Why Does llama.cpp’s Own API Give Xiaomi’s MiMo V2.6 Flash 9x the KV Cache?
Last Updated on September 25, 2026 by Editorial Team
Author(s): Chew Loong Nian – AI ENGINEER
Originally published on Towards AI.
At 128K tokens the model needs 2.96 GiB of cache. libllama’s default context settings would allocate 27.19 GiB, llama-cpp-python’s 28.31. Two flags fix it.
If you load Xiaomi’s new MiMo V2.6 Flash through llama.cpp’s C API, you get a KV cache 9.2x larger than the model needs at 128K context; through llama-cpp-python, 9.6x.
The article explains that MiMo V2.6 Flash is designed so only 9 of its 48 attention layers use global attention; the other 39 are limited to a 128-token window, meaning the KV cache should stay small as context grows. It then shows how llama.cpp can accidentally allocate the much larger “full-size SWA cache” when callers use the C API default or when llama-cpp-python inherits those defaults, with additional memory increases caused by default flash-attention and cache padding behavior. The author provides a small script to compute expected cache sizes from the model’s config, verifies the results against llama.cpp itself, and demonstrates that setting swa_full=False (and enabling flash attention) yields the intended 128K cache footprint (~2.96 GiB instead of ~27+ GiB). Finally, it discusses the tradeoff of the full-size cache for advanced prompt-caching scenarios and concludes with practical guidance for users to ensure their binding/app exposes the correct flags.
Read the full blog for free on Medium.
Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.
Published via Towards AI
Towards AI Academy
We Build Enterprise-Grade AI. We'll Teach You to Master It Too.
15 engineers. 100,000+ students. Towards AI Academy teaches what actually survives production.
Start free — no commitment:
→ 6-Day Agentic AI Engineering Email Guide — one practical lesson per day
→ Agents Architecture Cheatsheet — 3 years of architecture decisions in 6 pages
Our courses:
→ AI Engineering Certification — 90+ lessons from project selection to deployed product. The most comprehensive practical LLM course out there.
→ Agent Engineering Course — Hands on with production agent architectures, memory, routing, and eval frameworks — built from real enterprise engagements.
→ AI for Work — Understand, evaluate, and apply AI for complex work tasks.
Note: Article content contains the views of the contributing authors and not Towards AI.
Related posts
Recent Posts
