Local LLM Selection Guide: KAIHE Full Lineup Performance Testing & Recommendations

Published on: 2026-05-02

What's the hardest part about choosing AI hardware?

It's not budget. It's "What level of models can I actually run with this money?"

Today, let's break down the KAIHE AI Box lineup with real-world testing data — from the ¥999 A1 to the enterprise-grade G1.


Quick Reference Table

Model Compute Max Local Model Cloud Support Price Best For
A1 6 TOPS 2B params Full API suite ¥999 Personal/Trial
B1 10 TOPS 4B params Full API suite ¥1,699 Power users
C1 20 TOPS 7B params Full API suite ¥2,999 Small teams
E1 60 TOPS 13B params Optional ¥8,999 SMEs
G1 1000 TOPS 70B params Optional ¥34,999 Enterprise/R&D

Deep Dive: Real Experience at Each Tier

A1 (¥999): Gateway to the AI Ecosystem

6 TOPS might sound "limited", but here's the truth: A1 runs 2B models like Gemma smoothly for specific tasks (translation, summarization, basic conversation).

The real value: A1 supports all major cloud LLM APIs — MoonShot, Doubao, GPT, Claude. Configure your API keys in the panel, and A1 becomes your connected AI workstation. Local lightweight models for quick tasks, switch to cloud for deep reasoning.

Bottom line: ¥999 buys you the gateway to the entire AI ecosystem, not just an edge computing box.


B1 (¥1,699): Sweet Spot for Power Users

10 TOPS, up to 4B local models. This tier hits the sweet spot — 4B models are genuinely useful for writing assistance, information organization, and code completion. Nearly double the local inference speed of A1, smooth multi-agent execution.


C1 (¥2,999): Small Team Content Hub

20 TOPS, 7B local. The 7B class is where the open-source community truly thrives — Qwen2.5-7B, Mistral-7B, Llama-3.1-8B are all production-proven models. Ideal for small creator teams sharing one AI workstation, content creators needing AI-assisted writing/translation, and small businesses using AI for daily emails and documents.


E1 (¥8,999): True Private Deployment

60 TOPS, 13B local. At this tier, "local private deployment" becomes genuinely meaningful. For organizations with confidential data that absolutely cannot go to the cloud, teams needing custom fine-tuned models running locally, or serving 10-20 employees for daily AI assistance. 13B models approach GPT-4 level performance on most professional tasks.


G1 (¥34,999): Full-Scale LLM Fortress

1000 TOPS, 70B. Runs full-scale open-source models like Qwen-72B and DeepSeek-70B. Multiple concurrent 70B instances or blazing-fast single-model performance. Built for R&D centers, financial/medical institutions with extreme data security requirements, and lab environments needing AI cluster processing.


Selection Decision Tree

Tight budget, want to experiment → A1. Zero decision pressure at ¥999.

Heavy personal user → B1. The extra compute is noticeable in daily high-frequency use.

Small team sharing → C1. No queuing with concurrent multi-user calls.

Confidential data → E1 minimum. Only 13B+ local models deliver true enterprise-grade usability.

Extreme performance/compliance → G1. This is infrastructure investment, not a cost.


Final Words

KAIHE AI Box has a clear position: it's an AI computer, not an AI toy.

From A1 to G1, every model has a defined scenario and user base. No model is "not worth buying" — only "which one fits you best."

If you're still hesitating, remember this principle: buy the A1 and start using it. The value of AI tools is discovered through use, not calculated through spec comparisons.

© KAIHE AI - Agent Computer Specialist