Is MiMo V2.6 Free

Is MiMo V2.6 free? Yes, but only in specific forms: OpenCode announced a one-week MiMo-V2.6-Flash free route on September 21, 2026, and Xiaomi's two published checkpoints carry MIT metadata. Xiaomi's production API is still metered, Desktop and Token Plan access have separate terms, and self-hosting means supplying storage and compute yourself.

Is MiMo V2.6 Free? The Short Answer

MiMo-V2.6-Flash is temporarily free through OpenCode, but MiMo-V2.6 is not one universally free product. “Free” describes different things depending on which route you take.

That split is the decision. Use the promotion to evaluate a non-sensitive coding task. Use the official API for repeatable production access. Consider the checkpoint files only if your organization already knows how it will host a 178 GB or 573 GB repository.

What Actually Changed With MiMo-V2.6

The practical change is a new model family available through several products, with the hosted API retaining MiMo-V2.5 prices. Xiaomi's launch page is dated September 22, 2026, following the company's September 21 UTC announcement.

Xiaomi MiMo names two primary models. MiMo-V2.6-Pro is the flagship; MiMo-V2.6-Flash is the lower-cost model intended for higher-frequency use. Both publisher-owned model cards describe native text, image, video, and audio handling with a 1M-token context. UltraSpeed is a hosted Pro mode advertised at up to 20x output speed.

For an adopter, the important change is choice of operating model. The same name now sits on a temporary OpenCode route, paid Xiaomi API models, downloadable checkpoints, Desktop, and membership plans. Those surfaces do not share one price or one data policy.

MiMo V2.6 Flash Free: Where the Promotion Works

The promotional route works inside OpenCode, not as a general free API key you can drop into any application. OpenCode's public model feed listed mimo-v2.6-flash-free on September 22, and its live Zen documentation priced input, output, and cached reads as free.

OpenCode announced that route as free for the next week on September 21 at 21:19 UTC. No cutoff hour was published, so the honest description is “one week from the announcement,” not a fabricated expiry timestamp.

A dated access check found an important limit. The public selector exposed the model, but a direct unauthenticated chat request returned HTTP 403 with a FreeTierError saying the free tier can only be used from within OpenCode. No model output was produced.

OpenCode's standard Zen setup says to sign in, add billing details, obtain an API key, connect the client, and select a model. The public sign-in screen offered GitHub or Google. The public material did not say whether this promotion waives a payment card or initial credit purchase, so “free tokens” should not be rewritten as “no signup, no card, and no deposit.” Check the billing screen presented to your account before you proceed.

The privacy exception matters more than the token discount. OpenCode says most hosted models follow zero-retention and no-training policies, then explicitly excepts this free Flash route: data collected during the free period may improve the model. That makes it suitable for synthetic evaluation prompts, public code, and throwaway tasks, not confidential production work.

MiMo V2.6 License: What MIT Does and Does Not Cover

Both publisher-owned checkpoint repositories display License: mit, but that label applies to the released model files, not to free hosting everywhere. The Flash repository and Pro repository carried the same metadata when verified on September 22.

Neither visible file listing included a standalone LICENSE file. Both showed a README, technical report, configuration, tokenizer, deployment code, and weight shards. A legal or procurement review should preserve the exact repository snapshot and terms it approved rather than rely on a search snippet or assume that a metadata tag governs a separate hosted service.

The files also make “free to download” different from “free to run.” Hugging Face displayed the Flash repository at 178 GB and the Pro repository at 573 GB. The cards document Transformers, vLLM, SGLang, Docker Model Runner, and compatible quantized-model apps, but the model download does not include GPUs, storage, networking, monitoring, or engineering time.

This is the same open-weight distinction that matters in the GLM-5.2 review: permissive checkpoint access can remove a model-license bill while leaving the infrastructure bill untouched.

MiMo V2.6 API Cost After the Promotion

The paid fallback is inexpensive on Flash and sharply tiered above it. Xiaomi's launch API table and live API price page, both checked September 22, list these real-time rates per million tokens:

• Flash: $0.0028 cached input, $0.14 uncached input, and $0.28 output.
• Pro: $0.0036 cached input, $0.435 uncached input, and $0.87 output.
• Pro-UltraSpeed: $0.036 cached input, $4.35 uncached input, and $8.70 output.

For a normalized monthly workload of 100M uncached input tokens and 20M output tokens, Flash costs $19.60, Pro costs $60.90, and UltraSpeed costs $609. The same workload makes Pro 3.11x the Flash bill. UltraSpeed is 10x the Pro bill.

Those multiples create a clean rule: default to Flash, promote a task to Pro only when fewer retries or better accepted outputs recover a 3.11x token premium, and choose UltraSpeed only when latency is worth 10x the Pro rate. “Up to 20x faster” is a vendor speed claim, not an automatic cost saving.

Batch processing cuts the supported Pro and Flash rates by 50%. The normalized workload falls to $9.80 on Batch Flash or $30.45 on Batch Pro, provided the job can wait for non-real-time processing. Cache writes are temporarily free, while overseas web search is billed separately at $5 per 1,000 calls.

Xiaomi says V2.6 retained V2.5 API pricing. The post-promotion move is therefore predictable: point eligible workloads at mimo-v2.6-flash, keep a monthly limit, and measure cost per accepted result. The DeepSeek pricing breakdown explains the same product split: a free surface and a usage-billed API can share a brand without sharing an entitlement.

What It Means for Builders, Operators, and Buyers

The release matters differently depending on who owns the work.

Builders: evaluate the interface before the model

Builders should use the free route to answer one narrow question: does Flash complete your representative tool-calling or coding task well enough to merit integration work? Keep the prompt synthetic and portable. If the result clears your bar, repeat it through the paid Xiaomi API before committing, because the promotional endpoint and its data terms are not a production contract.

Operators: the batch switch can matter more than the model switch

Operators with asynchronous jobs can halve the listed token bill by using Xiaomi's Batch API. That savings is deterministic. A move from Flash to Pro is not: Pro must improve accepted outcomes enough to offset its 3.11x uncached-input and output rates. Track attempts, accepted completions, latency, and human review time together.

MiMo V2.6 Pro: pay for harder work, not the label

Pro belongs on difficult, high-value tasks after Flash misses a written quality threshold. Xiaomi reports a 46.32 Artificial Analysis Intelligence Index v4.3 score for Pro, but that is an attributed launch result, not an independent test performed here. A funded founder may value fewer failures on a complex repository; a solo builder with small jobs may never recover the premium.

MiMo Coding Plan and Desktop: separate budget lines

Buyers should treat Xiaomi's end-user products separately from API metering. MiMo Desktop left early access with this release, while existing early-access users were told that access would continue for one more week. Xiaomi's Individual Token Plan documentation displayed undiscounted monthly tiers of $6, $16, $50, and $100, with lower annualized or promotional figures also shown. Confirm the account, region, included credits, and renewal terms on the purchase screen instead of assuming the OpenCode week carries over.

Who Should Act Now, Wait, or Ignore It

Act now if you can evaluate with a synthetic task and you already use OpenCode. The promotion can remove the token charge from a controlled comparison, and one week is enough to measure completion quality, retries, and latency on a small task set.

Wait if your evaluation requires proprietary code, customer data, a service-level commitment, or verified no-card access. The free endpoint's data-use exception and account uncertainty are enough reasons to use the paid API or postpone the test.

Ignore the promotion if you already chose a provider on accepted-result economics and MiMo does not address a measured failure. A $0 endpoint does not justify migration work by itself. Likewise, an organization without inference infrastructure should not download hundreds of gigabytes merely because the repository metadata says MIT.

What's Overhyped

The biggest overstatement is “MiMo-V2.6 is free.” One temporary route is free, two repositories carry MIT metadata, and nearly every durable operating path still has a bill.

“MIT” also gets stretched too far. It does not make OpenCode, Xiaomi's API, a GPU fleet, or a Desktop subscription free. The missing standalone license file in the visible repositories is another reason for compliance teams to archive the exact terms they rely on.

The benchmark story needs restraint too. Xiaomi's launch materials report strong coding, agent, visual, and cyber results, but a vendor table cannot price your retry rate or validate your prompt distribution. Run the task you buy models to perform.

Finally, UltraSpeed's “up to 20x” output speed does not mean 20x better value. Its token rates are 10x Pro's. It can make sense for latency-sensitive work, but not for a queue that can use batch processing at half the real-time rate.

The Monday Move

Use the promotion as a one-task measurement window, not a migration event.

1. 1

   Write one safe task

   Choose a synthetic coding or agent task with a clear pass condition. Remove proprietary code, credentials, customer data, and internal documents.

2. 2

   Check access before testing

   Sign in through OpenCode, open the model selector, and confirm mimo-v2.6-flash-free is still present. Record any billing-detail, card, or credit requirement your account shows.

3. 3

   Run it once

   Use the OpenCode client, because a direct unauthenticated request was rejected. Record the output, elapsed time, retries, and whether the result passed. If the route is gone or account terms are unacceptable, stop there.

4. 4

   Price the durable route

   Run the same evaluation through paid Flash only when the data is safe for that provider. Estimate the monthly bill from uncached input and output, then escalate to Pro only if the quality gain beats 3.11x.

Frequently Asked Questions

Is MiMo free to use?

Temporarily, through OpenCode's MiMo-V2.6-Flash free route. The downloadable checkpoints also display MIT metadata, but Xiaomi's API, Desktop access, subscription plans, and self-hosting resources are separate.

Is Xiaomi MiMo free?

Not as one blanket product. OpenCode's Flash promotion is free for one week from its September 21 announcement, while Xiaomi's official API is metered and its end-user products follow separate account and plan terms.

How good is MiMo-V2?

Xiaomi reports that MiMo-V2.6-Pro scored 46.32 on Artificial Analysis Intelligence Index v4.3 in September 2026. Treat that as an attributed benchmark and test the model on your own acceptance criteria before changing a workflow.

How much does MiMo-V2.5-Pro cost?

Xiaomi's current pricing page lists V2.5 Pro beside V2.6 Pro at $0.0036 cached input, $0.435 uncached input, and $0.87 output per million tokens, while marking V2.5 Pro for deprecation. A new deployment should evaluate V2.6 Pro instead.

Which is better, MiMo-V2.5-Pro or DeepSeek V4 Pro?

MiMo-V2.5-Pro is the outgoing generation, so it is the wrong baseline for a new buying decision. Compare DeepSeek V4 Pro with MiMo-V2.6 Pro on the same task, then judge cost per accepted result rather than one vendor benchmark.

How much does MiMo Pro cost?

MiMo-V2.6-Pro costs $0.0036 per million cached input tokens, $0.435 per million uncached input tokens, and $0.87 per million output tokens on Xiaomi's real-time API. Batch rates are half those figures for supported jobs.

Is mimo worth paying for?

Pay for Pro only when it reduces retries or improves accepted outputs enough to recover a 3.11x token-price multiple over Flash. For non-real-time work, test Batch Flash first because its listed rates are 50% below real-time Flash.

Is mimo-V2.5-pro good?

Its historical capability matters less now that Xiaomi marks it for deprecation. Starting a new workflow on V2.5 Pro adds migration risk when V2.6 Pro is already available at the same listed API rates.

Is mimo better than Duolingo?

That question mixes two unrelated entities named Mimo or MiMo. It is not a meaningful comparison involving Xiaomi MiMo-V2.6 and should not influence a model-access decision.

Want the next model release translated into a budget and workflow decision? Join the newsletter.