Xiaomi and TileRT made a big move in June: MiMo V2.5 Pro UltraSpeed — a trillion-parameter MoE model running at 1000+ tokens/s on consumer GPUs.
This is not a lab figure or a scenario-specific peak. It is the sustained speed in general-purpose inference.
---
Key Numbers
| Metric | MiMo V2.5 Pro UltraSpeed | |--------|--------------------------| | Total parameters | 1 trillion (MoE) | | Active parameters | Not disclosed (estimated ~50-100B class) | | Inference speed | 1000+ tokens/s (consumer GPU) | | Speed gains from | FP4 quantization + DFlash speculative decoding | | Pricing | 3x the standard version | | Limited application | June 9-23 |
---
What 1000 Tokens/s Means
For comparison:
| Model / Service | Typical Speed | |-----------------|---------------| | GPT-4o (API) | ~50-100 tokens/s | | Claude 3.5 Sonnet | ~80-150 tokens/s | | Local Llama 3 70B (RTX 4090) | ~30-50 tokens/s | | MiMo V2.5 Pro UltraSpeed | 1000+ tokens/s |
1000 tokens/s means:
- A 2,000-character Chinese article generated in about 2 seconds
- Code completion that feels instantaneous
- Real-time conversation latency below human perception thresholds (<100ms)
- Standard inference uses FP16 (16-bit floating point)
- FP4 compresses weights to 4 bits — 4x memory savings
- With specialized quantization/dequantization kernels, accuracy loss stays within an acceptable range
- Traditional decoding: one token at a time
- Speculative decoding: a small "draft model" predicts multiple tokens, which the large model verifies in one pass
- DFlash is TileRT's improved variant, with a more efficient draft model and lower rejection rates
- Why did ChatGPT beat early Bard? Not because it was smarter — because it was faster
- Why does Cursor feel better than GitHub Copilot? Lower response latency
- Why is Claude Code addictive for developers? The rhythm of streaming output
- Price: 3x the standard version — unfriendly to individual developers
- Accuracy: FP4 quantization may lose accuracy on some complex reasoning tasks
- Ecosystem: MiMo's tooling and fine-tuning frameworks lag behind Llama/Qwen
- Sustainability: whether 1000 tokens/s holds during peak load remains to be seen
---
How It's Done: FP4 + DFlash
MiMo's speed isn't achieved with a "smaller model" — it comes from more aggressive inference optimization:
FP4 Quantization
DFlash Speculative Decoding
The combined effect: 3x the price for ~10x the speed.
---
"Speed Is the Product"
The release points to an emerging consensus:
> Once model capability passes a certain threshold, inference speed itself becomes the biggest product differentiator.
Consider:
As the gap in "intelligence" narrows, being fast becomes the decisive user experience.
---
The Business Logic of the Limited Application
MiMo V2.5 Pro UltraSpeed isn't openly available — it's a limited application window (June 9-23). That signals:
1. Constrained compute: FP4 saves memory, but inference still consumes massive resources 2. Market testing: how many users will pay 3x for speed? 3. Manufactured scarcity: limited time + application = higher brand premium
---
Limitations, Honestly Stated
One-Sentence Takeaway
> MiMo V2.5 Pro UltraSpeed isn't about "a smarter model" — it's about "AI is finally fast enough." When latency drops from seconds to milliseconds, AI shifts from "tool" to "environment" — like electricity, you no longer "use" it; you simply live inside it.