Xiaomi MiMo 2.5 Pro

What it is

Xiaomi MiMo 2.5 Pro is Xiaomi's previous-generation open model: efficient, on-device-friendly weights with a reasoning score that does not match its size or its price. Xiaomi is a consumer hardware company first, and the MiMo line reflects that — models designed to be small enough for the devices the company already sells, rather than to top leaderboards.

As of September 2026 the 2.5 Pro has been superseded by the MiMo 2.6 series at the same price, which changes what you should buy but not what this page measures. Verify the weights and licence on Xiaomi's channels before deploying.

How the composite is built

Overall 82 weights Coding 24%, Terminal 20%, Reasoning 20%, Tool use 15%, Context 13%, Speed 8%. A tight profile from 79 to 84 — no spikes, no holes, and a spread that is unusual at this price.

Coding 83

Two behind GLM-5's 85 and five behind Terra's 88. Everyday code work competently handled.

Terminal 79

Level with GLM-5 and DeepSeek R2, and sixteen behind Opus 5's 95. Supervised shell work only.

Reasoning 84

Its best axis and the note's claim made concrete: level with Qwen 4 Coder and Mistral Large 3, both more expensive propositions.

Tool use 79

Level with DeepSeek R2's 79 and one behind GLM-5's 80. Workable in short chains with validation in the harness.

Context 81

Mid-field. File and module scale; the long-context specialists like Kimi k3 sit ten points higher.

Speed 83

Nine above GLM-5's 74 and level with Falcon 3's 83. The efficient-weights story is measurable here.

Where it fits

Cost-controlled work at the sweet spot of this board: agent loops on a budget, product features that need reasoning without frontier pricing, and deployments where the model may end up running on-device rather than in a datacentre. It is the model on this board where the price-to-capability arithmetic is hardest to argue with.

Limits

• It has been superseded. The MiMo 2.6 series arrived in September 2026 at the same $0.43/$0.87 rate — new work should look there first and treat this page as the record for 2.5 Pro.
• Terminal 79 and Tool use 79 cap agentic autonomy. Keep humans or stricter models on anything destructive.
• No axis reaches 85. When quality rather than economics decides the project, the frontier tiers are a real step up.
• First-party hosting is operated from China. For regulated data, use the open weights or the third-party gateways rather than the direct endpoint.

Price and access

$0.43 per million input tokens and $0.87 per million output, blending to $0.54/M. The Artificial Analysis index is 26.0, giving 48.1 index points per blended dollar — the best value rating on this board after GPT-5.6 Luna, and the reason the column reads Good. Listed on OpenRouter and OpenCode Go, with weights published for self-hosting. The unscored MiMo 2.6 listings on this board span $0.14/$0.28 to $4.35/$8.70 depending on tier. Last scored 15 Sep 2026.

Alternatives on this board

• GLM-5 — 82 at $0.60/$1.92 with the same overall and three gateways. More expensive, slightly stronger on code.
• GPT-5.6 Luna — 81 at $0.20/$1.20 with Speed 90 and Context 89. Cheaper and closed.
• Mistral Small 3.2 — 79 at $0.09/$0.25 with open weights, for the tightest budgets.
• DeepSeek V4 — 86 at $0.78/$1.57, when the capability ceiling matters more than the rate.

Sources