MiMo-V2.6-Pro.
9.0× faster output.

PrimaLabs serverless delivers 407 output tokens per second on MiMo-V2.6-Pro at 10,000 input tokens. In the 24 September Artificial Analysis snapshot, Xiaomi’s public endpoint delivers 45 tokens per second and DeepInfra delivers 25. PrimaLabs leads the three-provider comparison at 9.0× Xiaomi’s output speed.

The lead holds across all four tested workloads: 1k, 10k, and 100k input tokens, and 1k input with ten parallel queries.

For applications generating code, documents, and multi-step agent responses, faster token generation means less time spent producing each answer. PrimaLabs brings that speed to a serverless API with zero data retention (ZDR): your prompts and outputs are not retained.

MiMo-V2.6-Pro scores 46 on the Artificial Analysis Intelligence Index. Xiaomi’s published model card describes a mixture-of-experts architecture with 1.02 trillion total parameters, 42 billion active parameters, a one-million-token context window, and support for text, images, video, and audio input. The weights are released under the MIT license.

Artificial Analysis model overview · Xiaomi model card

For teams building agents, that combination offers capability, room for long working contexts, and flexibility over deployment. Serving performance determines how much waiting those workflows introduce.

MiMo-V2.6-Pro and Grok 4.7 (xhigh) both score 46 on the Artificial Analysis Intelligence Index. Gemini 3.8 Flash (high), another closed model in the nearby score range, scores 41. PrimaLabs’ serverless pricing shows how competitive the economics of an open model can be.

Pricing and Intelligence Index snapshot: 23 September 2026. Prices are USD per million tokens. The MiMo row uses PrimaLabs’ serverless rates from the live Artificial Analysis provider page. The input rate is displayed rounded to two decimal places.

Its listed input-token rate is also lower than both. Those are token-price differences; application savings depend on the tokens each model needs to complete the work.

Artificial Analysis also lists PrimaLabs at $0.18 per million tokens blended, using its 7:2:1 mix of cached input, uncached input, and output tokens. The comparison table above keeps input and output prices separate so readers can apply their own workload mix. View provider pricing.

Scores: Artificial Analysis Intelligence Index v4.3.2 ( MiMo overview and closed-model pages linked above). Pricing: PrimaLabs provider listing, Grok pricing, and Google’s standard Gemini pricing. Gemini rates shown apply through 31 December 2026. The comparison uses uncached input and standard output rates, without batch discounts or tool charges. Similar aggregate scores do not guarantee equivalent results on every task; token usage, reasoning, and caching affect the cost of completing a task.

At 10,000 input tokens, PrimaLabs leads both Xiaomi and DeepInfra on output speed. The same comparison now covers three providers, with PrimaLabs first in every workload shown.

The advantage extends beyond the headline workload. PrimaLabs reaches 357 tokens per second at 1k input, 327 at 100k input, and 287 in the ten-query concurrency test.

Select any benchmark chart to enlarge it.

Output tokens per second; higher is better. Values are transcribed from the Artificial Analysis charts dated 24 September 2026. Multipliers are calculated from the displayed rounded values. The ten-query figure is the reported output-speed metric, not aggregate fleet throughput.

Across the four workloads, PrimaLabs’ output-speed advantage over Xiaomi ranges from 2.4× to 9.4×. The headline uses the 10k-input result: 9.0×.

The speed advantage carries through to total response time. For a 500-token answer with 10,000 input tokens, Artificial Analysis reports a calculated end-to-end time of 6.8 seconds on PrimaLabs, compared with 58.2 seconds on Xiaomi and 102.9 seconds on DeepInfra. That is an 8.6× end-to-end speedup over Xiaomi.

Answer generation accounts for approximately 1.2 seconds on PrimaLabs versus 11.0 seconds on Xiaomi. Reasoning time also falls, from 44.1 seconds on Xiaomi to 4.9 seconds on PrimaLabs.

Approximate answer-generation time at 10,000 input tokens. Excludes input processing and reasoning.

The first answer token arrives in 5.6 seconds on PrimaLabs, versus 47.2 seconds on Xiaomi and 82.5 seconds on DeepInfra. This metric includes the reasoning that happens before the answer begins.

Seconds; lower is better. Figures follow the 24 September charts. Rounding can cause individual components to differ slightly from displayed totals.

The chart below breaks that number down: on PrimaLabs, almost all of the 5.6 seconds is reasoning (4.9 s), while Xiaomi spends 44.1 seconds and DeepInfra 81.8 seconds thinking before the answer begins.

Faster token generation means code, documents, and responses arrive sooner. For agents that make dependent calls, the benefit adds up: each shorter response lets the next model-dependent step begin earlier.

Consider 20 sequential calls, each with 10,000 input tokens and a 500-token answer. Applying the September 24 end-to-end benchmark times, those calls would total about 2.3 minutes on PrimaLabs versus 19.4 minutes on Xiaomi—roughly 17 minutes less waiting for model responses.

Illustration: 20 × 6.8 seconds on PrimaLabs and 20 × 58.2 seconds on Xiaomi. This is model-response time, not a measured agent run. Actual completion times also depend on tools, retries, context lengths, and opportunities to run work in parallel.

The advantage also holds in the ten-query concurrency test: 287 tokens per second on PrimaLabs versus 51 on Xiaomi, or 5.6× the output speed. Across both sequential work and the tested concurrent workload, faster serving reduces the time applications spend waiting on the model.

Speed of implementation is a core PrimaLabs advantage. Our team moved rapidly from MiMo-V2.6-Pro’s release through integration, inference optimization, and independent benchmarking—putting new model capabilities into developers’ hands sooner.

PrimaLabs builds a runtime that tunes inference around the application’s workload and available hardware. It measures how context length, concurrency, caching, and service requirements interact, then evaluates changes across batching, scheduling, memory movement, and GPU execution.

For this endpoint, the configuration included continuous batching, chunked prefill, and an admission policy tuned to the workload. The benchmark measures the resulting endpoint performance; it does not isolate the contribution of each optimization.

The same approach can be applied to customer workloads: measure the relevant traffic, evaluate configurations against quality and service requirements, and deploy the configuration that meets those requirements.

We see this as a new era for open-model economics: an openly available model can match a major closed model’s aggregate intelligence score while carrying a substantially lower public API token price. For builders, that expands the set of workloads worth evaluating—from longer coding sessions to document pipelines and agents that make repeated model calls.

Open weights also let the same model compete across serving providers. Teams can evaluate inference speed, price, data-retention policies, and deployment options independently of the company that trained it. The capability comes from the model; the serving layer determines how efficiently an application can use it.

PrimaLabs brings those economics together with measurable output speed: 407 tokens per second at 10,000 input tokens, delivered through a serverless API with zero data retention. Our view is that the next phase of adoption will reward the combination of model quality, affordable inference, and fast execution. PrimaLabs’ published pricing, zero data retention, and this serving result show why that combination deserves a place in production evaluations.

Methodology: Performance figures and the four embedded charts use the Artificial Analysis snapshot dated 24 September 2026. The headline compares output speed at 10,000 input tokens: 407 tokens/s for PrimaLabs, 45 for Xiaomi, and 25 for DeepInfra. Other tests cover 1k and 100k input tokens and 1k input with ten parallel queries. Output speed is tokens per second during generation, not aggregate fleet throughput. Time to first answer token includes reasoning time. Artificial Analysis calculates end-to-end time for a 500-token answer from input latency, reasoning time, and output speed. Ratios in this article use the displayed rounded values. Pricing and Intelligence Index comparisons retain the 23 September snapshot identified above. Results compare the tested endpoints and do not establish equal hardware, equal cost, or performance on every workload. Live figures can change. See the Artificial Analysis provider benchmarks and metric definitions.