How-To List Model Features via curl and the Inference API

Introduction

Crusoe Managed Inference exposes an OpenAI-compatible /v1/models endpoint that returns the full catalog as JSON. Every model comes back with its architecture, parameter count, context length, quantization, tags, per-token pricing, and — most usefully — the exact list of request parameters it accepts.

That last field is the one worth knowing about. supported_parameters tells you up front whether a model will accept tools, response_format, reasoning_effort, or logprobs, so you can check before sending a request rather than debugging a rejection. The same applies to context_length and top_provider.max_completion_tokens when you are sizing a batch job.

This article covers listing the whole catalog, filtering to one model, and reading the fields that matter — along with the metadata the API does not return, which is just as important to know.

Prerequisites

Valid API Key for the Crusoe Inference API
curl Installed (Default on macOS and Linux; via WSL or Git Bash on Windows)
jq Installed (Optional, Recommended for Readable JSON Output) — brew install jq on macOS, apt install jq on Linux

Instructions

Step 1: List All Available Models

Set KEY to your API key in the shell first, then run:

curl -s "https://api.inference.crusoecloud.com/v1/models" \
  -H "Authorization: Bearer $KEY" | jq

This returns a JSON list of every model currently available, along with their properties.

ℹ️ Note: Export the key as a variable rather than pasting it into the command. A key typed inline is written to your shell history and is visible in the process list while the request runs.

Step 2: Filter to a Specific Model

If you only want details for one model, filter the response by model ID:

curl -s "https://api.inference.crusoecloud.com/v1/models" \
  -H "Authorization: Bearer $KEY" | \
  jq '.data[] | select(.id == "deepseek-ai/Deepseek-V4-Flash")'

Replace deepseek-ai/Deepseek-V4-Flash with the model ID you are looking for.

Example Output

{
  "id": "deepseek-ai/Deepseek-V4-Flash",
  "name": "deepseek-ai/Deepseek-V4-Flash",
  "created": 0,
  "description": "The DeepSeek-V4 series includes two strong Mixture-of-Experts (MoE) language models — DeepSeek-V4-Pro with 1.6T parameters (49B activated) and DeepSeek-V4-Flash with 284B parameters (13B activated) — both supporting a context length of one million tokens. Both models are pre-trained on more than 32T diverse and high-quality tokens, followed by a comprehensive post-training pipeline. The DeepSeek-V4 series includes two strong Mixture-of-Experts (MoE) language models — DeepSeek-V4-Pro with 1.6T parameters (49B activated) and DeepSeek-V4-Flash with 284B parameters (13B activated) — both supporting a context length of one million tokens. Both models are pre-trained on more than 32T diverse and high-quality tokens, followed by a comprehensive post-training pipeline. The post-training features a two-stage paradigm: independent cultivation of domain-specific experts (through SFT and RL with GRPO), followed by unified model consolidation via on-policy distillation, integrating distinct proficiencies across diverse domains into a single model.",
  "icon": "https://huggingface.co/api/organizations/deepseek-ai/avatar",
  "owned_by": "deepseek",
  "is_public": true,
  "type": "chat",
  "context_length": 1048576,
  "quantization": "fp8",
  "architecture": {
    "modality": "text",
    "tokenizer": "transformers",
    "instruct_type": "",
    "parameter_count": 158069433298
  },
  "tags": [
    "text to text",
    "tool calling"
  ],
  "supported_parameters": [
    "chat_template_kwargs",
    "frequency_penalty",
    "logit_bias",
    "logprobs",
    "max_completion_tokens",
    "max_tokens",
    "metadata",
    "min_p",
    "parallel_tool_calls",
    "presence_penalty",
    "prompt_cache_key",
    "reasoning_effort",
    "repetition_penalty",
    "response_format",
    "safety_identifier",
    "seed",
    "service_tier",
    "stop",
    "structured_outputs",
    "temperature",
    "tool_choice",
    "tools",
    "top_k",
    "top_logprobs",
    "top_p",
    "user"
  ],
  "is_billed_by_token": true,
  "pricing": {
    "prompt": "0.14",
    "completion": "0.28",
    "image": "0",
    "request": "0",
    "video": "0",
    "input_cache_reads": "0.03"
  },
  "top_provider": {
    "context_length": 1048576,
    "max_completion_tokens": 1048576,
    "is_moderated": false
  }
}

⚠️ Warning: Do not treat description as authoritative. It is upstream model-card prose passed through unchanged, and in the sample above it contradicts the structured data in the same response — the text claims 284B parameters while architecture.parameter_count reports roughly 158B. The opening two sentences are also repeated verbatim. Read the structured fields, not the blurb.

Notes and Limitations

No checkpoint or version date field. The API does not currently return a model's training checkpoint or release date as a structured field. If a model has multiple versions or checkpoints, this information, when available, is typically included as part of the model name, such as a date suffix like 0324, not as a separate field.
The created field is unreliable. Some models return 0 or a placeholder value. It should not be used to determine release or checkpoint date.
Version metadata varies by provider. If you need version-specific information that is not reflected in the model name or this listing, check the model documentation page for that specific model.

Example

A team is moving a structured-extraction job onto Crusoe and needs a model that will accept response_format for JSON-mode output and tools for function calling, with enough context to take a full document in one pass.

Rather than trial-and-error against the chat endpoint, they list the catalog once and filter it locally:

curl -s "https://api.inference.crusoecloud.com/v1/models" \
  -H "Authorization: Bearer $KEY" | \
  jq -r '.data[]
    | select(.supported_parameters | index("response_format") and index("tools"))
    | select(.context_length >= 131072)
    | "\(.id)\t\(.context_length)\t\(.pricing.prompt)"'

That returns only the models that can actually do the job, with their context length and prompt price, so the choice comes down to cost rather than guesswork. Checking pricing.input_cache_reads at the same time is worthwhile for a job that resends the same system prompt on every call.

Was this article helpful?

0 out of 0 found this helpful

Comments

0 comments

Article is closed for comments.