The Speed Milestone

Xiaomi and TileRT reported sustained throughput above 1,000 tokens per second, with demos peaking near 1,200 tokens per second, on a standard 8-GPU server . The achievement breaks through what Xiaomi calls the industry's "impossible triangle" of speed, capability, and general-purpose GPU compatibility . MiMo CEO Lei Jun highlighted the milestone in a social post, describing it as the industry's first time crossing 1,000 tokens/s on a trillion-parameter model .

The UltraSpeed mode is not a new model class but an engineering-driven serving mode layered on top of MiMo-V2.5-Pro, a 1.02-trillion-parameter Mixture-of-Experts architecture with 42 billion active parameters and a 1-million-token context window .

Three Techniques Behind the Speedup

Xiaomi's official documentation describes a full-stack model-system co-design that combines three coordinated techniques to push throughput past 1,000 tokens/s .

1. FP4 Mixed-Precision Quantization

Only the MoE (Mixture of Experts) expert layers are quantized to FP4 precision, while all other layers retain their original precision . Quantization-aware training (QAT) reduces the model's memory footprint and bandwidth pressure with the goal of maintaining near-lossless quality . This selective approach avoids degrading non-expert components that are more sensitive to precision loss.

2. DFlash Speculative Decoding

DFlash replaces traditional autoregressive draft generation with block-level masked parallel prediction . The draft model uses sliding-window attention (SWA) to keep prediction cost near constant, rather than scaling with sequence length . A Muon optimizer and self-distillation are used to improve acceptance rates, directly boosting inference throughput . In coding scenarios, reports indicate an average accepted length around 6.30 tokens per verification step .

3. TileRT Persistent Kernel Engine with Warp Specialization

The TileRT system abandons the conventional per-operator kernel launch model in favor of a persistent kernel engine where the compute pipeline stays resident on the GPU . Full-pipeline prefetching overlaps data movement with computation, dramatically reducing idle GPU cycles . The system also decomposes communication, data movement, and tensor computation across different warps with dedicated roles, effectively turning the GPU into a continuously flowing, heterogeneous execution system .

Pricing: "3× the Price, 10× the Output Experience"

The UltraSpeed API trial price is set at exactly 3× the standard MiMo-V2.5-Pro output price .

• Standard tier: $0.87 per million output tokens (international) or CN¥6 per million (domestic)
• UltraSpeed trial: $2.61 per million output tokens (international) or CN¥18 per million (domestic)

Input pricing follows the same 3× multiplier, with cache-hit input at $0.0108 per million tokens and cache-miss input at $1.305 per million tokens . Xiaomi markets this as "3× the price, 10× the output experience," emphasizing the roughly 10× throughput gain for 3× the token cost .

Limited Trial Window and Access Rules

The UltraSpeed trial period is explicitly time-boxed: June 9 through June 23, 2026, at 23:59 . Access is application-based due to limited high-speed inference resources, with priority given to enterprise and professional developer use cases .

Approved users receive a free chat experience during the two-week window, subject to fairness rules: a maximum of 10 successful queue entries per account per day, a 30-minute session limit, and automatic resource release after 5 minutes of idle time . Xiaomi does not guarantee review timeliness or approval rates .

Open-Source Release

The underlying model, referred to as MiMo-V2.5-Pro-FP4-DFlash, was released as open-source alongside the UltraSpeed announcement . The FP4-quantized weights and DFlash model checkpoints are available on HuggingFace, consistent with Xiaomi's documentation identifying FP4 quantization and DFlash speculative decoding as core system components .

What This Means for Developers

The UltraSpeed mode demonstrates that trillion-parameter inference at interactive speeds can run on commodity infrastructure without custom silicon, a departure from the specialized-hardware approach seen elsewhere in the industry . For developers building latency-sensitive agentic applications, tool-calling pipelines, or real-time code generation, the combination of high throughput and a 1-million-token context window signals a practical path toward faster, more capable production systems — provided they can gain access during the limited trial window.

The Speed Milestone

Xiaomi and TileRT reported sustained throughput above 1,000 tokens per second, with demos peaking near 1,200 tokens per second, on a standard 8-GPU server . The achievement breaks through what Xiaomi calls the industry's "impossible triangle" of speed, capability, and general-purpose GPU compatibility . MiMo CEO Lei Jun highlighted the milestone in a social post, describing it as the industry's first time crossing 1,000 tokens/s on a trillion-parameter model .

The UltraSpeed mode is not a new model class but an engineering-driven serving mode layered on top of MiMo-V2.5-Pro, a 1.02-trillion-parameter Mixture-of-Experts architecture with 42 billion active parameters and a 1-million-token context window .

Three Techniques Behind the Speedup

Xiaomi's official documentation describes a full-stack model-system co-design that combines three coordinated techniques to push throughput past 1,000 tokens/s .

1. FP4 Mixed-Precision Quantization

Only the MoE (Mixture of Experts) expert layers are quantized to FP4 precision, while all other layers retain their original precision . Quantization-aware training (QAT) reduces the model's memory footprint and bandwidth pressure with the goal of maintaining near-lossless quality . This selective approach avoids degrading non-expert components that are more sensitive to precision loss.

2. DFlash Speculative Decoding

DFlash replaces traditional autoregressive draft generation with block-level masked parallel prediction . The draft model uses sliding-window attention (SWA) to keep prediction cost near constant, rather than scaling with sequence length . A Muon optimizer and self-distillation are used to improve acceptance rates, directly boosting inference throughput . In coding scenarios, reports indicate an average accepted length around 6.30 tokens per verification step .

3. TileRT Persistent Kernel Engine with Warp Specialization

The TileRT system abandons the conventional per-operator kernel launch model in favor of a persistent kernel engine where the compute pipeline stays resident on the GPU . Full-pipeline prefetching overlaps data movement with computation, dramatically reducing idle GPU cycles . The system also decomposes communication, data movement, and tensor computation across different warps with dedicated roles, effectively turning the GPU into a continuously flowing, heterogeneous execution system .

Pricing: "3× the Price, 10× the Output Experience"

The UltraSpeed API trial price is set at exactly 3× the standard MiMo-V2.5-Pro output price .

• Standard tier: $0.87 per million output tokens (international) or CN¥6 per million (domestic)
• UltraSpeed trial: $2.61 per million output tokens (international) or CN¥18 per million (domestic)

Input pricing follows the same 3× multiplier, with cache-hit input at $0.0108 per million tokens and cache-miss input at $1.305 per million tokens . Xiaomi markets this as "3× the price, 10× the output experience," emphasizing the roughly 10× throughput gain for 3× the token cost .

Limited Trial Window and Access Rules

The UltraSpeed trial period is explicitly time-boxed: June 9 through June 23, 2026, at 23:59 . Access is application-based due to limited high-speed inference resources, with priority given to enterprise and professional developer use cases .

Approved users receive a free chat experience during the two-week window, subject to fairness rules: a maximum of 10 successful queue entries per account per day, a 30-minute session limit, and automatic resource release after 5 minutes of idle time . Xiaomi does not guarantee review timeliness or approval rates .

Open-Source Release

The underlying model, referred to as MiMo-V2.5-Pro-FP4-DFlash, was released as open-source alongside the UltraSpeed announcement . The FP4-quantized weights and DFlash model checkpoints are available on HuggingFace, consistent with Xiaomi's documentation identifying FP4 quantization and DFlash speculative decoding as core system components .

What This Means for Developers

The UltraSpeed mode demonstrates that trillion-parameter inference at interactive speeds can run on commodity infrastructure without custom silicon, a departure from the specialized-hardware approach seen elsewhere in the industry . For developers building latency-sensitive agentic applications, tool-calling pipelines, or real-time code generation, the combination of high throughput and a 1-million-token context window signals a practical path toward faster, more capable production systems — provided they can gain access during the limited trial window.