Xiaomi's standard V2.5 model combines native text, image, video, and audio understanding with a 1-million-token context window, tool use, and sparse Mixture-of-Experts efficiency. It is the better fit when multimodal input and cost per completed task matter. The Pro variant is larger and stronger on demanding coding and long-horizon agent work, but it is text-input focused and costs roughly three times as much per token at Xiaomi's published rates.
The practical decision is simple: choose the standard model for multimodal agents, document and media analysis, and high-volume automation; choose Pro when difficult software engineering or sustained autonomous execution is the bottleneck.
Key Takeaways
The standard model uses 310B total parameters while activating 15B per token.
It accepts text, images, video, and audio, then returns text.
Its API supports 1M context, up to 128K output, tool calls, web search, streaming, structured output, and context caching.
Reviewed for clarity, source attribution and current API terminology.
Publisher-reported results include 65.8 on Terminal-Bench 2.0 and 56.1 on SWE-Bench Pro.
Xiaomi’s current MiMo-V2.5-Pro model card lists 1.02T total and 42B active parameters. The publisher-listed SWE-Bench Pro scores are 56.1 for MiMo-V2.5 and 57.2 for Pro; these are dated, publisher-reported comparisons, not a guarantee for a specific coding workflow.
At current Xiaomi rates, standard input/output costs $0.14/$0.28 per million tokens; Pro costs $0.435/$0.87.
Both APIs in CometAPI are priced 20% below the official uncached price pairs. Information checked: September 28, 2026. Prices, benchmarks, limits, availability, and request formats can change. Xiaomi has announced that its official mimo-v2.5 and mimo-v2.5-pro API model names will be deprecated at 10:00 Beijing time on October 21, 2026, without automatic replacement. Plan migration and confirm the live route before deployment.
API lifecycle note: the official Xiaomi platform plans to retire both V2.5 model IDs on October 21, 2026. This notice concerns Xiaomi’s API platform; check any third-party gateway separately. Xiaomi model deprecation notice
What the Standard Model Is
MiMo-V2.5 is Xiaomi MiMo's efficiency-oriented foundation model for multimodal perception and agentic work. It provides native text, image, video, and audio input within one architecture and produces text output.
Xiaomi’s model release log dates the MiMo-V2.5 series public beta to April 23, 2026. The weights were subsequently released under the MIT license. MiMo-V2.5 has 310B parameters in total, but activates only about 15B for each token; it uses a large pool of specialists without running all of them at once.
MiMo-V2.5-Pro targets a different workload. It is a trillion-parameter, text-input model designed for the hardest coding, terminal, and long-running agent tasks. The two variants share a 1M maximum context but differ sharply in modality, active compute, and price.
MiMo-V2.5 Specifications and Features
Official architecture details
Value
Why it matters
Architecture
Sparse MoE
Large expert capacity with selective activation
Total / active parameters
310B / 15B
Active compute is far below total capacity
Transformer layers
48: 1 dense + 47 MoE
Most layers use routed experts
Routed experts
256; 8 active per token
Specialization without dense 310B execution
Attention
39 SWA + 9 global layers
Balances local efficiency and long-range flow
SWA window
128 tokens
Reduces long-context cache pressure
Context / maximum output
1M / 128K tokens
Supports long documents and extended traces
Input / output
Text, image, video, audio / text
Native perception across four input types
Vision encoder
729M ViT; 28 layers
Image and video understanding
Audio encoder
261M transformer; 24 layers
Native audio understanding
Multi-Token Prediction
3 modules; about 329M parameters
Supports speculative decoding efficiency
Training scale
The operating envelope includes 1M context and 128K output, plus published limits of 100 requests per minute and 10 million tokens per minute. Provider-level limits may differ by account and integration route.
Sparse Experts: Total Capacity Is Not Active Compute
The central efficiency fact is the 310B-total, 15B-active design. The router selects eight of 256 experts for each token. That gives the model access to a much larger parameter pool than a conventional 15B dense model without executing all 310B parameters every time.
This does not make self-hosting trivial. The complete weights still have to be stored and distributed, so memory capacity, interconnect bandwidth, expert routing, and serving software remain material constraints.
Hybrid Attention for Long Context
The backbone interleaves sliding-window and global attention at a 5:1 SWA-to-global ratio. Thirty-nine local layers control cache growth, while nine global layers preserve long-distance information flow. Xiaomi reports nearly a sixfold KV-cache reduction versus an all-global design.
A 1M-token maximum is capacity, not guaranteed accuracy. Retrieval quality, latency, tool-state growth, and attention over distant evidence still require workload-specific evaluation.
Native Encoders and Agent Features
A 729M-parameter vision transformer handles image and video inputs; a 261M-parameter audio transformer handles audio. The projectors map both streams into the language backbone, letting one agent combine a screenshot, a video segment, an audio track, text instructions, and tool results.
Three Multi-Token Prediction modules support speculative decoding and reinforcement-learning efficiency. The model was trained on about 48T tokens, with context progressively extended to 1M during post-training.
The open weights use an MIT license. Commercial deployment is permitted, but infrastructure cost and third-party dependencies still need separate review.
MiMo-V2.5 Benchmarks: Coding, Agents, and Multimodal Results
Benchmark note:
the figures below are publisher-reported. They are directional evidence, not a guarantee for a specific repository, media format, tool stack, latency target, or safety policy.
Coding and Agent Results
Official Xiaomi MiMo coding and agent benchmark chart.
Official coding benchmark
Reported score
Decision signal
MiMo Coding Bench
71.8
Broad coding-agent capability
Claw-Eval Text
62.3
General text-agent completion
Terminal-Bench 2.0
65.8
Interactive terminal execution
SWE-Bench Pro
56.1
Real-world software engineering
Claw-Eval Multi-Turn
63.2
Longer multi-step agent work
ResearchClawBench
16.91
Autonomous research workflows
Result: the strongest case is practical tool use rather than isolated code completion. A 65.8 Terminal-Bench result supports terminal-oriented agents, while 56.1 on SWE-Bench Pro suggests useful repository-level ability. The much lower research score is a reminder to test evidence gathering and citation behavior separately.
Multimodal Results
Official Xiaomi MiMo image, video, and multimodal-agent benchmark chart.
Official multimodal benchmark
Reported score
Capability tested
CharXiv RQ
81.0
Chart and document reasoning
MMMU-Pro
77.9
Expert-level multimodal reasoning
HR-Bench 4K
88.5
High-resolution image understanding
OmniDocBench
87.2
Document understanding
Claw-Eval Multimodal
23.8
Multimodal agent completion
Video-MME
87.7
Video understanding
DailyOmni
83.5
Everyday audiovisual reasoning
VideoHolmes
64.0
Temporal video reasoning
Result: document, high-resolution image, and video understanding are the clearest strengths. The 23.8 multimodal-agent score is much lower than the perception scores, so a system that must both understand media and reliably operate tools should be evaluated end to end.
MiMo-V2.5 vs MiMo-V2.5-Pro: Multi-Dimensional Comparison
Official architecture comparison
MiMo-V2.5
MiMo-V2.5-Pro
Official Pro specifications
Official Pro benchmarks
Practical implication
Total parameters
310B
1.02T
Pro has over 3× the total capacity
Activated parameters
15B
42B
Pro uses about 2.8× more active parameters
Layers / routed experts
48 / 256
70 / 384
Pro is a larger serving target
Model-card context ceiling
1M
1M
Both list up to 1M tokens; test retrieval and latency at your workload size
Official API maximum output
128K
128K
Both current Xiaomi API model pages list 128K; provider-specific limits may differ
Native input
Text, image, video, audio
Text
MiMo-V2.5 is the documented multimodal choice; verify the exact Pro endpoint before sending media
SWE-Bench Pro
56.1
57.2
Pro leads by 1.1 points
Terminal-Bench 2.0
65.8
68.4
Pro leads by 2.6 points
Primary fit
Efficient multimodal agents
Complex coding and long-horizon agents
Choose by tested workload; the model-level margins do not prove a general quality lead
Comparison result: Pro's benchmark advantage is modest on the two shared coding-agent tests, while the standard variant adds native image, video, and audio input and uses far fewer active parameters. Pro is justified when small gains in hard task completion are more valuable than multimodal input and lower unit cost.
How Much Do MiMo-V2.5 and MiMo-V2.5-Pro Cost?
Price basis: May 27, 2026
Official standard API
Official Pro API
MiMo-V2.5 API in CometAPI
MiMo-V2.5-Pro API in CometAPI
Input, cache miss / MTok
$0.14
$0.435
$0.112
$0.348
Output / MTok
$0.28
$0.87
$0.224
$0.696
Input, cache hit / MTok
$0.0028
$0.0036
Check live billing
Check live billing
Discount vs official uncached rate
Baseline
Baseline
20%
20%
At official rates, Pro costs about 3.1× as much as standard for both uncached input and output. The provider prices shown above reduce each uncached pair by 20%, but the relative gap between the variants remains essentially unchanged.
Price per token is not total cost. Tool retries, context size, output length, latency, failed task recovery, and human review determine cost per completed task. Run a representative workload sample before selecting a default production model.
What Is MiMo-V2.5 Best For?
Multimodal agents: combine screenshots, documents, video, audio, instructions, and tools in one workflow.
Long-document analysis: process repositories, contracts, research archives, logs, and support histories.
Media understanding: summarize videos, extract events, interpret charts, and answer audiovisual questions.
Coding and terminal agents: inspect files, run commands, modify code, and iterate on test results.
High-volume automation: use sparse activation and lower token pricing for repeated production tasks.
Limitations and Risks
Only 15B parameters are active per token, but self-hosting still requires the full 310B-weight system.
Multimodal output is text; image, speech, and video generation require other models.
A 1M context ceiling does not guarantee uniform retrieval accuracy across the full window.
Publisher benchmarks may not transfer to custom tools, repositories, prompts, or safety constraints.
Provider modality exposure can differ from the underlying open-weight model's full capability.
Production gate:
validate accuracy, tool-call completion, latency, token consumption, modality handling, and fallback behavior on representative tasks before committing traffic.
API Example
The following Python example uses an OpenAI-compatible CometAPI endpoint and the standard model ID. Confirm the current endpoint and supported request schema before deployment.
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["COMETAPI_KEY"],
base_url="https://api.cometapi.com/v1",
max_retries=0,
)
response = client.chat.completions.create(
model="mimo-v2.5",
max_tokens=256,
messages=[
{
"role": "user",
"content": "Summarize the main findings in this technical report.",
}
],
)
print(response.choices[0].message.content)
Decision Guide
Workload signal
Start with
Switch when
Images, video, or audio are first-class inputs
Standard
Do not switch unless multimodal preprocessing is acceptable
Large document or mixed-media volume
Standard
Pro only if reasoning failures dominate cost
Difficult repository engineering
Test both
Choose Pro if its completion gain offsets 3.1× unit cost
Long autonomous terminal trajectories
Pro
Return to standard if gains are not measurable
High-volume routine automation
Standard
Escalate only failed or high-value tasks
Recommended routing:
use standard as the default for multimodal and high-volume work, then route only difficult text-first coding or long-horizon tasks to Pro. This preserves capability while controlling total cost.
Conclusion
The standard model is not merely a smaller Pro. It is a distinct optimization point: native multimodal input, 1M context, strong tool-oriented benchmarks, and 15B active parameters at a much lower token price.
Pro is the specialist option for difficult software engineering and sustained autonomous execution. Its extra capacity produces measurable but not universal gains, so the strongest deployment pattern is workload-based routing rather than selecting one variant for every request.
FAQ
Is the standard model open source?
Its weights are released under the MIT license, allowing commercial use, modification, fine-tuning, and redistribution subject to the license terms.
How many parameters does it use?
The sparse MoE contains 310B total parameters and activates 15B for each token. Active parameters describe per-token computation, not the storage size of the complete model.
Does it support images, video, and audio?
Yes. The standard variant accepts text, images, video, and audio and returns text. The Pro variant's official specification lists text input.
What is the maximum context window?
Both model cards specify a context window of up to 1M tokens. Xiaomi’s current API pages list a 128K maximum output for MiMo-V2.5 and MiMo-V2.5-Pro. Actual usable limits can vary by endpoint, provider, account, and request format; verify the deployed route before relying on those ceilings.
Pro reports higher results on shared coding and terminal benchmarks, but the margin is modest. Test both on the target repository and choose based on task completion, latency, and total cost.
Which variant is better for multimodal agents?
The standard variant is the natural choice because it natively accepts image, video, and audio input. Pro is text-input focused.
How should a production team choose?
Start with standard, measure failures, and route only the difficult text-first tasks that benefit from Pro. Compare cost per completed task rather than price per token alone.