DeepInfra raises $107M Series B to scale the inference cloud — read the announcement

Updated November 2026.
MiMo-V2.6-Pro is the kind of model that gets a pricing guide worth reading: it combines open weights, a 1M-token context window, and pricing that undercuts the typical large open-weight reasoning model. At list rates of $0.43 per million input tokens and $0.87 per million output tokens ($0.387 / $0.783 on DeepInfra’s current promotion; pricing as of 2026-11-17), it comes in below the median pricing Artificial Analysis reports for comparable models, while posting an Artificial Analysis Intelligence Index score of 30. For teams trying to balance capability against infrastructure and API spend, that makes it more than just another big checkpoint.
MiMo-V2.6-Pro is Xiaomi’s flagship model in the MiMo-V2.6 family, released on September 21, 2026. It is a sparse Mixture-of-Experts model with roughly 1 trillion total parameters and 42 billion active parameters per token, built for reasoning-heavy workloads and natively designed for text, image, video, and audio input with text output. The model ships under an MIT license, its weights are available on Hugging Face as XiaomiMiMo/MiMo-V2.6-Pro-RL, and hosted access is available through providers including DeepInfra and OpenRouter — see the full DeepInfra model catalog for everything else hosted alongside it. For developers, that combination matters: you can evaluate it as an API, self-host it, or do both.
What makes MiMo-V2.6-Pro stand out is not just scale, but the shape of the tradeoff. It pairs a very large context window — 1,048,576 to 1,050,000 tokens depending on the source — with competitive inference pricing, a 99% cache discount, and built-in support for function calling and structured outputs. On Xiaomi’s own published benchmark summaries, it scores 71.9 on DeepSWE v1.1, 53.1 on AutomationBench v1.0.6, and 34.9 on Terminal Bench 4.0, putting it in the same conversation as models like Claude Opus 5, GPT-5.6 Sol, and Claude Fable 5 on several agent and coding tasks. The tradeoff is speed: Artificial Analysis rates it below average on output throughput and time-to-first-token versus similar open-weight models, so this is less about raw responsiveness and more about capability per dollar.
If you are evaluating MiMo-V2.6-Pro for production, the real question is not whether it is interesting. It is whether its mix of long context, multimodal input, open-weight licensing, and aggressive token pricing fits your workload better than faster or more locked-down alternatives. That is especially relevant now that there is an actual provider market around the model, with OpenRouter listing five providers and DeepInfra exposing a managed endpoint with JSON mode, function calling, multimodal support, and private deployment options.
MiMo-V2.6-Pro sits in a useful middle ground: cheaper than many large-model alternatives, open-weight, and available through multiple hosted routes. If you need long-context reasoning, agent workflows, or multimodal input without giving up the option to self-host later, it is a strong candidate; if you care most about the absolute lowest price, GMICloud via OpenRouter is cheapest today, while DeepInfra is the cleaner fit for teams that want a managed API with a 0% error rate on Artificial Analysis, developer-friendly features and deployment flexibility.
| Best For | Provider Recommendation | Why |
|---|---|---|
| Easiest onboarding / fastest time-to-first-call | DeepInfra | DeepInfra exposes the model directly as XiaomiMiMo/MiMo-V2.6-Pro and includes managed API access with JSON mode, function calling, multimodal support, zero data retention, and private endpoint deployment. |
| RAG, document-heavy, or long-context production workloads | DeepInfra | The hosted model supports the full 1,048,576-token context window and low cached-input pricing at $0.00324/M tokens, which fits repetitive retrieval and document workloads well. |
| Lowest price / cost-sensitive workloads | GMICloud via OpenRouter | OpenRouter lists GMICloud at the lowest rate for this model: $0.32625/M input, $0.6525/M output and $0.003/M cache read (as of 2026-11-17). |
| Proprietary or managed model access | OpenRouter | OpenRouter gives managed access across five providers with routing modes for price, speed, and failover, which is useful if you want API convenience without committing to a single backend. |
| Teams comparing against closed-model alternatives | MiMo-V2.6-Pro on DeepInfra or OpenRouter | Xiaomi’s published benchmark summaries show it staying competitive with Claude Opus 5, GPT-5.6 Sol, and Claude Fable 5 on several agent and coding evaluations while remaining open-weight and MIT-licensed. |
MiMo-V2.6-Pro is priced the standard API way: you pay for tokens in, tokens out, and sometimes cached tokens when the provider can reuse prompt state. With a 1M-token context window, token math stops being academic pretty quickly.
| Token Type | What It Is | Why It Matters |
|---|---|---|
| Input tokens | Everything you send to the model: system prompt, user message, tool schemas, retrieved context, conversation history, and multimodal content after tokenization | This is the part that grows quietly. RAG chunks, long chats, and large tool definitions can dominate cost before the model writes a single token back. |
| Output tokens | Everything the model generates: answers, reasoning-heavy responses, JSON, tool calls, and long code blocks | MiMo-V2.6-Pro charges 2x more for output than input at list price, so verbose responses can move cost faster than expected. |
| Cached input tokens | Previously seen input tokens the provider can reuse at a discounted rate — see how prompt caching works | This is where MiMo-V2.6-Pro gets unusually attractive for repeated long prompts. If your provider supports cache reads well, the economics improve a lot. |
| Context window tokens | The total active prompt budget available in a request, up to about 1,048,576 to 1,050,000 tokens depending on provider listing | Big context is useful, but it also makes it easier to accidentally ship giant prompts into production and normalize a bad cost pattern. |
For MiMo-V2.6-Pro, the headline prices are:
DeepInfra lists discounted pricing at:
OpenRouter’s lowest-priced route today (GMICloud) lists:
A practical way to think about it:
MiMo-V2.6-Pro also trends somewhat verbose in Artificial Analysis testing. That does not mean it is wasteful by default. It does mean you should expect output tokens to add up faster on reasoning-heavy or tool-using workflows than they would on a terse classification model.
The annoying part of model pricing is that “same model” does not always mean “same bill shape.” MiMo-V2.6-Pro is a good example: token rates are close across providers, but the best choice changes depending on whether you care most about raw price, cache economics, routing, or operational simplicity.
| Provider | Input | Output | Cached Input / Cache Read | Advantages | Disadvantages |
|---|---|---|---|---|---|
| DeepInfra (direct) | $0.387/M | $0.783/M | $0.00324/M | 0% error rate on Artificial Analysis, and a promotional price below Xiaomi’s $0.435 / $0.87 list rate. Among the lowest cache rates in the listed set. Good fit for long-lived prompts, RAG, and repeated agent scaffolding. Direct model access is simple, so billing is easier to reason about. | Not the absolute lowest input/output price if you optimize purely for cheapest routed access. |
| OpenRouter via GMICloud | $0.32625/M | $0.6525/M | $0.003/M | Lowest listed input, output and cache pricing for this model on OpenRouter (as of 2026-11-17). | Available through OpenRouter’s routing layer; check quality and availability on your own workload before committing. |
| OpenRouter via Xiaomi | $0.435/M | $0.87/M | $0.0036/M | First-party route from the model’s developer. | List-price route: above DeepInfra’s promotional rate on every token type. |
| OpenRouter via DeepInfra (discounted) | $0.387/M | $0.783/M | $0.00324/M | Same pricing as going direct to DeepInfra, including the $0.00324/M cache rate. Useful if you want OpenRouter’s routing layer or consolidated billing while still getting DeepInfra’s discounted rate. | No cheaper than DeepInfra direct on any token category; you add OpenRouter’s layer for routing convenience. |
| OpenRouter via StreamLake | $0.435/M | $0.87/M | $0.0036/M | Another list-price route, useful if you need this specific backend for routing or availability reasons. | No token-price advantage over Xiaomi’s list price; above DeepInfra’s promotional rate. |
| OpenRouter via PrimaLabs | $4.00/M | $8.00/M | $0.036/M | Fastest output speed Artificial Analysis measures for this model (about 256 tokens/second). | Roughly 10× the price of the other routes. Hard to justify for cost-sensitive workloads. |
A few grounded rules for MiMo-V2.6-Pro specifically:
DeepInfra pairs aggressive token pricing with infrastructure built for inference efficiency. DeepInfra owns its GPU infrastructure, which helps it deliver better performance-per-dollar and keep pricing tight. Full current rates are on the DeepInfra pricing page. If you want a managed endpoint without paying a premium just for convenience, this is a provider worth evaluating first.
| Listing | Best Use Case | Context Window | Input / 1M | Output / 1M |
|---|---|---|---|---|
| MiMo-V2.6-Pro (discounted) | Long-context reasoning, multimodal agents, production API workloads | 1,048,576 | $0.387 | $0.783 |
| MiMo-V2.6-Pro-RL (discounted) | Flagship RL checkpoint for advanced reasoning and tool use | 1,048,576 | $0.387 | $0.783 |
| XiaomiMiMo/MiMo-V2.6-Pro (discounted) | Direct DeepInfra hosted endpoint for developers | 1,048,576 | $0.387 | $0.783 |
| MiMo-V2.6-Pro (standard list price) | Teams comparing discounted vs. non-discounted DeepInfra billing | 1,048,576 | $0.43 | $0.87 |
| MiMo-V2.6-Flash (lighter sibling; $0.003/M cached input) | Routine agent steps, extraction and high-volume calls where Pro’s extra capability isn’t needed | 1,048,576 | $0.14 | $0.28 |
On DeepInfra, MiMo-V2.6-Pro is $0.387/M input and $0.783/M output at the current discounted rate — below the model’s standard $0.43/M input and $0.87/M output list pricing (pricing as of 2026-11-17). If you are running sustained volume, going direct to DeepInfra instead of paying a routed markup is the cleaner cost move. The cached-input rate is also just $0.00324/M, which is attractive for repeat-prompt and long-context workloads.
For teams planning real usage rather than one-off tests, DeepInfra is the kind of provider that gets more compelling as volume grows. If your workload is large, repetitive, or cost-sensitive, its pricing and deployment flexibility are worth factoring in.
The case for DeepInfra is not that it is always the absolute cheapest on every token category. It is that it stays cheap while also giving developers direct model access, JSON mode, function calling, multimodal support, zero data retention, and private deployment options. For production teams, that combination often matters more than squeezing out the very last fraction of a cent through a routed marketplace.
Below are practical scenarios where MiMo-V2.6-Pro on DeepInfra (at its current discounted rate of $0.387/M input, $0.783/M output, $0.00324/M cached input) makes a strong case. All figures below are recalculated directly from those rates, and each comparison uses DeepInfra’s list price ($0.43/M input, $0.87/M output, $0.0036/M cached).
You are building a document assistant that ingests large internal docs, keeps a stable system prompt and tool schema, and reuses a lot of prompt structure across requests. This is exactly the kind of workload where DeepInfra’s $0.00324/M cached input rate stands out.
| Volume | Provider | Input Tokens | Output Tokens | Monthly Cost |
|---|---|---|---|---|
| 100M cached + 20M fresh input + 10M output / month | DeepInfra | 120M total | 10M | $15.89 |
Cost breakdown:
The same workload at DeepInfra’s list price would cost $17.66 instead of $15.89 (20M × $0.43 + 100M × $0.0036 + 10M × $0.87).
This is the classic developer workflow: repo context, tool definitions, intermediate reasoning, and structured outputs for an agent loop. MiMo-V2.6-Pro is attractive here because it pairs long context with strong benchmark results on coding and agent tasks, while DeepInfra exposes the model with the features developers actually need.
| Volume | Provider | Input Tokens | Output Tokens | Monthly Cost |
|---|---|---|---|---|
| 50M input + 25M output / month | DeepInfra | 50M | 25M | $38.93 |
Cost breakdown:
The same workload at DeepInfra’s list price would cost $43.25 (50M × $0.43 + 25M × $0.87), versus $38.93 at the discounted rate.
If your developers are processing screenshots, screen recordings, logs, and user-submitted audio alongside text, MiMo-V2.6-Pro’s native multimodal input support becomes more than a spec-sheet feature. DeepInfra is appealing here because it gives you managed multimodal access without forcing you into a proprietary stack.
| Volume | Provider | Input Tokens | Output Tokens | Monthly Cost |
|---|---|---|---|---|
| 30M input + 12M output / month | DeepInfra | 30M | 12M | $21.01 |
Cost breakdown:
The same workload at DeepInfra’s list price would cost $23.34 instead of $21.01.
This is where DeepInfra gets especially compelling: lots of repeated scaffolding, lots of tool definitions, and enough monthly volume that small pricing differences become visible in the bill.
| Volume | Provider | Input Tokens | Output Tokens | Monthly Cost |
|---|---|---|---|---|
| 500M cached + 100M fresh input + 50M output / month | DeepInfra | 600M total | 50M | $79.47 |
Cost breakdown:
The same workload at DeepInfra’s list price would cost $88.30 instead of $79.47 (100M × $0.43 + 500M × $0.0036 + 50M × $0.87).
Some teams start with hosted inference, then move specific workloads to private endpoints once governance, customer isolation, or predictable throughput starts to matter more. DeepInfra is strong here because it supports both the managed API path and private endpoint deployment, so you do not have to switch model providers just because your deployment model changes.
| Volume | Provider | Input Tokens | Output Tokens | Monthly Cost |
|---|---|---|---|---|
| 200M input + 80M output / month | DeepInfra | 200M | 80M | $140.04 |
Cost breakdown:
The same workload at DeepInfra’s list price would cost $155.60 instead of $140.04.
Not every step in an agent needs the flagship. Xiaomi’s lighter MiMo-V2.6-Flash (309B total / 15B active parameters) shares Pro’s 1M-token context, multimodal input and RL training recipe, and on DeepInfra it costs $0.14/M input, $0.28/M output and $0.003/M cached input (pricing as of 2026-11-17). That is about 64% below Pro’s discounted rate on both input and output. It’s the same price as Xiaomi’s own API, with no Flex or Priority tier, and DeepInfra’s Flash endpoint showed a 0.0% error rate on Artificial Analysis.
| Volume | Model | Input Tokens | Output Tokens | Monthly Cost |
|---|---|---|---|---|
| 50M input + 25M output / month | MiMo-V2.6-Pro on DeepInfra | 50M | 25M | $38.93 |
| 50M input + 25M output / month | MiMo-V2.6-Flash on DeepInfra | 50M | 25M | $14.00 |
Cost breakdown (Flash):
That is $24.93 less than the same workload on Pro ($38.93). A common middle path is to route by difficulty: send routine steps to Flash and hard ones to Pro. With 80% of this traffic on Flash and 20% on Pro, the month costs about $18.99 ($11.20 + $7.79). Test both on your own tasks before you split traffic.
For developers, the pattern is pretty consistent: if you want MiMo-V2.6-Pro with strong managed API ergonomics and you expect real production usage rather than occasional experiments, DeepInfra is easiest to justify on workloads that are long-context, cache-friendly, tool-heavy, or likely to evolve into more controlled deployments.
Choosing a provider for MiMo-V2.6-Pro is less about finding the one correct answer and more about matching the model’s strengths to your specific cost structure. The model itself is a strong fit for long-context, reasoning-heavy, and multimodal workloads — but what you pay per input token, and whether your cache hits cost $0.00324/M or $0.0036/M, depends entirely on which provider you route through and how deliberately you set up your prompt architecture.
The practical criteria that should drive that decision are straightforward. If your workload involves repeated large prompts — RAG pipelines, shared agent scaffolding, stable system prompts with growing conversation history — cache pricing is the variable that moves the bill most at volume. DeepInfra’s $0.00324/M cached input rate sits at the low end of the listed set, and the gap between that and list-price routes ($0.0036/M) compounds fast once you are processing hundreds of millions of cached tokens per month. If raw token cost is the only constraint and you are comfortable with the operational tradeoffs, GMICloud’s route through OpenRouter is the cheapest listed option today.
One thing worth keeping in mind as you plan: MiMo-V2.6-Pro’s 1M-token context window is genuinely useful, but it also makes it easy to normalize large prompts early in development and only notice the cost pattern later in production. Getting your token math right at the design stage — not after you have shipped — is where the real savings come from. The MiMo-V2.6-Pro model page on DeepInfra gives you direct API access with JSON mode, function calling, and multimodal support — a clean starting point for testing your actual workload before optimizing for scale. For a sense of how its pricing compares to another recent large reasoning model, see our DeepSeek V4 Pro pricing breakdown.
Why Open Models Get Faster Weeks After Launch<p>Qwen3.8-Flash-Next was released on August 26, 2026. The same day, SGLang announced day-0 support, built with the Qwen, NVIDIA and AMD teams, and shipped a dedicated container image. vLLM served the model from its own dedicated image while the architecture-support work was still being merged. llama.cpp, the engine most people use to run models on […]</p>
What Is Google TurboQuant and What Does It Mean for Open Source Inference? - Deep Infra<p>In late March 2026, Google Research published a paper that got more attention outside of academic circles than most AI research does. TurboQuant, a new compression algorithm for the key-value cache in large language models, landed with enough noise that Cloudflare CEO Matthew Prince called it Google’s DeepSeek moment. The Silicon Valley Pied Piper comparisons […]</p>
Kimi K3.1 Pro on DeepInfra: Pricing, Benchmarks & API<p>Moonshot AI’s Kimi K3.1 Pro posts the top scores in its own comparison table on BrowseComp (92.1) and Humanity’s Last Exam with tools (70.0), ahead of the closed frontier models Moonshot tested against, according to the Kimi K3.1 Pro model card (vendor-reported). Independent testing backs up the headline: Artificial Analysis gives it 46 on its […]</p>
© 2026 DeepInfra. All rights reserved.