SN-X

Model · Flagship

SN-X Papers

Working papers across the five research axes. Published in arXiv, Nat. Phys., PNAS, and journals of record.

Try in playgroundRead the system card
Context window
1,048,576 tokens
Modalities
Text · Vision
Tool-use
JSON schema · streaming
Languages
18 source · 90 read

Capabilities

What Papers is good at.

  • Long-horizon reasoning

    Deliberative rollouts let M3 plan, commit, and recover across hundreds of tool calls without drifting off-task.

  • Native tool-use

    A first-class JSON tool-call channel with structured refusal, idempotency hints, and automatic schema recovery.

  • One-million-token context

    1M tokens of working memory with sparse mixture routing — cost-effective for long documents and codebases.

  • Code generation

    92.4 on HumanEval+. Native understanding of multi-file diffs and toolchains across 18 languages.

  • Vision

    Charts, diagrams, screenshots, and UI mockups — with grounded region citations and refusal on ambiguous inputs.

  • Streaming & function calls

    Server-sent token streaming with interleaved tool calls and traceable reasoning.

API

Drop-in OpenAI-compatible endpoint.

Same request shape as the OpenAI Chat Completions API. Switch yourbaseURLandmodeland you're running Papers.

curlbash
curl https://api.minimax.dev/v1/chat/completions \
  -H "Authorization: Bearer $MINIMAX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax-m3",
    "temperature": 0.7,
    "messages": [
      { "role": "system", "content": "You are a careful research assistant." },
      { "role": "user", "content": "Summarise the safety properties of M3." }
    ]
  }'
client.tsts
import OpenAI from "openai"

const client = new OpenAI({
  apiKey: process.env.MINIMAX_API_KEY,
  baseURL: "https://api.minimax.dev/v1",
})

const response = await client.chat.completions.create({
  model: "minimax-m3",
  temperature: 0.7,
  messages: [
    { role: "system", content: "You are a careful research assistant." },
    { role: "user", content: "Summarise the safety properties of M3." },
  ],
})

console.log(response.choices[0].message.content)

Pricing

Pay for what you use.

Token-based pricing across three model sizes. Volume discounts available above 10B tokens / month.

M3-nano

$0.05 / 1M input · $0.20 / 1M output

Sub-second latency for high-volume classification, extraction, and autocomplete workloads.

M3-mini

$0.30 / 1M input · $1.20 / 1M output

The everyday workhorse. Tool-use, 256k context, and most enterprise workflows at a sensible price.

minimax M3

Flagship

$2.50 / 1M input · $10.00 / 1M output

The full flagship. 1M-token context, deliberative rollouts, vision, and the strongest reasoning profile.

Availability

Where you can run it.

  • Hosted APIapi.minimax.dev — global, 99.9% SLA
  • RegionUS-East · US-West · EU-West · AP-South
  • Self-hostedAWS, GCP, Azure — via the minimax runtime image
  • On-deviceM3-nano on Apple Silicon, Snapdragon X, and Linux/ROCm

Changelog

Releases.

  1. M3

    minimax M3 — general availability

    Reasoning v2, vision, 1M context, structured tool calls, and an updated safety stack.

  2. M3-mini

    M3-mini — context doubled to 256k

    Same price, twice the working memory. Latency unchanged.

  3. M3-nano

    M3-nano — released

    A 3B sparse-mixture model purpose-built for high-volume, low-latency workloads.

  4. Preview

    Public preview

    Limited preview under the codename "M3-preview". Replaced by GA on 2026-01-15.

Try it now.

No account required for the playground — just type a prompt and watch the stream.

Open playground