import { Matrix } from "@/components/matrix"

🏗️ Junze Zhiku · Agent & Model Matrix

1. Agent System Overview

Junze Zhiku runs 5 Agents, routed and distributed through Feishu Bots, each running a different model provider.

Agent	Feishu Bot	Model	Context
UltraClaw 🫡	Main assistant	deepseek-v4-pro	1M
Coder-Qwen	Coder Qwen	qwen3.7-max	1M
Coder-DeepSeek	Coder DeepSeek	deepseek-v4-pro	1M
Coder-MiniMax 🆕	Coder MiniMax	MiniMax-M3	1M
Coder-GLM	Coder GLM	glm-5.1	200K

2. Model Provider Matrix

2.1 Shared Capabilities (Possessed by All Models)

🇨🇳 Native Chinese

All models have native support for Traditional Chinese and do not need a translation layer. DeepSeek / Qwen / MiniMax / GLM are all developed by Chinese teams.

🔌 OpenAI Compatible

All use an OpenAI-compatible API (`/v1/chat/completions`), so the same codebase can switch between them seamlessly.

🛠️ Tool Calling

All support function calling / tool use and work normally within the OpenClaw Agent framework.

🧠 Deep Thinking

All support thinking / reasoning modes (DeepSeek: thinking_max, Qwen: thinking, MiniMax: adaptive, GLM: enabled).

💰 Pay-as-You-Go

Except for the OMLX local model, all are billed per token with no monthly lock-in.

🌐 Available in Hong Kong

All API endpoints are directly accessible from Hong Kong without a VPN.

2.2 Core Differences (Common Ground, Distinct Strengths)

Dimension	DeepSeek V4 Pro	Qwen 3.7 Max	MiniMax M3	GLM 5.1
Context Window	1,000,000	1,000,000	1,000,000	200,000
Max Output	384,000	65,536	131,072	128,000
Parameter Scale	1.6T / 49B active	MoE (undisclosed)	MoE (undisclosed)	Undisclosed
Open Source	✅ MIT	❌ Closed source	❌ Closed source	✅ MIT
API Protocol	OpenAI + Anthropic	OpenAI Compat	OpenAI Compat	OpenAI Compat
Multimodal	Text	Text + images + video	Text	Text
Pricing Highlights	Long-context cost only 27% of V3	Low price within 256K	High cost-performance ratio	Coding Plan billed separately
Ecosystem Maturity	⭐⭐⭐⭐⭐	⭐⭐⭐⭐	⭐⭐⭐	⭐⭐⭐

2.3 Three Shared Advantages

🥇 A Million-Token Context for Everyone

DeepSeek V4, Qwen 3.6+, and MiniMax M3 all have a 1,000,000 token context window. This means a codebase or document containing 750,000 Chinese characters can be loaded in one pass without chunking. GLM-5.1's 200K is already sufficient for ordinary tasks and falls slightly short only when analyzing very large codebases.

🥈 All Suited to Agent Workflows

All four models have tool calling / function calling capability and can reliably execute multi-step tasks within OpenClaw's Agent framework. DeepSeek has the most mature Agent ecosystem (native support in Claude Code / OpenClaw), MiniMax M3 has the most cutting-edge Agentic Reasoning design, and Qwen has the best Agent stability in Chinese-language scenarios.

🥉 Native Chinese Across the Board

All are models developed by Chinese teams, with native-level support for Traditional Chinese. There is no need to go through an English translation layer; using Chinese prompts directly yields the best results. In scenarios such as handling Hong Kong company secretary documents, legal provisions, and Traditional Chinese financial reports, they all outperform Western models such as GPT/Claude.

2.4 The Real Differences

Difference	Explanation
Anthropic API	Only DeepSeek supports both the OpenAI and Anthropic protocols. It is more stable when Claude Code connects natively.
Multimodal Input	Only Qwen supports image and video input. The other three are text only.
Open Source Licensing	DeepSeek V4 (MIT) and GLM-5.1 (MIT) are open source and can be deployed locally. Qwen and MiniMax are closed source.
Output Ceiling	DeepSeek's 384K output dominates the field. Qwen's 65K output is the lowest. When large volumes of code must be generated, DeepSeek has a clear advantage.
Billing Model	DeepSeek/MiniMax are pay-as-you-go. Bailian Token Plan is a subscription model. GLM Coding Plan is billed separately. The choice depends on your usage pattern.
Local Deployment	DeepSeek V4 and GLM-5.1 are open source and can be deployed locally (requires substantial GPU resources). OMLX already runs Qwen3-Coder-Next (8bit quantized, 32K context, completely free).

3. Provider Panorama

Provider	Number of Models	Context Range	Endpoint	Status
DeepSeek	4	1M	api.deepseek.com	✅
Qwen (Bailian)	9	262K ~ 1M	DashScope International	✅
Bailian Token Plan	5	1M	Alibaba Cloud MaaS	✅
MiniMax	3	1M	api.minimax.io	✅
z.ai (Zhipu)	4	200K	api.z.ai (Coding)	✅
MaaS (Alibaba Cloud)	15	1M	Alibaba Cloud MaaS	⚠️ Pending activation
OMLX (local)	1	32K	Mac Studio	✅ Free

4. Scenario Selection Recommendations

Everyday Conversation & Business Consulting

🥇 UltraClaw (DeepSeek V4), full features + memory system + all skills

🥈 Any coder agent will do, depending on which model you want to use

Hong Kong Stock Research & Reports

🥇 UltraClaw (DeepSeek V4), the AK-SDD skill lives here

🥈 Coder-Qwen, the most natural expression for Chinese reports

Frontend UI & Full-Stack Development

🥇 Coder-Qwen, the strongest Chinese UI generation and the best frontend aesthetics

🥈 Coder-DeepSeek, more stable reasoning when the logic is complex

Backend & Algorithms & APIs

🥇 Coder-DeepSeek, the strongest reasoning ability + 384K output

🥈 Coder-MiniMax, cutting-edge Agentic reasoning design

Large Codebase Analysis

🥇 Coder-MiniMax (M3), 1M context + 131K output

🥈 Coder-DeepSeek, 1M context + 384K output, better suited to generation

High-Frequency Batch Coding

🥇 Coder-GLM, Coding Plan billed separately, cost isolation

🥈 OMLX Local, completely free, suited to non-critical tasks

5. Routing Architecture

Boss → Feishu
        ├─ Main Bot ──→ UltraClaw (General Manager · DeepSeek V4)
        ├─ Qwen Bot ──→ Coder-Qwen (Frontend/Full-stack)
        ├─ DeepSeek Bot ──→ Coder-DeepSeek (Backend/Algorithms)
        ├─ MiniMax Bot ──→ Coder-MiniMax (Large Codebase)
        └─ GLM Bot ──→ Coder-GLM (Backup · Coding Plan)

Each Agent runs independently, and you switch models by choosing a different Feishu Bot. There is no need for a /model command; you simply talk to a different one.

Junze Zhiku Agent Matrix · 2026-06-04 · UltraClaw 🫡

Junze Zhiku Agent & Model Matrix