From Niche Metric to Front-Line Differentiator
The fastest AI models are systems that can generate or process large numbers of tokens per second, and their GPU inference speed has shifted from a niche benchmark to a core measure of AI model performance that decides which tools feel instant and which feel slow in real-world use. For years, speed sat behind accuracy and model size in marketing decks; now, laggy responses mean lost users and lower productivity. Data from the Artificial Analysis index, quoted by OfficeChai, shows leading models spanning a wide range from 134 to 306 tokens per second, and “this difference adds up fast” when workflows hit AI hundreds of times a day. As a result, speed-per-dollar, not only quality-per-token, is shaping model choice. Developers and enterprises are building evaluation checklists where throughput, latency, and concurrency sit alongside reasoning benchmarks.
Xiaomi MiMo-V2.5-Pro and the 1,000 Tokens-Per-Second Breakthrough
Xiaomi’s MiMo-V2.5-Pro in UltraSpeed Mode highlights how quickly the ceiling for GPU inference speed is rising. Co-designed with TileRT, the 1-trillion-parameter model reaches over 1,000 tokens per second on general-purpose GPUs, far beyond the 150 tokens per second reported for MiMo-V2-Flash in late 2025. Xiaomi describes this as the result of “ultimate co-design” between the model and its underlying system, and claims roughly 10 times faster output compared with standard MiMo-V2.5-Pro API access. The UltraSpeed tier comes with a 3x price increase over the regular API and is limited to an application-based trial window, underscoring how scarce high-speed inference resources still are. Even with limits on queue entries, session length, and idle time, the experiment signals a new era: hitting four-digit tokens per second on standard GPUs is no longer hypothetical, but a commercial product direction.
How Speed Changes What AI Can Do in Practice
Once AI models cross certain tokens-per-second thresholds, new applications become practical rather than aspirational. At around 150 tokens per second, as with Xiaomi’s earlier MiMo-V2-Flash, an assistant can output text much faster than a person can read or speak, removing the sense of waiting for the model. Pushing toward and beyond 1,000 tokens per second enables richer scenarios: summarizing large documents in real time, powering interactive coding tools that feel as responsive as a local editor, or driving conversational agents that can consume long histories without a noticeable pause. The fastest AI models from the Artificial Analysis index, such as GPT-oss 120B at 306 tokens per second, show that even without experimental modes, high throughput is already reachable. Higher speed also means lower latency for small replies and higher throughput for bulk workloads, both critical for customer-facing products and internal automation.
Standard GPUs, Not Exotic Hardware, Power the New Wave
A striking part of Xiaomi’s MiMo-V2.5-Pro UltraSpeed story is that the model hits over 1,000 tokens per second on general-purpose GPUs rather than exotic, custom accelerators. That aligns with a broader trend in the fastest AI models rankings, where companies compete not only on raw model size but on how well they squeeze performance from common GPU infrastructure. OpenAI’s GPT-oss 120B reaching 306 tokens per second on its high-compute tier, and NVIDIA’s Nemotron 3 Super hitting 153 tokens per second, both signal that software optimizations and scheduling strategies now matter as much as hardware selection. As these techniques mature, high GPU inference speed becomes accessible beyond a few hyperscale platforms. Smaller teams can rent standard GPUs yet still aim for hundreds of tokens per second, narrowing the gap between experimental labs and mainstream developers.
Speed Benchmarks Are Redefining Model Selection
Speed benchmarks are reshaping how developers and enterprises evaluate AI model performance. Leaderboards that once focused on reasoning and coding scores now prominently display tokens per second, with GPT-oss, Gemini, Qwen, Grok, Mistral, and others compared side by side on throughput as much as capability. The spread from 134 to 306 tokens per second in the Artificial Analysis chart illustrates why: in high-volume workflows, a 2x speed difference can halve infrastructure usage for the same output or enable far richer experiences within the same time budget. Xiaomi’s decision to price UltraSpeed at 3x the standard MiMo-V2.5-Pro API while promising a “10x output experience” shows how speed itself is becoming a product feature. As more providers roll out fast and ultra-fast tiers, model choice will often start with a simple question: does it feel instant enough for the job?
