nsheaps/glm-models

GLM Model Family

The GLM (General Language Model) family is developed by z.ai (formerly Zhipu AI / 智谱AI). These models support text generation, vision, code, embeddings, image generation, and video generation. All recent models are open-weight under MIT license.

• Docs: https://docs.z.ai/
• API Base URL: https://api.z.ai/api/paas/v4/
• Pricing: https://docs.z.ai/guides/overview/pricing

Flagship Text Models

| Model | Architecture | Context | Key Features | | --------------- | ---------------------- | ------------------ | ------------------------------------------------------------- | | glm-5 | ~745B MoE (44B active) | 200K in / 128K out | Agentic engineering, tool streaming, long-horizon tasks, MIT | | glm-5-turbo | Same, optimized | 200K in / 128K out | Improved stability for long-chain agent tasks | | glm-4.7 | ~400B MoE | 200K in / 128K out | Coding-focused, Preserved Thinking, Turn-level Thinking, MIT | | glm-4.7-flash | Lightweight | Reduced | Free tier, lighter capability | | glm-4.6 | 355B total | 200K | Strong code benchmarks, agent frameworks, MIT | | glm-4.5 | 355B / 32B active | 128K | Hybrid reasoning (thinking/non-thinking modes), deep thinking | | glm-4.5-x | Premium tier | 128K | Higher capability, premium pricing | | glm-4.5-air | 106B / 12B active | 128K | Compact variant of GLM-4.5 | | glm-4.5-flash | Lightweight | 128K | Free tier |

Thinking Mode

GLM-4.5+ models support hybrid reasoning — toggle between deep thinking and instant response:

{
  "model": "glm-4.7",
  "messages": [{ "role": "user", "content": "Solve this step by step" }],
  "thinking": { "type": "enabled" }
}

• Preserved Thinking (GLM-4.7): Retains thinking blocks across multi-turn conversations
• Turn-level Thinking (GLM-4.7): Per-turn control — disable for lightweight requests, enable for complex tasks
• Tool Streaming (GLM-5): Stream output during tool calling (tool_stream: true)

Vision / Multimodal Models

| Model | Parameters | Context | Description | | ---------------- | ----------------- | ------- | -------------------------------------- | | glm-4.6v | 106B / 12B active | 128K | Vision understanding, function calling | | glm-4.6v-flash | 9B | — | Free, open weights, commercial license | | glm-4.5v | 106B VLM | — | Vision-language model |

Vision API Example

curl "https://api.z.ai/api/paas/v4/chat/completions" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-4.6v",
    "messages": [{
      "role": "user",
      "content": [
        {"type": "text", "text": "What is in this image?"},
        {"type": "image_url", "image_url": {"url": "https://example.com/image.jpg"}}
      ]
    }]
  }'

Specialized Models

| Model | Category | Description | | ----------------- | ---------------- | -------------------------- | | glm-image | Image generation | Text-to-image (Jan 2026) | | glm-ocr | OCR | Document and image OCR | | cogview-3-plus | Image gen | High-quality text-to-image | | cogvideox | Video gen | Text-to-video generation | | cogvideox-flash | Video gen | Fast video generation |

Embedding Models

| Model | Dimensions | Description | | ------------- | ---------- | ------------------------------- | | embedding-3 | 2048 | General-purpose text embeddings | | embedding-2 | 1024 | Previous generation embeddings |

curl "https://api.z.ai/api/paas/v4/embeddings" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "embedding-3",
    "input": "What is machine learning?"
  }'

Model Selection Guide

| Use Case | Recommended Model | Why | | ------------------- | ----------------- | --------------------------------------- | | Agentic tasks | glm-5 | Tool streaming, long-horizon planning | | Coding | glm-4.7 | Coding-focused, Preserved Thinking | | Complex reasoning | glm-4.5 | Hybrid reasoning with deep thinking | | General chat | glm-4.5-flash | Free, good quality | | High throughput | glm-4.5-air | Compact, fast inference | | Image understanding | glm-4.6v | Best vision model with function calling | | Embeddings/search | embedding-3 | Latest generation | | Image creation | glm-image | Latest generation (Jan 2026) | | Budget-conscious | glm-4.5-flash | Free tier available |

Claude Code Model Mapping

When using z.ai's Anthropic-compatible endpoint with Claude Code, map models to slots:

| Claude Code Slot | Recommended GLM Model | Rationale | | ---------------- | --------------------- | ---------------------------- | | Opus | glm-5 | Most capable, agentic | | Sonnet | glm-4.7 | Strong coding, balanced cost | | Haiku | glm-4.5-air | Fast, cost-effective |

Pricing (per 1M tokens, USD)

| Model | Input | Output | | ---------------- | ------ | ------ | | glm-5 | ~$1.00 | ~$3.20 | | glm-4.7 | $0.60 | $2.20 | | glm-4.7-flash | Free | Free | | glm-4.5 | ~$0.20 | ~$1.10 | | glm-4.5-x | — | $8.90 | | glm-4.5-flash | Free | Free | | glm-4.6v | ~$0.14 | ~$0.41 | | glm-4.6v-flash | Free | Free |

Prices approximate; see docs.z.ai/guides/overview/pricing for current rates. Batch API available at 50% cost.

Unique Features

• MIT license: GLM-4.5, 4.6, 4.7, and 5 are all open-weight under MIT
• 200K context: GLM-4.6, 4.7, and 5 support 200K input with up to 128K output
• Hybrid reasoning: Toggle deep thinking on/off per request or per turn
• Tool streaming: GLM-5 streams output during tool calls for real-time agent UX
• Free tiers: glm-4.5-flash, glm-4.7-flash, glm-4.6v-flash are free
• Domestic chip training: GLM-5 trained on Huawei Ascend chips, GLM-4.6 on Cambricon — zero NVIDIA dependency
• Bilingual strength: Particularly strong in Chinese + English tasks
• Anthropic-compatible API: Native Claude Code integration without proxies
• Native function calling: OpenAI-style tool description format in all recent models

References

• z.ai Developer Docs
• GLM-5 Overview
• GLM-4.7 Docs
• GLM-4.6 Docs
• GLM-4.5 Docs
• Pricing
• z.ai on GitHub
• z.ai on Hugging Face