Xiaomi MiMo Price Cut Technical Optimizations | Generated by AI
Question: What is this Xiaomi MiMo announcement saying, and what are the key technical optimizations behind the price reduction?
Answer:
Summary
Xiaomi just sent a developer newsletter announcing massive price cuts for their MiMo-V2.5 reasoning model API, effective today (2026-05-27 Beijing time). The headline: up to 99% price reduction.
Pricing Breakdown
| Model | Cache Hit Input | Cache Miss Input | Output |
|---|---|---|---|
| MiMo-V2.5-Pro | ¥0.025/M tokens | ¥3/M tokens | ¥6/M tokens |
| MiMo-V2.5 | ¥0.02/M tokens | ¥1/M tokens | ¥2/M tokens |
| MiMo-V2.5-TTS | Free (limited time) | — | — |
Compare: MiMo-V2.5 output at ¥2/M ≈ $0.28/M tokens — cheaper than GPT-4o-mini output ($0.60/M), aggressive pricing in the Chinese LLM API market.
The huge spread between cache-hit (¥0.02) vs cache-miss (¥1) input pricing incentivizes you to structure prompts with long stable system prompts to maximize cache hits — same pattern as Anthropic’s prompt caching.
The Technical Story (Most Interesting Part)
The price cut is backed by real engineering. Let me break down what they actually did:
1. SGLang + HiCache with Sliding Window Attention (SWA)
SWA (used in Mistral-style architectures) means each token only attends to a fixed window of recent tokens, not the full context. This limits KV cache size per layer:
Standard attention: KV cache grows as O(seq_len × d_model) per layer
SWA: KV cache capped at O(window_size × d_model) per layer
HiCache is SGLang’s hierarchical KV cache system — it tiered storage across:
- GPU HBM (fast, expensive, small)
- CPU RAM (slower, cheap, large)
- SSD (slowest, very cheap, very large)
Their claim: data movement reduced to 1/7 of before, cacheable tokens increased 5x. This is the core reason cache-hit pricing is 50x cheaper than cache-miss — serving a cached token is nearly free.
2. Expert Parallelism Optimization
MiMo is a MoE (Mixture of Experts) model (based on their architecture). Expert parallelism means different GPUs handle different experts. The optimization likely involves:
- Better load balancing across experts
- Reducing expert routing communication overhead
- Possibly fused dispatch/combine kernels
3. Input Length Bucketing
Instead of padding every sequence to max length, they bucket inputs by length range:
# Conceptually:
buckets = [512, 1024, 2048, 4096, 8192]
# Batch requests within same bucket → less padding waste → higher GPU utilization
This improves throughput per GPU, directly reducing cost per token.
Token Plan Reset (Business Side)
Existing paid users get their consumed credits reset to zero — effectively a free refill. Since V2.5 tokens go 5-8x further than V2 at these new prices, existing plans are now worth much more. Smart retention move.
What This Means for You
If you’re building anything with Chinese LLM APIs (for bank work, agents, CLI tools), MiMo-V2.5 is now worth benchmarking against:
- DeepSeek-V3 (¥2/M output, similar tier)
- Qwen-Plus
- Your current Claude/GPT-4o usage for non-English tasks
The SWA + HiCache architecture insight is also directly applicable to your own inference work — if you’re running SGLang on your RTX 4070 or AMD MI300X, these are the knobs to understand.