AI Tools

Xiaomi MiMo V2.6 pricing: what do Pro, Flash and UltraSpeed cost per million tokens?

Published 2026-09-22 · Researched and price-checked on official sources

Xiaomi MiMo V2.6 pricing starts at $0.14 per million input tokens for Flash and $0.435 for Pro, identical to the V2.5 series. Cache-hit input drops to $0.0028 and $0.0036. Pro-UltraSpeed lists at $4.35 input and $8.70 output for up to 20x the output speed, and every tier keeps a 1M-token context window.

Key facts

What matters

MiMo-V2.6-Flash costs $0.14 per million input tokens and $0.28 output; Pro costs $0.435 and $0.87, the same rates Xiaomi charged for the V2.5 series.
Cache-hit input is the biggest cost lever: $0.0036 per million on Pro against $0.435 on a cache miss, about 120x cheaper for prompts you keep stable.
The open weights are real but large: Pro is a 1.02T-parameter sparse MoE with 42B activated, Flash is 309B total with 15B activated, and both model cards declare the MIT license.
Only the 9.4B Distill-Qwen-9B is a realistic local download, and its Hugging Face card lists no license field, so commercial reuse needs checking first.
Pro-UltraSpeed costs exactly 10x Pro per token for Xiaomi's claimed 20x output speed, and a Batch API covers Flash and Pro for non-realtime jobs.

What Xiaomi MiMo V2.6 actually is

MiMo-V2.6 is Xiaomi's open-weight model series, published on September 22, 2026, and sold through three API tiers: MiMo-V2.6-Pro, MiMo-V2.6-Flash and MiMo-V2.6-Pro-UltraSpeed.

The series is natively omnimodal. Xiaomi's launch post describes text, image, video and audio input in a single model with a 1M-token context window, and the model cards add the architecture detail: a hybrid sliding-window attention backbone, sparse mixture-of-experts feed-forward layers and an MTP block for speculative decoding.

Xiaomi says Pro scores 46.32 on the Artificial Analysis Intelligence Index v4.3 (September 2026), which it presents as the best score for an open-source model, ahead of Kimi K3 and Qwen3.8 Max. That comparison is Xiaomi's own reading of a third-party index, so treat the phrase "strongest open-source model" as a vendor claim rather than a settled fact.

Distribution is wider than the API. MiMo Studio, MiMo Code, MiMo Desktop and the Xiaomi MiMo Open Platform all carry the V2.6 models, and OpenRouter lists all three tiers under the xiaomi provider.

MiMo V2.6 API pricing, tier by tier

Every rate below is USD per million tokens, taken from Xiaomi's launch post and cross-checked against OpenRouter's per-model listing. The two agree on all three tiers, which matters because Xiaomi's own RMB pricing page is JavaScript-rendered and returns nothing to a plain fetch.

Two columns decide your bill. Cache-miss input is fresh text you pay full price for. Cache-hit input applies when the same prefix repeats, which is the normal pattern for coding agents that resend a long system prompt and tool trace on every turn. Xiaomi's post notes that cache writes are free for a limited time, so the caching discount is promotional and can change.

Beyond pay-as-you-go, a Batch API covers mimo-v2.6-flash and mimo-v2.6-pro for large non-realtime jobs, and a Token Plan subscription handles predictable high-volume usage for individuals and teams.

Pro, Flash or UltraSpeed: picking a tier

Flash for high-frequency calls, bulk extraction and classification. At $0.28 per million output tokens, a caching mistake costs you cents.
Pro for long-horizon agent runs, repository-wide refactors and research tasks, where a wrong answer costs more than the tokens that produced it.
Pro-UltraSpeed only when latency is the product: voice interfaces, live editing, anything where a person waits on the token stream.

UltraSpeed is a mode of Pro rather than a separate checkpoint, so you are buying speed at a 10x rate, not extra capability. If your workload can wait, the Batch API on Flash or Pro is the cheaper path. Pick UltraSpeed when a slow answer breaks the interaction, not because the label sounds faster.

What a realistic agent workload costs

Rates only become a decision when you run them through a workload. Take 10 million input tokens and 2 million output tokens a month, the shape of a busy coding agent, and price it with Xiaomi's listed rates: Flash comes to $1.96, Pro to $6.09 and Pro-UltraSpeed to $60.90. Those totals are arithmetic on published rates, not vendor benchmarks.

Caching changes the picture more than tier choice does. If 80% of that input arrives as cache hits, Flash falls to about $0.86 and Pro to about $2.64, because the repeated prefix stops being billed at the miss rate. That is the practical argument for keeping system prompts and tool definitions byte-identical across calls, and the reason the cache-hit column is worth more attention than the headline input price.

Compare that spread with the price gap between tiers: moving from Flash to Pro multiplies your bill roughly 3x, while fixing a cache-unsafe prompt can cut it by more than half. Optimize prompt stability before you shop for a cheaper model.

Are the open weights usable? For most teams, no

Xiaomi open-sourced the checkpoints, and the honest reading is that they are for clusters, not laptops. The MiMo-V2.6-Pro-RL card describes a sparse MoE with 1.02T total and 42B activated parameters, 384 routed experts and a 1M-token context. The repository itself carries roughly 524B parameters of safetensors spread over more than a hundred expert-parallel shards plus a DFlash draft model, and the card tags the release as fp8, which explains why the stored count is about half the stated total.

Flash is smaller but not small: the Flash-RL card lists 309B total and 15B activated parameters with 256 routed experts. Both cards declare the MIT license, though neither repository ships a LICENSE file, so the license lives in the model card rather than the repo root.

The one variant you can actually run on a workstation is MiMo-V2.6-Distill-Qwen-9B: 9.4B parameters, built on Qwen3.5-9B, image-text-to-text, with GGUF and MLX quantizations already published by third parties. Its Hugging Face card lists no license field, and the repository has no LICENSE file, so clear the terms before shipping it inside a product. A free download is not the same as a commercially clear one.

What to watch before you commit

Free cache writes are temporary. Xiaomi's post says cache writes are free for a limited time. Once that ends, cheap cache reads get more expensive to set up.
Price parity with V2.5 is a promise, not a contract. The V2.6 rate card matches V2.5 exactly, and the next generation is the obvious place for a change.
Vendor index claims move. The 46.32 score is Xiaomi citing Artificial Analysis v4.3 for September 2026. Check the live leaderboard before you quote it in a procurement note.
1M context is where costs concentrate. OpenRouter lists 1,048,576 tokens of context for the V2.6 models. Long contexts and long agent traces are exactly the workload that bills the most cache-miss input.

None of that argues against the series. It argues for measuring your own cache-hit rate for a week before you move a production workload onto it.

At a glance

Model	Input (cache hit)	Input (cache miss)	Output	Context (OpenRouter)
MiMo-V2.6-Flash	$0.0028	$0.14	$0.28	1,048,576 tokens
MiMo-V2.6-Pro	$0.0036	$0.435	$0.87	1,048,576 tokens
MiMo-V2.6-Pro-UltraSpeed	$0.036	$4.35	$8.70	1,048,576 tokens

Xiaomi MiMo V2.6 API rates, USD per million tokens (launch post and OpenRouter, checked September 22, 2026).

FAQ

Is Xiaomi MiMo V2.6 free to use?

The API is paid: Flash starts at $0.14 per million input tokens and Pro at $0.435. The weights are free to download, and the Pro and Flash model cards declare MIT, but serving a 1.02T or 309B parameter checkpoint needs cluster hardware, so "free" applies to the download rather than to running it.

Did MiMo V2.6 change the API prices from V2.5?

No. Xiaomi's launch post states that V2.6 keeps V2.5 pricing, and OpenRouter lists MiMo-V2.5-Pro at the same $0.435 input and $0.87 output as V2.6-Pro. The change is in capability and in the new UltraSpeed tier, not in the rate card.

Can I run MiMo V2.6 locally?

Only the 9.4B Distill-Qwen-9B is realistic on a single machine, since it is built on Qwen3.5-9B and already has community GGUF and MLX quantizations. Its model card lists no license, so check the terms before commercial use. Pro and Flash are 1.02T and 309B parameter MoE models published as fp8 shards for multi-GPU serving.