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.
Not lab data, not peak speeds in cherry-picked scenarios. This is sustained speed in general-purpose inference.
---
Core 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 gain source | FP4 quantization + DFlash speculative decoding | | Pricing | 3x the standard version | | Limited application window | June 9–23 |
---
What Does 1000 tokens/s Mean?
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 2000-character Chinese article generated in 2 seconds
- Code completion that feels instantaneous
- Real-time conversation latency below human perception threshold (<100ms)
- Standard inference uses FP16 (16-bit floating point)
- FP4 compresses weights to 4 bits — a 4x memory saving
- With dedicated quantization-dequantization kernels, accuracy loss stays within an acceptable range
- Traditional decoding: generate 1 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 version, with a more efficient draft model and lower rejection rate
- Why was ChatGPT more popular than 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 precision on some complex reasoning tasks
- Ecosystem: MiMo's ecosystem (tooling, fine-tuning frameworks) is less mature than Llama/Qwen
- Sustainability: whether 1000 tokens/s can be maintained during peak load remains to be verified
---
How It's Done: FP4 + DFlash
MiMo's speed isn't achieved by using a "smaller model," but by more aggressive inference optimization:
FP4 Quantization
DFlash Speculative Decoding
The combined effect: 3x the price for 10x the speed.
---
"Speed Is the Product"
MiMo's release reveals an emerging consensus:
> Once model capability passes a certain threshold, inference speed itself becomes the biggest product differentiator.
Consider:
As the gap in "how smart" models are narrows, "fast" becomes the decisive user experience.
---
The Business Logic of the Limited Application Window
MiMo V2.5 Pro UltraSpeed isn't directly open — instead it uses a limited application window (June 9–23).
This signals several things: 1. Limited compute: FP4 saves memory, but inference still consumes massive compute 2. Market testing: how many users will pay 3x for speed? 3. Manufactured scarcity: limited time + application = higher brand premium
---
Limitations and Honesty
One-Line Summary
> MiMo V2.5 Pro UltraSpeed isn't about "a smarter model" — it's about "AI finally being fast enough." When latency drops from "seconds" to "milliseconds," AI goes from "tool" to "environment" — like electricity, you no longer "use" it; you simply live inside it.