Token Factory

Models

Text models share one endpoint and one request format; video models share the video API. One API key reaches all of them; name the model in each request.

Text models

Chat completions at /v1/chat/completions, whole or streamed. Prices are on the Pricing page.

Qwen-SEA-LION-v4.5-27B-IT

Live

qwen-sea-lion-v4.5-27bAI Singapore

AI Singapore's instruction-tuned model for Southeast Asian languages, alongside English, fine-tuned from Qwen3.6-27B. AI Singapore is a MarsCompute collaborator.

  • 128K context
  • JSON output

Always reasons before it answers, and reasoning counts as output. Set max_tokens to at least 2,048, or leave it out for the default of 65,536. Runs on GPU servers that MarsCompute operates.

Model card

Kimi K3

Live

k3Moonshot AI

Moonshot AI's open-weight flagship, a mixture-of-experts model built for long-horizon coding, knowledge work and reasoning.

  • 1M context
  • Tools
  • JSON output

Reasons before it answers without showing the reasoning, and reasoning counts as output: about half of its output tokens in our tests. A small max_tokens can leave the answer empty, so leave it out for the default of 65,536. Each request also counts about 80 input tokens beyond your messages.

Model card

Kimi K2.7 Code

Live

kimi-k2.7-codeMoonshot AI

The coding model of Moonshot AI's Kimi K2 family, built to carry programming tasks through to the end over long contexts.

  • 256K context
  • Tools
  • JSON output

Reasons before it answers, and reasoning counts as output. Set max_tokens to at least 2,048, or leave it out for the default of 65,536.

Model card

Kimi K2.6

Live

kimi-k2.6Moonshot AI

Moonshot AI's Kimi K2.6, for long-horizon coding, front-end generation and multi-agent work.

  • 256K context
  • Tools

Reasons before it answers, and reasoning counts as output. Set max_tokens to at least 2,048, or leave it out for the default of 65,536.

Model card

GLM-5.3

Live

glm-5.3Z.ai

Z.ai's flagship GLM-5.3: a 753-billion-parameter open-weights mixture-of-experts model, built above all for coding, with a context of up to a million tokens.

  • 1M context
  • Tools
  • JSON output

It may reason before it answers, and reasoning counts as output. Set max_tokens to at least 2,048, or leave it out for the default of 65,536.

Model card

GLM-5.3 Flash

Live

glm-5.3-flashZ.ai

The fast, low-cost member of Z.ai's GLM-5.3 family, a mixture-of-experts model with hybrid attention for long inputs.

  • 1M context
  • Tools
  • JSON output

Always reasons before it answers, and reasoning counts as output. Set max_tokens to at least 2,048, or leave it out for the default of 65,536. It may answer in text when a tool call is forced with tool_choice.

Model card

GLM-5.2

Live

glm-5.2Z.ai

Z.ai's 750-billion-parameter mixture-of-experts model for project-level coding, long-context work and agents, with a context of up to a million tokens.

  • 1M context
  • Tools
  • JSON output

Reasons before it answers, and reasoning counts as output. Set max_tokens to at least 2,048, or leave it out for the default of 65,536.

Model card

GLM-5.1

Live

glm-5.1Z.ai

Z.ai's GLM-5.1, a 750-billion-parameter mixture-of-experts model with strong coding and long-horizon task handling.

  • 200K context
  • Tools
  • JSON output

Reasons before it answers, and reasoning counts as output. Set max_tokens to at least 2,048, or leave it out for the default of 65,536.

Model card

GLM-5 Turbo

Live

glm-5-turboZ.ai

Z.ai's GLM-5 Turbo, built for fast inference in agent workflows.

  • 198K context
  • Tools

Reasons before it answers, and reasoning counts as output. Set max_tokens to at least 2,048, or leave it out for the default of 65,536. It does not take stop, and it refuses a forced tool_choice.

DeepSeek V4 Pro

Live

deepseek-v4-proDeepSeek

DeepSeek's large V4 mixture-of-experts model, in its generally available 0813 release, for advanced reasoning, coding, tool use and long agent workflows.

  • 1M context
  • Tools
  • JSON output

Reasons before it answers, and reasoning counts as output. Set max_tokens to at least 2,048, or leave it out for the default of 65,536. A request can also count up to about 80 input tokens beyond your messages.

Model card

DeepSeek V4.1 Flash

Live

deepseek-v4.1-flashDeepSeek

DeepSeek's fast, low-cost V4.1 model for everyday chat, coding and agent workflows.

  • 1M context
  • Tools
  • JSON output

It may reason before it answers, and reasoning counts as output. A small max_tokens can then leave the answer empty, so leave it out for the default of 65,536.

Model card

MiniMax M3

Live

minimax-m3MiniMax

MiniMax's 428-billion-parameter mixture-of-experts reasoning model for coding and agent workflows.

  • 512K context
  • Tools
  • JSON output

Reasons before it answers, and reasoning counts as output. Set max_tokens to at least 2,048, or leave it out for the default of 65,536. Each request also counts about 180 input tokens beyond your messages. It does not take stop.

Model card

MiniMax M3.1 Flash

Live

minimax-m3.1-flashMiniMax

MiniMax's fast mixture-of-experts model for coding and agent workflows: about 428 billion parameters with 23 billion active per token.

  • 1M context
  • Tools
  • JSON output

Always reasons before it answers, and reasoning counts as output: a 120-word answer used about 1,800 output tokens in our tests. Leave max_tokens out for the default of 65,536, or allow more. Each request also counts about 175 input tokens beyond your messages.

Model card

Qwen3.8-27B

Live

qwen3.8-27bQwen

Qwen's dense 27-billion-parameter reasoning model for coding, professional work, research and long agent tasks.

  • 256K context
  • Tools
  • JSON output

Reasons before it answers, and reasoning counts as output. Set max_tokens to at least 2,048, or leave it out for the default of 65,536. A stop sequence can match inside its reasoning and end the answer empty.

Model card

Qwen3.6-35B-A3B

Live

qwen3.6-35b-a3bQwen

An open-weight mixture-of-experts model from Qwen, 35 billion parameters with 3 billion active for each token.

  • 256K context
  • Tools
  • JSON output

Reasons before it answers, and reasoning counts as output. Set max_tokens to at least 2,048, or leave it out for the default of 65,536.

Model card

Qwen3-30B-A3B-Instruct-2507

Live

qwen3-30b-a3b-instruct-2507Qwen

A small, fast mixture-of-experts model, 30.5 billion parameters with 3.3 billion active, for chat, coding and tool use.

  • 256K context
  • Tools
  • JSON output

It does not reason first, so replies are direct and low-cost. It does not take stop.

Model card

MiMo-V2.6-Flash

Live

mimo-v2.6-flashXiaomi

Xiaomi's fast MiMo model for chat, coding and agent workflows.

  • 1M context
  • Tools
  • JSON output

Thinking is switched off, so replies are direct. It does not take stop.

Model card

Gemma 3 27B

Live

gemma-3-27b-itGoogle

Google's mid-size open model for instruction following, coding and many languages, with a 110K-token context. It answers directly and takes no tools.

  • 110K context
  • JSON output

It answers directly, without reasoning first. It does not take tools.

Model card

Hermes 4 405B

Live

hermes-4-405bNous Research

Nous Research's 405-billion-parameter hybrid-reasoning model, trained on verified reasoning traces for maths, coding and step-by-step work. It answers directly by default.

  • 128K context
  • JSON output

It answers directly, without reasoning first. It does not take tools.

Model card

Nemotron 3 Ultra 550B-A55B

Live

nemotron-3-ultra-550b-a55bNVIDIA

NVIDIA's 550-billion-parameter hybrid mixture-of-experts model, 55 billion active, for demanding multi-agent work and complex reasoning, with a context of about a million tokens.

  • 1M context
  • Tools
  • JSON output

Reasons before it answers, and reasoning counts as output. Set max_tokens to at least 2,048, or leave it out for the default of 65,536.

Model card

Nemotron 3.5 Lightning

Live

nemotron-3.5-lightningNVIDIA

NVIDIA's fast 30-billion-parameter hybrid mixture-of-experts model, 3 billion active, for agentic reasoning, tool use and coding over contexts of about a million tokens.

  • 1M context
  • Tools
  • JSON output

Reasons before it answers, and reasoning counts as output. Set max_tokens to at least 2,048, or leave it out for the default of 65,536.

Model card

Video models

Video generation at /v1/videos: submit a job, poll it, then download the MP4 within 90 days. You pay per second of video delivered; see Pricing.

MiniMax H3

Live

minimax-h3MiniMax

MiniMax's open-weights video model. It turns a prompt, or a first and last frame, into a 2K clip, and renders text and brands precisely.

  • 2K, 5 to 15 s
  • 16:9, 9:16, 1:1
  • Sound on or off

You pay for the seconds delivered; a generation that fails is not charged. A first frame, a last frame or both can guide the clip.

Examples

Pika H3 Flex

Live

pika-h3-flexPika

Pika's six-step video model, built on MiniMax H3. It turns a prompt and up to 12 image, video and audio references into a 3 to 15 second clip with sound, keeping characters, places and motion from the references.

  • 480p to 4K, 3 to 15 s
  • 11 aspect ratios
  • Up to 12 references

$0.02 a second at 480p, $0.05 at 768p, $0.065 at 1080p, $0.07 at 2K and $0.09 at 4K; reference video is billed per second at the same rate, and a generation that fails is not charged. A 10-second 768p clip with references is typically ready in 35 to 55 seconds; 1080p, 2K and 4K are made at 768p and upscaled once, which adds about 30 seconds, 50 seconds and 3 minutes. Pika's own API request works too. Powered by MiniMax H3, under the MiniMax H3 Community License, whose use restrictions apply.

Examples

Call any model

Switch models by changing one field.

The request is the same for every text model: set model to any text model ID above. Video models work as jobs instead; see Generate a video.

curl https://api.marscompute.ai/v1/chat/completions \
  -H "Authorization: Bearer $MARSCOMPUTE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4.1-flash",
    "messages": [
      {"role": "user", "content": "Say hello in Malay."}
    ]
  }'

Call a model in a few minutes.

Sign up, add credit and create one key for every model.