All models are served from PrivateMind's secure, sovereign private infrastructure. Open-weight models hosted on PrivateMind-controlled hardware within your environment. No traffic is routed to public clouds, and no data is used for training.
The catalog differs by environment and organisation, and it changes over time. Call GET /v1/models to see what you can use, and check that a model supports a feature before you rely on it.
List models
curl -s "https://api.privatemind.com/v1/models" \
-H "Authorization: Bearer $PMIND_KEY"The response follows OpenAI's format and adds these fields to each model:
model_type: what kind of model it is, which tells you the endpoint to callcapabilitiesandsupported_parameters: what the model acceptsaliases: role names that point to the model, such asfastcost: the published pricebatch_eligible: whether the Batch API accepts the modelowned_by: the organisation that publishes itopen_weights:truefor the open-weight models PrivateMind hostsdescription,model_full_nameandmodel_icon_url: optional, for building a model picker
{
"object": "list",
"data": [
{
"id": "qwen3-8-27b-nvfp4",
"object": "model",
"context_length": 262144,
"model_type": "vision-chat",
"model_full_name": "Qwen 3.8 27B",
"model_icon_url": "/model-icons/qwen.svg",
"capabilities": {
"tools": true,
"response_format": true,
"reasoning_effort": true,
"image_input": true
},
"supported_parameters": [
"max_tokens", "temperature", "top_p", "top_k", "seed",
"frequency_penalty", "presence_penalty", "stream", "stop",
"reasoning", "include_reasoning",
"tools", "tool_choice", "response_format", "reasoning_effort"
],
"aliases": ["fast", "vision"],
"batch_eligible": true,
"cost": {
"input_per_m_token": 0.27,
"output_per_m_token": 0.85
}
}
]
}Get one model
GET /v1/models/{id} returns a single entry of the list, by id or alias, with the same visibility rules. An id you cannot use returns 404 with code model_not_found.
curl -s "https://api.privatemind.com/v1/models/fast" \
-H "Authorization: Bearer $PMIND_KEY"Messages API
Send an anthropic-version header, as the Anthropic SDKs always do, and the same paths answer in Anthropic's format. The list then holds only the models the Messages API can serve: the chat and vision-chat models PrivateMind hosts. Models from external providers are left out. Each entry uses the model's real id. Aliases such as fast are not listed, but they work for retrieve and as model:
import anthropic
client = anthropic.Anthropic(base_url="https://api.privatemind.com", api_key="PMIND...:...")
for m in client.models.list():
print(m.id, m.display_name, m.max_input_tokens)
print(client.models.retrieve("fast")){
"data": [
{
"type": "model",
"id": "qwen3-8-27b-nvfp4",
"display_name": "Qwen 3.8 27B",
"created_at": "1970-01-01T00:00:00Z",
"max_input_tokens": 262144,
"max_tokens": 40000
}
],
"has_more": false,
"first_id": "qwen3-8-27b-nvfp4",
"last_id": "qwen3-8-27b-nvfp4"
}display_nameismodel_full_name, or the id when none is set.max_input_tokensiscontext_length;max_tokensis the model's output cap, ornullwhen it declares none.created_atis always the Unix epoch, which Anthropic uses for an unknown date. Our deployments carry no release date.- The list pages with
limit(1 to 1000, default 20) andbefore_idorafter_id. The SDK pages for you.
Capabilities, pricing and supported_parameters are only in the OpenAI-shaped list above.
Capability discovery
Check capabilities and supported_parameters for the feature you need, not the model id. Models are added and removed, so a check like if (model === 'fast') breaks without warning. Test the feature instead: model.capabilities.tools, or model.supported_parameters.includes('reasoning_effort').
capabilities
What the model accepts. A model lists only the capabilities it has, so a missing flag means false.
| Flag | Meaning |
|---|---|
tools |
Accepts tools / tool_choice for function calling. See Tool use |
response_format |
Accepts response_format for JSON mode / structured outputs |
reasoning_effort |
Supports the reasoning_effort parameter (hybrid and thinking-only models) |
reasoning_effort_levels |
The effort levels this model honours, in order, for example ["off", "low", "medium", "high"]. Missing when reasoning_effort is only an on/off switch. Includes "off" when thinking can be turned off |
reasoning_effort_default |
The level applied when a request omits reasoning_effort, for example "high" |
reasoning_mandatory |
Present and true when thinking cannot be turned off. reasoning_effort: "off" returns 400 on these models |
image_input |
Accepts images in chat messages. See Vision |
supported_parameters
Every request parameter the model accepts, in the format OpenRouter uses. It includes reasoning and include_reasoning when the model returns its chain of thought (see Streaming → Reasoning output), and reasoning_effort, tools and response_format when the matching flag above is set. Embeddings, speech and image models have an empty list, because they take none of the chat sampling parameters. OCR, guard and reranker models list the sampling controls they accept.
batch_eligible
true when you can name the model, by id or alias, in a Batch API request. A batch line naming any other model fails with model_not_eligible.
model_type
What kind of model it is, and so which endpoint to call:
model_type |
Endpoint |
|---|---|
chat, vision-chat |
/v1/chat/completions or /v1/messages (vision-chat also accepts image input) |
embeddings |
/v1/embeddings |
reranker |
/v1/rerank |
ocr, ocr-vision |
/v1/chat/completions (document extraction) |
tts |
/v1/audio/speech |
asr |
/v1/audio/transcriptions |
image-gen |
/v1/images/generations |
guard |
/v1/moderations |
In the table at the top of this page, a Hybrid badge marks a model that can either think first or answer directly, chosen per request. Set reasoning_effort on /v1/chat/completions: "off" turns thinking off and any other accepted value turns it on.
How much thinking each level gives differs by model, so read reasoning_effort_levels and don't assume a scale. A model with no list treats every value that turns thinking on the same way. On a model with a list, two neighbouring levels can still think to the same depth.
Display metadata
Two optional fields help you label models in a picker. Either can be missing, which is normal. Whatever you display, send the id as model.
| Field | Meaning |
|---|---|
model_full_name |
A readable name, for example "Qwen 3.8 27B". Show the id when it is missing. |
model_icon_url |
Relative path to the brand icon the PrivateMind app shows for the model, for example /model-icons/qwen.svg. It points at our own assets, so your app will usually use its own icons. Missing when the model has no brand icon. |
Pricing
A model with a published price has a cost object. Token rates are in US dollars per million tokens:
| Field | Meaning |
|---|---|
input_per_m_token |
USD per 1M prompt (input) tokens |
output_per_m_token |
USD per 1M completion (output) tokens |
image_per_generation |
USD per generated image. Present only on image-generation models; omitted otherwise |
Use these list prices to estimate spend before a call: prompt_tokens / 1e6 × input_per_m_token + completion_tokens / 1e6 × output_per_m_token. Embeddings bill input tokens only.
What a request actually cost is its row in /v1/usage, priced from the token counts in the response's usage block. See also Rate limits & budgets.
Context length
context_length is the most tokens a request can hold, prompt and completion together. Keep max_tokens within it. A request that goes over returns 400.
What the list contains
The list holds exactly the models you can call: those live in this environment that your organisation is allowed to use.
Cache the list for a few minutes if you call it often, and let your code fall back to another model if one disappears.
Model register
The register lists each model offered to every organisation: who publishes it, its licence and version, where it runs and the checks it has had. How we test models explains each check.
- Publisher is whoever publishes the exact weights we serve. When another company made the model, for example when the publisher made a quantised build of it, the maker is named under the publisher.
- Version is the first 12 characters of the commit that identifies those weights in the publisher's repository.
- Weights check compares every file we serve with the hash recorded for it in the publisher's repository at that commit.
Updated 13 Nov 2026.
Source and hosting
| Model | Publisher | Licence | Version | Hosting | Support |
|---|---|---|---|---|---|
Chatterbox Turbo 350Mchatterbox-turbo |
Resemble AI | MIT | 749d1c1a46eb |
London | Supported |
Cosmos 3 Super 64Bcosmos3-super-text2image |
NVIDIA | OpenMDW-1.1 | 754f0cd5a7eb |
London | Supported |
DeepSeek V4.1 Flash 552B-16Bdeepseek-v4-1-flash |
DeepSeek | MIT | 1bae0e108899 |
London | Supported |
GLM 5.3 753B-40Bglm-5-3-nvfp4 |
RadixArk Made by Z.ai |
GLM-5.3 License | 11af4cba759e |
London | Supported |
Kimi K3.1 Pro 2.8T-104Bkimi-k3-1-pro |
Moonshot AI | Kimi K3.1-Pro License | 6d0a02029567 |
London | Supported |
KX qSQL 9Bqsql-qwen35-9b |
KX | Apache-2.0 | 7dd55d3f873c |
London | Supported |
Nemotron 3 Embed 8Bnemotron-3-embed-8b |
NVIDIA | OpenMDW-1.1 | d1f2f25730bb |
New York | Supported |
Nemotron 3.5 Super VL 120B-12Bnemotron-3-5-super-vl-nvfp4 |
NVIDIA | OpenMDW-1.1 | 59b8721646c3 |
London | Supported |
olmOCR 2 7Bolmocr-2-7b-1025-fp8 |
Allen AI | Apache-2.0 | 40bd7202494b |
New York | Supported |
Qwen 3.8 27Bqwen3-8-27b-nvfp4 |
NVIDIA Made by Alibaba |
Apache-2.0 | 482ca0f38322 |
London | Supported |
Qwen3 Reranker 4Bqwen3-reranker-4b |
Alibaba | Apache-2.0 | 22e683669bc0 |
New York | Supported |
Qwen3Guard 8Bqwen3guard-gen-8b |
Alibaba | Apache-2.0 | 4505cb1a6f18 |
London | Supported |
Voxtral Realtime 4Bvoxtral-mini-4b-realtime |
Mistral AI | Apache-2.0 | 2769294da956 |
New York | Supported |
Checks
| Model | Weights check | Safety tests | Capability tests | Approval date |
|---|---|---|---|---|
| Chatterbox Turbo 350M | Verified on 11 Nov 2026 10 of 10 files matched |
Not yet run | Completed | 13 Nov 2026 |
| Cosmos 3 Super 64B | Verified on 11 Nov 2026 83 of 83 files matched |
Not yet run | Completed | 13 Nov 2026 |
| DeepSeek V4.1 Flash 552B-16B | Verified on 11 Nov 2026 83 of 83 files matched |
Completed | Completed | 13 Nov 2026 |
| GLM 5.3 753B-40B | Verified on 11 Nov 2026 58 of 58 files matched |
Completed | Completed | 13 Nov 2026 |
| Kimi K3.1 Pro 2.8T-104B | Verified on 11 Nov 2026 112 of 112 files matched |
Completed | Completed | 13 Nov 2026 |
| KX qSQL 9B | Verified on 11 Nov 2026 14 of 14 files matched |
Not yet run | Completed | 13 Nov 2026 |
| Nemotron 3 Embed 8B | Verified on 11 Nov 2026 18 of 18 files matched |
Not yet run | Completed | 13 Nov 2026 |
| Nemotron 3.5 Super VL 120B-12B | Verified on 11 Nov 2026 114 of 114 files matched |
Completed | Completed | 13 Nov 2026 |
| olmOCR 2 7B | Verified on 11 Nov 2026 20 of 20 files matched |
Not yet run | Completed | 13 Nov 2026 |
| Qwen 3.8 27B | Verified on 11 Nov 2026 19 of 19 files matched |
Completed | Completed | 13 Nov 2026 |
| Qwen3 Reranker 4B | Verified on 11 Nov 2026 17 of 17 files matched |
Not yet run | Completed | 13 Nov 2026 |
| Qwen3Guard 8B | Verified on 11 Nov 2026 15 of 15 files matched |
Not yet run | Completed | 13 Nov 2026 |
| Voxtral Realtime 4B | Verified on 11 Nov 2026 9 of 9 files matched |
Not yet run | Completed | 13 Nov 2026 |
Where next
- Chat completions to send a request with one of these ids or aliases.
- Tool use and Vision for two of the capabilities above.
- Errors for what happens when a model is unavailable or out of budget.
- How we test models for the checks behind the model register.