization: Bearer $TENSTORRENT_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Wan2.2-T2V-A14B-Diffusers",
    "prompt": "A serene Korean rice field at sunrise",
    "negative_prompt": "low quality, blurry, watermark"
  }'

Poll GET /v1/video/jobs/{id} until status: completed, then download video_url first, or the first non-empty returned presigned/artifact URL if video_url is absent. In live use, status: completed can appear briefly before an artifact URL; keep polling until a URL appears. If job creation returns no id, surface the API response body instead of replacing it with a generic error.

TTS

| Model ID | Notes | |---|---| | tts-1 | Endpoint discovery order: try OpenAI-compatible POST /v1/audio/speech, then POST /v1/tts/jobs. If both are unavailable, report a visible TBD instead of silently falling back. |

Expected OpenAI-compatible body:

{"model":"tts-1","input":"안녕하세요","voice":"alloy"}

STT

| Model ID | Notes | |---|---| | whisper-large-v3 | Endpoint discovery: POST /v1/audio/transcriptions multipart upload with file and model. If unavailable, report visible TBD. |

Prompt and Artifact Rules

• Video prompt limit: keep prompts at or below 2000 characters.
• Image default: sdxl, 1024x1024, 20 steps.
• Video default: Wan2.2-T2V-A14B-Diffusers.
• Audio defaults: tts-1, whisper-large-v3.
• Presigned artifact URLs expire; download immediately into local output paths.
• No silent fallback: if an endpoint is missing, report the missing endpoint explicitly.

CLI Companion

The companion CLI is aldegad/tenstorrent-cli.

tenstorrent
/model
/image a cat
/video A serene Korean rice field at sunrise
/tts 안녕하세요
/stt ./audio.wav

Source

Catalog source: console.tenstorrent.com inference usage page as captured for the 2026-05-18 v0.2 Torendissi patch, plus live CLI endpoint probes for image/video/audio behavior.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: aldegad
• Source: aldegad/tenstorrent
• License: MIT

Install and usage instructions live in the source repository linked above.