Xiaomi’s MiMo-V2.6-Pro combines a 1T-parameter MoE with low API prices

Artificial Analysis ranks Xiaomi’s MiMo-V2.6-Pro as the leading open-weight model on its Intelligence Index at the time of release. Its score of 46 exceeds MiMo-V2.5-Pro’s 26, while Xiaomi lists output pricing below $1 per million tokens. The combination makes the model relevant to teams evaluating coding agents, research systems, browser automation, and other tool-heavy workloads.

The Intelligence Index aggregates several capability evaluations into one score. Rankings can change as models and test suites are updated, so the 20-point generational gain carries more weight than the temporary leaderboard position. Open weights also refers to checkpoint availability; licensing terms, training-data disclosure, and reproducibility remain separate considerations.

Sparse compute keeps inference cheap

MiMo-V2.6-Pro uses a sparse mixture-of-experts architecture with 1.02 trillion total parameters and 42 billion active parameters. An MoE model routes each token through a subset of specialized expert layers, reducing per-token computation while retaining a much larger pool of weights.

The active parameter count affects inference compute, while the full checkpoint still determines storage needs. At four bits per parameter, 1.02 trillion weights would occupy roughly 510 GB before metadata, higher-precision layers, runtime memory, and the key-value cache. Xiaomi’s mixed-precision build requires more, making API access considerably easier than self-hosting for most teams.

A full 99% cache discount would reduce cached input to about $0.00435 per million tokens. That rate matters for agents that repeatedly send the same system prompt, repository context, policies, or tool definitions. Actual savings depend on Xiaomi’s prefix-matching, expiration, and billing rules, which teams should verify against the current API documentation.

Artificial Analysis places the model on its intelligence-versus-cost Pareto frontier, meaning it offers one of the strongest measured capability and price combinations in the evaluated set. Production costs will also depend on output length, retries, tool-call loops, cache-hit rates, and failed runs.

One family, three serving profiles

Xiaomi is releasing V2.6 as a family with separate profiles for capability, cost, and throughput:

MiMo-V2.6-Pro supports a one-million-token context window and native multimodal input. A maximum context length describes how much the API can accept; retrieval accuracy and reasoning quality across that window still require workload-specific testing. Xiaomi also needs to document supported media formats, size limits, structured-output behavior, and tool-call compatibility for developers comparing it with existing providers.

The company describes the family as an engine for agentic workflows, where a model plans multiple steps, invokes tools, reads the results, and continues until a task is complete. Those systems place unusual pressure on instruction retention, error recovery, tool selection, and cost control because one user request can trigger hundreds of model calls.

The RL run was visible

Xiaomi exposed live logs from two reinforcement-learning post-training jobs on its own domain within two days of starting them. The dashboard showed step counts, reward curves, token throughput, sandbox availability, GPU faults, and a running cost estimate. Reinforcement-learning post-training uses scored model outputs to improve behavior after the main pretraining phase.

Each step generated 25,088 trajectories across the 1,568 prompts. The figures describe Xiaomi’s reported training spend; API inference uses the separate token rates above. Public operational logs provide useful evidence about scale and failure handling, although the underlying prompts, reward models, datasets, and complete training configuration remain necessary for reproduction.

FP4 carries the speed tier

Xiaomi created the UltraSpeed variant by applying the MXFP4 four-bit format selectively to the MoE expert layers. Its tests found that quantizing the entire model reduced reasoning and code-generation quality, so routers and other modules retain higher precision. The experts contain most of the weights and tolerate compression better, yielding substantial memory-bandwidth savings with a smaller reported quality loss.

Working with TileRT, Xiaomi reports generation above 1,000 tokens per second for the trillion-parameter model on commodity GPUs through model-system co-design. Throughput depends heavily on hardware, batch size, prompt length, output length, concurrency, and measurement method. Developers need those test conditions, along with time-to-first-token and tail-latency figures, before comparing the claim with another serving stack.

Benchmarks leave production questions

The previous MiMo-V2.5-Pro performed strongly on ClawEval, GDPVal, and SWE-bench Pro, which test agent behavior, economically valuable tasks, and software engineering. Xiaomi also reported that it could complete long-running professional tasks involving more than 1,000 tool calls. Those results describe V2.5-Pro; V2.6-Pro still needs model-specific results on the same suites and independent replication.

Earlier Flash evaluations identified several issues worth retesting:

• Long-context fidelity: The earlier 256,000-token model compressed distant context into attention sinks, which can lose information compared with full attention. V2.6’s one-million-token window requires fresh retrieval and reasoning tests.
• Training transparency: Xiaomi has withheld the datasets and detailed configuration behind its Multi-Teacher On-Policy Distillation pipeline, limiting full reproduction even when weights are available.
• Task balance: Earlier releases performed better on coding and reasoning than on creative writing and single-prompt generation.
• Agent reliability: Aggregate benchmark scores reveal little about malformed tool calls, repeated actions, recovery from tool failures, or instruction drift during long runs.

Choose by workload, then verify

A useful production evaluation should cover the behavior and costs hidden by a leaderboard score:

1. Check the checkpoint license, commercial-use terms, API availability, and supported deployment formats.
2. Run representative coding, research, browser, and tool-use tasks with the same harness used for current models.
3. Measure success rate, retries, tool-call accuracy, time to first token, output throughput, and p95 latency.
4. Test retrieval from the beginning, middle, and end of long prompts rather than relying on the advertised context limit.
5. Confirm which prompt prefixes qualify for caching and calculate costs from observed cache-hit rates.
6. Validate multimodal formats, structured outputs, streaming behavior, and failure handling.
7. Sandbox generated code and tools, especially for cybersecurity or autonomous workflows.

MiMo-V2.6-Pro’s clearest advantages are its measured generational gain, low published token prices, large context window, and serving options. Repeated-context agents stand to benefit most from caching, while self-hosted deployments face substantial memory and infrastructure requirements. Direct tests will determine whether its long-context accuracy, tool reliability, and real-world latency match the economics.