MiMo-V2.6-Pro Is Now Available on DeepInfra

From 19.0 to 71.9 on DeepSWE — that’s the jump Xiaomi reports between MiMo-V2.5-Pro and MiMo-V2.6-Pro, its new flagship, on real-world software engineering tasks. That puts it, on Xiaomi’s own published figures, alongside Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0) on a benchmark designed to test repository-level coding rather than toy problems. Built by Xiaomi’s MiMo team, it’s a sparse Mixture-of-Experts model with 1.02T total parameters and 42B activated per forward pass, designed for long-horizon agentic workloads across a 1-million-token context window. It handles text, image, video, and audio natively — no bolt-on adapters.

The training approach is what makes the architecture choices pay off. Xiaomi calls it “You Only RL Once”: a single RL run that mixes coding, general agent, visual, and cybersecurity tasks in the same batch rather than running separate per-domain phases — so strategies transfer across domains within one training run. That shows up in Xiaomi’s own reported numbers: AutomationBench goes from 16.0 to 53.1, CyberGym from 40.0 to 94.0, versus the previous version. It ships under an MIT license, which removes the licensing friction that typically complicates production deployments at this parameter scale. MiMo-V2.6-Pro is now available on DeepInfra.

What Makes This Model Different

MiMo-V2.6-Pro is a sparse Mixture-of-Experts model with 1.02T total parameters, but only 42B are activated per forward pass. The LLM backbone runs 70 layers (60 sliding-window attention / 10 global attention), with 384 routed experts and 8 activated per token — no shared experts. It’s natively omnimodal, processing text, image, video, and audio through dedicated encoders: a 681M-parameter MiMo ViT for vision and two audio encoders (an AudioTokenizer at 308M params and an audio patch encoder at 127M). The context window is 1,048,576 tokens (1M), which matters practically for long repository ingestion, multi-tool agent traces, and multi-session runs.

The training story is worth understanding. Xiaomi calls it “You Only RL Once”: a single mixed RL run across coding, general agent tasks, visual, and cybersecurity domains simultaneously — not separate per-domain fine-tuning passes. Tasks are mixed in the same batch, so the model can transfer reasoning strategies across domains. This is complemented by Groupwise Agentic Grading, a self-improvement loop that builds task-specific rubrics from contrasting rollouts and re-ranks passing trajectories to steer toward shorter, more token-efficient solutions. A speculative decoder (5-layer SWA MTP drafter) predicts 7 subsequent tokens per forward pass, improving throughput without changing output quality.

The improvement over MiMo-V2.5-Pro is substantial across the board, per Xiaomi’s own benchmark comparisons:

Those aren’t incremental gains — the cybersecurity and agentic terminal benchmarks in particular suggest a substantial capability shift rather than tuning on familiar evals.

Against frontier competitors, V2.6-Pro is competitive on agentic tasks while trailing on others. On Xiaomi’s published benchmark comparisons, it beats Claude Opus 5 (50.3) and GPT-5.6 Sol (45.8) on AutomationBench with a score of 53.1. On MiMo VisualCoding (72.3) it edges out Claude Opus 5 (70.0). It falls behind on ExploitBench (47.9 vs. 70–78.5 for Claude Opus 5 and GPT-5.6 Sol) and Terminal Bench 4.0 (34.9 vs. 39.9–49.0). The model’s strength is clearly in multi-step agentic workflows rather than single-turn reasoning benchmarks.

This release also introduces MiMo-V2.6-Pro-UltraSpeed — a serving mode built for substantially faster output than standard Pro. It’s not a separate model; it’s the same Pro checkpoint served differently, designed for real-time and latency-sensitive workflows. Pricing and provider availability for UltraSpeed vary by provider and are worth checking directly on the DeepInfra pricing page before committing to a workload. It runs at a higher cost per token but is available both via API and through MiMo Desktop.

Getting Started on DeepInfra

MiMo-V2.6-Pro is available now on DeepInfra. Pricing is $0.435 per million input tokens (cache miss), $0.0036 per million cached input tokens, and $0.87 per million output tokens.

If your workload involves long, repetitive context — agent traces, large codebases, multi-turn sessions — the prompt cache makes a real difference at that input price.

DeepInfra exposes the model through an OpenAI-compatible API with usage-based billing and no infrastructure to manage. It plugs in wherever you already use the OpenAI SDK. Private endpoint deployment is also available through the DeepInfra dashboard for workloads that need dedicated capacity. DeepInfra operates with a zero-retention policy on all inference requests and is SOC 2 and ISO 27001 certified.

Here’s a minimal example to get your first call running. Swap in your DeepInfra API key and you’re set:

Python

from openai import OpenAI
 
client = OpenAI(
    api_key="$DEEPINFRA_TOKEN",
    base_url="https://api.deepinfra.com/v1/openai",
)
 
response = client.chat.completions.create(
    model="XiaomiMiMo/MiMo-V2.6-Pro",
    messages=[{"role": "user", "content": "Outline a research plan for synthesizing a MOF optimized for PFAS adsorption."}],
    temperature=1.0,
    top_p=0.95,
)
print(response.choices[0].message.content)
copy

Shell

curl "https://api.deepinfra.com/v1/openai/chat/completions" \
    -H "Content-Type: application/json" \
    -H "Authorization: Bearer $DEEPINFRA_TOKEN" \
    -d '{
          "model": "XiaomiMiMo/MiMo-V2.6-Pro",
          "messages": [
            {
                "role": "user",
                "content": "Outline a research plan for synthesizing a MOF optimized for PFAS adsorption."
            }
          ],
          "temperature": 1.0,
          "top_p": 0.95
      }'
copy

JavaScript

import OpenAI from "openai";
 
const openai = new OpenAI({
  apiKey: "$DEEPINFRA_TOKEN",
  baseURL: "https://api.deepinfra.com/v1/openai",
});
 
const response = await openai.chat.completions.create({
  model: "XiaomiMiMo/MiMo-V2.6-Pro",
  messages: [{ role: "user", content: "Outline a research plan for synthesizing a MOF optimized for PFAS adsorption." }],
  temperature: 1.0,
  top_p: 0.95,
});
console.log(response.choices[0].message.content);
copy

Xiaomi’s own deployment guidance recommends temperature=1.0 and top_p=0.95 — worth sticking to those defaults, especially for agentic and reasoning workloads where the model benefits from the full output distribution. Head to the MiMo-V2.6-Pro page on DeepInfra to grab your API key and start building.

Conclusion

MiMo-V2.6-Pro is a serious agentic model — the benchmark jumps reflect genuine capability changes rather than eval overfitting, and the MIT license means you can ship it without legal back-and-forth. The combination of a 1M-token context window, native function calling, and training that explicitly optimizes for multi-step tool use makes it a reasonable fit for the class of problems that single-turn models still struggle with: long-running coding agents, automated security workflows, and complex multi-session tasks that require coherent reasoning across large contexts.

If you’re building anything in the agentic space, this is worth running through your eval suite — and worth comparing against other recent releases like DeepSeek V4 Flash and Qwen3.6 or the Kimi K3 vs. Claude Opus 4.8 vs. GPT-5.6 Sol lineup. Start at the MiMo-V2.6-Pro model page on DeepInfra, or browse the full model catalog for other options.