Overview
Xiaomi MiMo V2.5 entered public beta on April 23, 2026 — a new V2.5 model family led by MiMo-V2.5 and MiMo-V2.5-Pro on the text/omnimodal side, with V2.5-TTS Series and V2.5-ASR announced alongside. It replaces the one-flagship setup from the V2-Pro release just five weeks earlier. The redesign pushes on two axes at once: V2.5-Pro targets very long agentic runs where Claude Opus 4.6 and GPT-5.4 currently lead, while the non-Pro V2.5 folds what used to be the separate V2-Omni multimodal model into the main text reasoning model and drops it to roughly half the price.
The launch also swaps the Token Plan credit scheme, removes the 256K/1M tier penalty, and comes with a full Credits reset for existing subscribers before 22:00 CST on April 22, 2026.
What's New
MiMo-V2.5-Pro: Longer-Horizon Agents
Xiaomi positions V2.5-Pro explicitly against Claude Opus 4.6 and GPT-5.4 for long-horizon agentic work rather than single-turn benchmarks — framing it as able to "go head-to-head" on agentic scenarios rather than citing third-party consensus. Xiaomi reports stable execution across roughly 1,000 tool calls in a single session within a 1M-token context window, with maximum output extended to 128K tokens. Two long-running tasks were disclosed as evidence rather than leaderboard scores:
- Rust SysY compiler (Peking University compiler-principles project). V2.5-Pro wrote a full compiler pipeline — lexer, parser, AST, Koopa IR codegen, RISC-V assembly backend, and performance optimization — in 4.3 hours across 672 tool calls. Hidden test score: 233/233, with perfect sub-scores on Koopa IR (110/110), RISC-V backend (103/103), and optimization (20/20). Cold-start compile pass rate was 59% (137/233) on the first build, meaning the architecture was correct before the test harness ever ran.
- Video editor web app. From the prompt "build a video editor web app," V2.5-Pro produced a working 8,192-line application with multi-track timeline, clip trimming, crossfade, audio mixing, and export — across 1,868 tool calls over 11.5 hours of autonomous work.
Instruction-following in long agent sessions was also specifically called out — V2.5-Pro is described as picking up implicit requirements from earlier context and staying logically consistent across very long runs, which is where AI agent frameworks typically break down in production.
MiMo-V2.5: Native Omnimodal at 1M Context
The non-Pro V2.5 is a native omnimodal model that handles image, audio, and video input in the same reasoning loop rather than routing through a separate vision or speech pipeline. Two claims frame the upgrade:
- Agent capability surpasses MiMo-V2-Pro on Claw-Eval — meaning the cheaper, omnimodal sibling now beats the previous flagship on Xiaomi's primary agent benchmark, while cutting API cost by roughly 50% at matched throughput.
- Multimodal perception surpasses MiMo-V2-Omni on VideoMME, CharXiv, and MMMU-Pro, approaching or exceeding top closed-source models on those evaluations.
Context window is 1M tokens with 128K max output — matching V2.5-Pro on capacity but at a lower price tier. V2.5 is positioned for the "everyday" Agent scenarios where V2.5-Pro's long-horizon reliability is overkill.
Token-Efficiency Gains Across the Family
At matched ClawEval scores, Xiaomi measures:
- V2.5-Pro saves 42% tokens vs Kimi K2.6 at the same agent benchmark score.
- V2.5 saves 50% tokens vs Muse Spark at the same agent benchmark score.
This matters because agent billing scales with tool-call verbosity as much as with per-million-token rates — a 40-50% reduction in tokens consumed can outweigh headline input/output price comparisons for multi-turn workloads.
Four-Model Family + Token Plan Reset
Alongside the two LLMs, V2.5-TTS Series covers three speech-synthesis variants (base, voice-clone, voice-design) and V2.5-ASR handles speech recognition. Xiaomi's pricing page currently lists all TTS/ASR variants as free during the beta window, though individual endpoints may roll out after the core V2.5 / V2.5-Pro launch. Token Plan credit rates flattened to V2.5 = 1×, V2.5-Pro = 2× with the 256K/1M credit multiplier removed; overnight (00:00–08:00 CST) calls get an additional 20% off; new annual plans come in at 88% of list price.
Performance Benchmarks
The two long-horizon evidence items are worth restating because they describe a capability profile not captured by standard single-turn evals.
| Task | Duration | Tool Calls | Outcome |
|---|---|---|---|
| Peking University SysY compiler in Rust | 4.3 hours | 672 | 233/233 on hidden tests (Koopa IR 110/110, RISC-V 103/103, perf 20/20); 59% cold-start compile pass rate |
| "Build a video editor web app" | 11.5 hours | 1,868 | 8,192 lines of working code; multi-track timeline, crossfade, audio mixing, export |
For MiMo-V2.5 on the omnimodal side, Xiaomi cites gains on VideoMME, CharXiv, and MMMU-Pro over MiMo-V2-Omni but has not yet published the exact numbers — treat those as directional until the official model card lands.
The agent-benchmark claim "V2.5 surpasses V2-Pro on Claw-Eval" is the most load-bearing comparison in the release, since it supports the pricing delta: the cheaper model is being marketed as strictly better on the metric Xiaomi cares about.
Availability & Access
V2.5 launched as a public beta rather than a general-availability release, which shapes how access works today:
| Access Path | Details |
|---|---|
| platform.xiaomimimo.com | API key issuance + console + usage dashboard |
| Xiaomi AI Studio | Free interactive testing without billing setup |
| OpenAI / Anthropic API protocols | Drop-in compatibility at the SDK level |
| IDE integrations | OpenCode, Claude Code, Cline, Kilo Code, Roo Code, Codex, Cherry Studio, Zed, Qwen Code, OpenClaw |
| Open-source weights | Xiaomi says MiMo-V2.5-Pro and MiMo-V2.5 will be open-sourced "soon"; release venue and timing have not been specified at launch |
Rate limiting is shared across all API keys on a single account: 100 RPM and 10M TPM per model. Cache-write pricing is free during the beta window, and network-search plugins bill separately ($5 per 1,000 calls overseas).
Pricing & Plans
V2.5 uses per-token API pricing; Token Plan subscriptions are optional. All overseas rates are per 1M tokens.
| Model | Tier | Input | Cached Input | Output |
|---|---|---|---|---|
| V2.5-Pro | ≤256K tokens | $1.00 | $0.20 | $3.00 |
| V2.5-Pro | 256K–1M tokens | $2.00 | $0.40 | $6.00 |
| V2.5 (omnimodal) | ≤256K tokens | $0.40 | $0.08 | $2.00 |
| V2.5 (omnimodal) | 256K–1M tokens | $0.80 | $0.16 | $4.00 |
| V2.5-TTS / voiceclone / voicedesign | — | Free (beta) | — | — |
| V2.5-ASR | — | Free (beta) | — | — |
V2.5-Pro pricing matches V2-Pro at both tiers — new flagship, same cost. V2.5 comes in at 40% of V2.5-Pro's input rate and 67% of its output rate, so the practical decision is "use V2.5 unless the job explicitly needs V2.5-Pro's long-horizon reliability."
For Token Plan subscribers, the key changes vs the V2-Pro-era scheme are:
- Credit cost now flat at 1× (V2.5) / 2× (V2.5-Pro) — the old 4× Credits multiplier on the 256K–1M tier is gone.
- Overnight window (00:00–08:00 CST) applies an additional 20% discount on Credits consumed.
- Auto-renewal adds 30% off next month for existing users (23% for new users), each limited to one use; annual plans land at 88% of list.
- All Credits balances purchased before 22:00 CST on April 22, 2026 were reset, but subscription time remaining was preserved.
Best For
- Engineering teams running agent pipelines that exceed 100 tool calls per session and hit reliability walls on shorter-context models
- Teams migrating off Claude Opus 4.6 or Claude Sonnet 4.6 where long-context work dominates the bill
- Developers who want an omnimodal model with 1M context at a sub-dollar input rate — a pricing tier Kimi, GLM-5, and Western frontier models currently don't match
- Token Plan subscribers who benefit from the new flat credit scheme plus overnight and annual discounts
- Researchers evaluating Chinese frontier LLM progress alongside GLM-5 and Kimi K2.x series at comparable capability tiers
FAQ
Is MiMo V2.5 a replacement for MiMo-V2-Pro?
Practically yes, but the upgrade path depends on workload. V2.5-Pro supersedes V2-Pro on long-horizon agent reliability at identical pricing. For workloads that previously ran on V2-Pro for general capability, the non-Pro V2.5 is a cheaper option and is claimed to beat V2-Pro on Claw-Eval while cutting API cost by roughly half.
How does V2.5-Pro compare to Claude Opus 4.6 and GPT-5.4?
Xiaomi positions V2.5-Pro as able to "go head-to-head" with those models on general agentic ability, complex software engineering, and long-horizon tasks — specifically highlighting ~1,000-tool-call stability and the two demonstrated multi-hour tasks. Independent third-party benchmarks have not yet been published, so treat the positioning as vendor-stated until external evaluations arrive.
What is the difference between MiMo-V2.5 and MiMo-V2-Omni?
MiMo-V2.5 replaces V2-Omni as the omnimodal line. It extends the context window from V2-Omni's 256K to 1M tokens, adds the 256K–1M tier pricing, and claims gains over V2-Omni on VideoMME, CharXiv, and MMMU-Pro — while also surpassing V2-Pro on Claw-Eval, which V2-Omni did not.
Are V2.5 and V2.5-Pro open source?
Not yet. Xiaomi has announced open-source weight releases for both V2.5 and V2.5-Pro but has not published them — or committed to a specific venue or timing — at launch. Earlier MiMo releases (V1 / MiMo-7B and V2-Flash) are open-weight under the XiaomiMiMo organization on Hugging Face and GitHub, but V2.5/V2.5-Pro destinations have not been confirmed.
What changed in the Token Plan at the V2.5 launch?
Three things: the 256K–1M Credit multiplier was removed (all tiers now cost 1× for V2.5 or 2× for V2.5-Pro per token), overnight calls 00:00–08:00 CST get an extra 20% discount, and annual/auto-renewal plans add 12–30% off. All existing Credits balances were reset to zero at 22:00 CST on April 22, 2026, without affecting subscription duration.
