Xiaomi MiMo-V2.5 vs Claude: Superseded by V2.6 (2026)

On April 22, 2026, Xiaomi released two open-weight models that landed with very little blog coverage and a very loud vendor claim: that MiMo-V2.5-Pro "sits next to Claude on coding." Both models — MiMo-V2.5 and the larger MiMo-V2.5-Pro — are fully open-sourced under an MIT license, which alone makes them worth a serious look for any team running agentic coding at scale.

This article does the unglamorous part: separating what Xiaomi claims from what independent measurement shows, and answering the only question that matters for an engineering team — can you actually replace Claude with this for production coding work, and at what cost? Short version: when this released it was the strongest open-weight agentic coder available, the efficiency story was real, and it still trailed Claude on the hard benchmarks. Five months later Xiaomi superseded it with MiMo-V2.6-Pro, at the same list price and roughly 20 Intelligence Index points higher, so the "is the gap worth the price" question has a different answer now than it did in April. Everything below is the record on V2.5 — read the V2.6-Pro guide if you are picking a model today rather than maintaining a deployment on this one.

What are MiMo-V2.5 and MiMo-V2.5-Pro?

They are two distinct open-weight models in the same release, not a small/large pair of the identical architecture:

• MiMo-V2.5 — a 310B-parameter sparse Mixture-of-Experts (MoE) model with ~15B active parameters per token. It is omnimodal: native text, image, video, and audio understanding in one model. Trained on roughly 48T tokens. 1M-token context. MIT license.
• MiMo-V2.5-Pro — a 1.02-trillion-parameter MoE with 42B active parameters per token, 384 routed experts, 8 experts activated per token. Text-focused, tuned for long-horizon agentic coding. Trained on 27T tokens. 1M-token context. MIT license.

The Pro variant is the one Xiaomi positions against Claude. The smaller MiMo-V2.5 is positioned as the efficient daily driver — Xiaomi states it matches V2.5-Pro on everyday coding at roughly half the inference cost, with multimodal capability the Pro model does not have.

Both are downloadable from Hugging Face (XiaomiMiMo/MiMo-V2.5 and XiaomiMiMo/MiMo-V2.5-Pro) with weights, tokenizer, and full model cards. GGUF quantizations (community, via Unsloth) exist for both.

How is MiMo-V2.5-Pro architected for efficiency?

The efficiency angle is the most technically interesting part of the release, and it is the thing that makes the cost story credible rather than just a price cut.

MiMo-V2.5-Pro uses a hybrid attention architecture: it interleaves Sliding Window Attention (SWA) and full Global Attention (GA) layers at a 6:1 ratio (per the model card, 60 SWA layers to 10 full-attention layers, SWA window size 128, GQA with 8 KV heads). Xiaomi reports this reduces KV-cache storage by roughly 7x versus an all-global-attention model of the same size while preserving long-context performance through a learnable attention-sink bias.

Why this matters in practice: for long agentic coding trajectories — the kind where an agent reads a repo, edits, runs tests, re-reads, and iterates for hours — KV-cache memory and per-token cost are the binding constraints, not raw parameter count. A 7x smaller KV cache is the difference between a trajectory that fits on a given GPU configuration and one that does not.

How does MiMo-V2.5-Pro benchmark against Claude Opus 4.7?

This is where the careful reading matters. There are three categories of number in circulation, and they tell different stories.

1. Xiaomi's vendor-reported benchmarks (treat as vendor-claimed): Xiaomi reports MiMo-V2.5-Pro at 78.9 on SWE-bench Verified, 57.2 on SWE-Bench Pro, and 68.4 on Terminal-Bench 2.0, with ~64% on the general ClawEval subset. Xiaomi also reports a head-to-head on its own internal coding benchmark of 73.7 for MiMo-V2.5-Pro vs 77.1 for Claude Opus 4.6 — i.e. by Xiaomi's own measure it trails the previous Claude generation. Note all of these are Xiaomi's internal numbers; the "sits next to Claude on coding" framing is Xiaomi's marketing, not an independent finding.

2. Anthropic's reported numbers for Claude Opus 4.7 (released April 16, 2026, vendor-reported): 87.6% on SWE-bench Verified and 64.3% on SWE-Bench Pro. Even comparing vendor-claim to vendor-claim across labs — which is methodologically loose — Claude Opus 4.7 leads MiMo-V2.5-Pro on both SWE-bench tracks.

3. Independent third-party measurement (Artificial Analysis): on AA's current Intelligence Index v4.3.2 (captured 5 October 2026) MiMo-V2.5-Pro scores 25.99, at $0.05 per Index task and a $138.26 whole-suite cost. The plain MiMo-V2.5 scores 25.17. For context on the same index version, Claude Opus 4.7 scores 40.69 and MiMo's own V2.6-Pro scores 46.32. We previously published 54 here. That figure was a v4.1.1-era capture, and AA has since rebased the index — on a rebase every score moves because the benchmark basket changed, not because the models did, and there is no conversion factor between versions. The honest v4.3.2 statement is narrower than what we wrote in April: V2.5-Pro was competitive among open-weight models at release and has since been passed by its own successor, by GLM-5.3 (Max, 44.78) and by Kimi K3 (Max, 43.59).

The honest read: MiMo-V2.5-Pro does not match Claude Opus 4.7 on the hard, widely-cited coding benchmarks — even on Xiaomi's own numbers it trails Claude Opus 4.6. (This is the same pattern we saw in the DeepSeek V4 vs Claude Opus 4.7 head-to-head: a strong open-weight model that lands close but not ahead on the hardest tasks.) What it does is land within striking distance while being open-weight and dramatically cheaper per token. "Can it match Claude?" — on raw capability, no. On capability-per-dollar for many agentic workloads, it is genuinely competitive.

What is the real cost and efficiency difference vs Claude?

This is where MiMo-V2.5-Pro is most defensible, and it is a two-part story: lower per-token price and fewer tokens per task.

• Per-token price. Artificial Analysis tracks MiMo-V2.5-Pro at $0.435 per 1M input tokens and $0.87 per 1M output — the same rate OpenRouter lists for xiaomi/mimo-v2.5-pro, and the same rate Xiaomi later carried over unchanged to V2.6-Pro. (We previously published $1.00/$3.00 here; that was wrong, and in MiMo's favour to correct.) Claude Opus 4.7's published list pricing is an order of magnitude higher per token ($5 input / $25 output per 1M, per Anthropic's rate card). To compare like-for-like on the cache-hit scenario: Anthropic prices cached input at $0.50 per 1M tokens (10% of list, per Anthropic's rate card), so even discounting both sides for caching, MiMo-V2.5-Pro remains materially cheaper per token.
• Tokens per task. Xiaomi reports MiMo-V2.5-Pro completes ClawEval agentic trajectories using ~70K tokens each — roughly 40–60% fewer tokens than Claude Opus 4.6, Gemini 3.1 Pro, and GPT-5.4 at comparable capability (vendor-claimed; not independently verified). Because the hybrid-attention KV-cache reduction is architectural, the "fewer tokens" claim is at least plausible rather than purely marketing.

Multiply those two effects together and the cost gap for a high-volume agentic-coding workload becomes large — frequently cited as an order of magnitude or more cheaper per completed task. The flip side: Xiaomi's token-efficiency comparisons are against Claude Opus 4.6, and they are Xiaomi's measurements. Treat the direction (much cheaper) as solid; treat the exact multiplier as vendor-flavored until independent harnesses confirm it on your task distribution.

Can you realistically self-host MiMo-V2.5?

Yes — and the MIT license means you can self-host, fine-tune, and commercially deploy without negotiation. But "can you" depends heavily on which variant.

MiMo-V2.5 (310B / 15B active) is the realistic self-host target for most teams. It runs under SGLang with FP8 quantization on a multi-GPU node. Per the model card, a representative SGLang launch:

python3 -m sglang.launch_server \
  --model-path XiaomiMiMo/MiMo-V2.5 \
  --served-model-name mimo-v2.5 \
  --pp-size 1 --dp-size 2 --tp-size 8 \
  --enable-dp-attention \
  --moe-a2a-backend deepep \
  --quantization fp8 \
  --context-length 262144 \
  --attention-backend fa3

MiMo-V2.5-Pro (1.02T / 42B active) is realistically a cluster, not a workstation. The model card's reference SGLang config uses tensor parallelism of 16 and expert parallelism of 16 (e.g. --tp-size 16 --ep-size 16 --quantization fp8 --context-length 1048576). That is a serious multi-node deployment. vLLM is also supported via prebuilt images (vllm/vllm-openai:mimov25-cu130 / cu129) with tensor parallelism and expert parallelism enabled; the vLLM Recipes site documents the current config, and our vLLM vs Ollama vs LM Studio production benchmark covers how that runtime choice plays out at scale.

Practical guidance: if you want to own the model on your own hardware, target MiMo-V2.5 (the 310B). If you want MiMo-V2.5-Pro's capability without a cluster, use one of the hosted API providers serving it — you still get the open-weight pricing benefit ($0.435/$0.87 per 1M tokens via tracked providers) and the MIT license still means no vendor lock-in on the weights themselves. Recommended sampling for both: temperature=1.0, top_p=0.95.

Has MiMo-V2.5 been superseded?

Yes. On 21 September 2026 Xiaomi published the MiMo-V2.6 series, and it changes the answer to almost every question on this page. V2.6-Pro keeps the same 1.02T/42B MoE shape and the same $0.435/$0.87 list price, and on Artificial Analysis's Intelligence Index v4.3.2 it scores 46.32 against V2.5-Pro's 25.99 — the largest single-generation jump in the open-weights field this year, and enough to make it the top-scoring downloadable-weights model AA tracks. On Xiaomi's own agent benchmarks the gap is wider still: DeepSWE v1.1 goes from 19.0 to 71.9 and Terminal Bench 4.0 from 1.5 to 34.9.

Two practical notes if you act on that. First, Xiaomi publishes only suffixed repositories — XiaomiMiMo/MiMo-V2.6-Pro-RL and XiaomiMiMo/MiMo-V2.6-Pro-MOPD; there is no bare MiMo-V2.6-Pro repo, so a download command using the plain name fails. Second, prefer the MOPD checkpoint for agent work — it carries a distillation pass that fixes a repeated-tool-call failure mode in the launch weights. Full detail in our MiMo-V2.6-Pro guide.

V2.5 still matters if you already run it: the weights are MIT and nothing revokes them, the 310B variant remains the cheaper self-host, and the architecture discussion below carries straight over to V2.6, which reuses it.

What are the limitations and caveats?

• It is not at Claude Opus 4.7's capability tier. On the hard benchmarks it trails — by Xiaomi's own numbers it trails Claude Opus 4.6. For the most demanding autonomous-engineering work, Claude Opus 4.7 is still ahead.
• Most headline numbers are vendor-reported. The 78.9 SWE-bench Verified, the 40–60% token savings, and the "sits next to Claude" framing are Xiaomi's. Independent measurement (Artificial Analysis Intelligence Index v4.3.2: 25.99) puts it well below frontier tier and below its own successor.
• Generation speed is mid-pack. Artificial Analysis measures MiMo-V2.5-Pro at 43.83 median output tokens per second with a 2.14-second time to first chunk — below average for its class. For latency-sensitive interactive use this matters; for batch/background agentic sweeps it usually does not.
• Pro is a cluster-scale self-host. The 1.02T variant is not a single-box deployment. The economically realistic self-host is MiMo-V2.5 (310B).
• Xiaomi's Claude comparison uses Opus 4.6, not 4.7. Anytime you see "matches Claude," check which Claude. The gap to the current Claude generation (4.7) is wider than the gap to 4.6.

What's the verdict for an engineering team?

At release MiMo-V2.5-Pro was the most credible open-weight agentic coding model available — ahead of the field we cover in the Kimi K2.6 vs DeepSeek V4 vs GLM-5.1 open-weights coding verdict — and the efficiency engineering behind it is real, not just a price tag. But "can it match Claude?" deserves a straight answer: no, not on raw capability — it trails Claude Opus 4.7 on the hard benchmarks, and trails even Claude Opus 4.6 on Xiaomi's own coding bench.

Where it wins is the economics. For high-volume, cost-sensitive agentic workloads — large refactor sweeps, CI-driven agents, internal tooling that runs thousands of trajectories — the combination of $0.435/$0.87 token pricing, fewer tokens per task, and an MIT license made MiMo-V2.5-Pro the rational choice even though it is the less capable model — though V2.6-Pro now offers the same pricing with a far higher score, so that argument has moved to the newer model. For frontier-difficulty autonomous engineering where capability dominates cost, stay on Claude Opus 4.7 — and if you go that route, our breakdown of Claude Opus 4.7 task budgets for agentic coding helps keep that spend predictable. Many teams will end up running both: MiMo for volume, Claude for the hard 10%.

FAQ

Is Xiaomi MiMo-V2.5 actually open source?

Yes. Both MiMo-V2.5 and MiMo-V2.5-Pro are released under the MIT license with weights, tokenizer, and full model cards on Hugging Face. MIT permits commercial use, fine-tuning, and redistribution without a separate agreement — it is one of the most permissive licenses an open-weight model of this size has shipped under.

Does MiMo-V2.5-Pro beat Claude Opus 4.7 at coding?

No. On the widely-cited coding benchmarks, Claude Opus 4.7 leads — Anthropic reports 87.6% on SWE-bench Verified and 64.3% on SWE-Bench Pro versus Xiaomi's reported 78.9 and 57.2 for MiMo-V2.5-Pro. By Xiaomi's own internal coding benchmark it even trails Claude Opus 4.6 (73.7 vs 77.1). It competes on cost-efficiency, not raw capability.

What is the difference between MiMo-V2.5 and MiMo-V2.5-Pro?

MiMo-V2.5 is a 310B-parameter omnimodal MoE (text, image, video, audio) with ~15B active parameters — the efficient daily driver and the realistic self-host target. MiMo-V2.5-Pro is a 1.02T-parameter text-focused MoE with 42B active parameters and a hybrid attention architecture, tuned for long-horizon agentic coding and positioned against Claude.

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

Artificial Analysis tracks MiMo-V2.5-Pro at $0.435 per 1M input tokens and $0.87 per 1M output — the rate OpenRouter also lists. Claude Opus 4.7 list pricing is $5 input / $25 output per 1M tokens (Anthropic's published rate), so roughly 11x the input and 29x the output cost per token, before factoring MiMo's claimed lower token count per task. On AA's own per-task measure, V2.5-Pro runs the Intelligence Index for $0.05.

Can I self-host MiMo-V2.5-Pro?

Technically yes (the MIT license allows it), but the 1.02T-parameter Pro variant realistically needs a multi-node cluster — the reference SGLang config uses tensor and expert parallelism of 16. For most teams the practical self-host is the 310B MiMo-V2.5; for Pro, use a hosted provider serving the open weights.

Are the MiMo-V2.5 benchmarks independently verified?

Mostly not. The SWE-bench, Terminal-Bench, ClawEval, and token-efficiency figures are Xiaomi-reported. The main independent data point is Artificial Analysis's Intelligence Index v4.3.2 score of 25.99, which puts it below frontier closed models and below Xiaomi's own MiMo-V2.6-Pro (46.32). Treat vendor benchmarks as directional and validate on your own task distribution before switching production workloads.

Should I replace Claude with MiMo-V2.5 for agentic coding?

For high-volume, cost-sensitive workloads where the capability gap is tolerable — yes, it can be the rational choice. For frontier-difficulty autonomous engineering, keep Claude Opus 4.7. The common pattern is a tiered setup: MiMo-V2.5 for the bulk of trajectories, Claude for the hardest tasks where capability outweighs cost.

If you are building this kind of multi-model agentic infrastructure — routing, harnesses, cost guardrails, and evals across open and closed models — and want engineers who have shipped it in production, Codersera matches you with vetted remote developers experienced with agentic coding stacks. We run a risk-free trial so you can validate technical fit before committing.