Supported Models
LingXizhi IDE natively supports 15+ mainstream models including Kimi K3, GLM-5.2, DeepSeek-V4, Qwen3.6, Doubao and MiniMax, via official direct and Volcano Engine channels. The model list is pulled from the cloud by the kernel in real time — newly released models are available without upgrading the IDE.
The IDE is the source of truth for available models: open "Settings → Models"; the dropdown lists every model available to your account. The table below lists representative models enabled by default and their key parameters.
Model List
| Model | Vendor | Tier | Context | Vision | Highlights |
|---|---|---|---|---|---|
kimi-k3 | Kimi (Moonshot AI) | Flagship | 1M | ✓ | 2.8T parameters, always-on thinking, adjustable thinking depth via reasoning_effort (low / high / max), automatic prefix caching |
kimi-k2.7-code | Kimi (Moonshot AI) | Standard | 256K | ✓ | Coding-focused model, thinking always on, image / video input, built for Agent tasks |
glm-5.2 | Zhipu GLM | Flagship | 1M | — | IndexShare index-reuse architecture, five thinking levels (Low / Medium / High / Extra / Max), native MCP tool calling |
deepseek-v4-pro | DeepSeek | Flagship | 1M | — | 1.6T-parameter MoE, thinking mode by default, open-source SOTA for Agentic Coding |
qwen3.8-max-preview | Qwen | Flagship | 1M | — | Thinking-only mode (cannot be disabled), strongest reasoning, structured output |
qwen3.7-max | Qwen | Flagship | 1M | — | Hybrid thinking mode (on by default), built-in tools |
qwen3.7-plus | Qwen | Standard | 1M | ✓ | Hybrid thinking mode, high cost-performance, structured output |
minimax-m3 | MiniMax | Flagship | 1M | ✓ | MSA sparse attention, native multimodal (image / video), max output 512K |
minimax-m2.7 | MiniMax | Standard | 200K | — | 229B-parameter MoE (10B active), thinking mode enforced |
Thinking modes vary by model: Kimi K3 keeps thinking always on, with reasoning_effort controlling the depth; GLM-5.2's deep thinking requires thinking enabled with clear_thinking: false. The kernel adapts these parameters automatically per model — no manual work needed after switching models.
Access Channels
The same model can be served by multiple channels; the IDE picks an available one based on configuration:
| Channel | Type | Description |
|---|---|---|
| Official direct | api | Official APIs of Zhipu GLM, DeepSeek, Alibaba Qwen, Moonshot AI Kimi and MiniMax |
| Volcano Engine | api | ByteDance Volcano Engine aggregation channel; all mainstream models above can be accessed through it |
| Coding Plan / Agent | plan / agent | Plan-based channels such as Zhipu Plan and Volcano Engine Plan; require the corresponding vendor's subscription |
Aggregation channels such as SiliconFlow are also integrated. Channel and model availability is managed centrally on the server side — the model picker in the IDE is the source of truth for actual combinations.
Switching Models
- Open "Settings → Models".
- Pick a model in the "Default model" dropdown — the choice applies to AI chat, inline edit and RUSH tasks at the same time.
- If the dropdown is empty, the network or the kernel is usually not ready yet — wait a moment or restart the IDE.
Bigger context is not always better: very long context increases the latency of each request and (with BYOK) the vendor cost. Use standard-tier models for everyday chat, and switch to flagship models for RUSH deep tasks.
Next Steps
- Learn about built-in plan quotas and the daily reset rule: Plans & Quota.
- Already have a vendor key? Bind it to get billed by the vendor with zero platform markup: BYOK.
- Use thinking mode and multimodal capabilities in chat: AI Chat Guide.