On our own GPUs
MiniMax M3 API
229.9B MoE that puts frontier coding, a 1M window and native multimodality in one model.
Context
1M
Max output
128K
Serving precision
BF16
Licence
Open weights
MiniMax M3 was released with open weights in June 2026: a 229.9B-parameter mixture-of-experts with 256 experts and 9.8B active per token, extending context from 256K to 1 million tokens. MiniMax describes it as the first open-weight model to combine frontier-level coding, a million-token window and native multimodality across text, image and video input. Its new MSA sparse attention reports roughly a 9.7× prefill and 15.6× decode speedup at 1M context over the previous generation.
Open weights, deployed in facilities we build out and lease directly, operated by us. We resell nobody else’s account quota.
Specifications
- Model ID
- minimax-m3
- Published by
- MiniMax
- How it is served
- On our own GPUs
- Licence
- Open weights
- Total parameters
- 229.9B MoE
- Active parameters
- 9.8B / 256 experts
- Context
- 1M
- Max output
- 128K
- Serving precision
- BF16
- Input modalities
- Text · Image · Video
- Released
- 2026-06-01
- Price
- See pricing
What it is good at
- It reads video directly
- Text, image and video input are all base capabilities. Analysing screen recordings, surveillance clips or product demos does not need a separate frame-extraction-and-captioning pipeline in front.
- Long context was re-engineered, not just extended
- MSA sparse attention rebuilds attention for very long sequences — the vendor reports ~9.7× faster prefill and ~15.6× faster decode at 1M context. On long-document work that difference decides whether the experience is usable at all.
- Unusually balanced for an open model
- Coding, long context and multimodality normally force a trade-off across different models. M3 puts them in one weight set, so a multimodal agent needs one fewer model to operate.
Where it fits
- Multimodal agents that look before they act
- Automated testing and QA with screenshot or recording evidence
- Structured summaries of long videos and meetings
- Mixed coding-plus-vision pipelines, with fewer models to maintain
Three steps
Change one line of base_url, put the model ID below in the `model` field, and leave the rest of your code alone.
model: "minimax-m3"
curl https://api.nexaix.net/v1/chat/completions \
-H "Authorization: Bearer $NEXAIX_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "minimax-m3",
"stream": true,
"messages": [
{ "role": "user", "content": "Summarise this contract." }
]
}'import os
from openai import OpenAI
client = OpenAI(
# the only new line
base_url="https://api.nexaix.net/v1",
api_key=os.environ["NEXAIX_API_KEY"],
)
# everything below is your existing code
stream = client.chat.completions.create(
model="minimax-m3",
stream=True,
messages=[
{"role": "user", "content": "Summarise this contract."}
],
)
for chunk in stream:
print(chunk.choices[0].delta.content or "", end="")import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://api.nexaix.net/v1", // the only new line
apiKey: process.env.NEXAIX_API_KEY,
});
// everything below is your existing code
const stream = await client.chat.completions.create({
model: "minimax-m3",
stream: true,
messages: [
{ role: "user", content: "Summarise this contract." }
],
});
for await (const chunk of stream) {
process.stdout.write(chunk.choices[0]?.delta?.content ?? "");
}About this model
Is the context 512K or 1M?
The vendor states the window extends from 256K to 1M tokens with a guaranteed floor of 512K. The number published on this page is what we actually serve — treat that as authoritative.
How is video input billed?
Video is billed as the image tokens of its sampled frames, so frame rate and resolution drive cost directly. The docs list recommended parameters; estimate on a short clip before scaling up.
M3 or Kimi K3 for multimodal work?
Both are natively multimodal. K3 is far larger and stronger on hard reasoning; M3 activates only 9.8B, so unit cost is lower and long-context throughput is higher. For vision-heavy batch work that is not reasoning-extreme, M3 is usually the better economics.
Also worth a look
- On our own GPUsKimi K32.8T mixture-of-experts, a 1M-token window, and vision built into the base.kimi-k3
- On our own GPUsGLM-5.2750B sparse MoE, a 1M window, MIT licensed — tuned hard for software engineering.glm-5.2
- On our own GPUsDeepSeek V4 Pro1.6 trillion parameters, thinking and non-thinking modes, MIT licensed.deepseek-v4-pro
Run your own eval on it
Take the trial credit, run the evals you already have against this model, and then decide whether it gets your production traffic.
