GLM 5
Zhipu AI's 744B-parameter open-weight MoE flagship built for agentic engineering, reasoning, and long-horizon coding tasks.
Overview
What is GLM 5
GLM 5 is Zhipu AI's open-weight flagship LLM, a 744B-parameter Mixture-of-Experts model designed for agentic engineering, advanced reasoning, and end-to-end software engineering. Released in February 2026, it integrates DeepSeek Sparse Attention and asynchronous reinforcement learning to deliver frontier-class coding and tool-use performance with open-source sovereignty.
Running it privately on Venice
On Venice, GLM 5 runs under a private zero-retention tier — your prompts are not stored, profiled, or used to train closed models. You get the full open-weight model with native tool use, reasoning, web search, and structured output, preserving data sovereignty without Big-Tech surveillance.
Assessment
Strengths and limitations
- Open-source weights enable self-hosting and full model sovereignty, with 744B parameters routed efficiently via 40B active experts.
- State-of-the-art agentic coding and reasoning among open models, delivering strong end-to-end software engineering performance.
- Native tool use / function calling, reasoning, web search, and structured JSON output for autonomous agentic workflows.
- Competitive with closed frontier models on coding benchmarks while costing a fraction of the price.
- DeepSeek Sparse Attention reduces inference cost while preserving long-context fidelity.
- Self-hosting requires massive GPU resources at full precision, making local inference impractical for most users.
- Not uncensored: safety alignment is present, which may limit certain edge-case generations.
- Venice hosts the text-only variant; vision or image-generation capabilities are not available here.
- FP8 quantization on Venice trades marginal precision for throughput compared to full-precision deployment.
- As a Chinese lab model, some enterprise procurement workflows may require additional compliance review despite open weights.
Capabilities
What it supports
- Tool use / function calling
- Vision (image input)
- Reasoning
- Web search
- Code-optimized
- Structured output (JSON schema)
- Audio input
- Video input
- Multiple image inputs
- Log probabilities
Specifications
Datasheet
- Maker
- Zhipu AI
- Released
- February 12, 2026
- Architecture
- Mixture-of-Experts (MoE) with DeepSeek Sparse Attention
- Parameters
- 744B total · 40B active
- Open weights
- Yes
- Modality
- Text
- Context window
- 198K tokens
- Max output
- 32K tokens
- Capabilities
- Function calling, Reasoning, Web search, Code-optimized
- Privacy on Venice
- Private — zero retention
- Available on Venice since
- Feb 2026
API
Call it from your code
Venice exposes an OpenAI-compatible API. Point your base URL at Venice and pass the model id.
curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "zai-org-glm-5",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Pricing
What it costs on Venice
Billed per token on Venice: $1 per 1M input tokens and $3.20 per 1M output tokens.
New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.
Alternatives
How it compares
| Model | Price (Venice) | Context window | Open weights | Best for |
|---|---|---|---|---|
| GLM 5 | $1 in · $3.20 out / 1M | 198K tokens | Yes | Agentic coding & reasoning |
| DeepSeek V3.2 | $0.33 in · $0.48 out / 1M | 160K tokens | Yes | Efficiency & low-cost inference |
| Kimi K2.6 | $0.75 in · $3.50 out / 1M | 256K tokens | Yes | Long-context agentic tasks |
| Claude Sonnet 4.6 | $3.60 in · $18 out / 1M | 1M tokens | No | General reasoning (closed) |
Open-weight MoE with tool use, reasoning, and web search. The best open choice for long-horizon engineering tasks.
Use cases
What it is good for
- 01Autonomous software engineering with iterative tool use, debugging, and multi-file refactoring.
- 02Long-horizon agentic workflows that combine web search, reasoning, and structured data extraction.
- 03Self-hosted deployment for organizations requiring data sovereignty and zero third-party access.
- 04Code generation and analysis across large codebases within a 198K-token context.
- 05Conversational AI with function calling for internal enterprise tools and APIs.
Prompting
Getting better results
Provide explicit function schemas and tool descriptions — GLM 5 is optimized for function-calling accuracy.
Break complex coding tasks into steps and leverage its reasoning capability for multi-turn planning.
Use structured output (JSON schema) when you need deterministic, parseable responses.
For long-context prompts, repeat key instructions near the end of the input to improve recall.
Version history
Predecessor — smaller 355B MoE.
Current open-weight flagship for agentic engineering.
FAQ
Frequently asked questions
GLM 5 is Zhipu AI's open-weight flagship LLM, a 744B-parameter Mixture-of-Experts model released in February 2026. It is designed for agentic engineering, advanced reasoning, and end-to-end software engineering with native tool use and web search.
Venice bills GLM 5 at $1 per 1M input tokens and $3.20 per 1M output tokens, with cached input at $0.20 per 1M tokens. There is no subscription required.
Yes. GLM 5 ships open weights under a permissive license, with model files available on Hugging Face. You can self-host or fine-tune it, though full-precision deployment requires significant GPU resources.
Yes. On Venice, GLM 5 supports function calling, reasoning, web search, and structured JSON output, making it suitable for autonomous agentic workflows.
Choose GLM 5 for agentic coding and reasoning with a 198K context. Choose DeepSeek V3.2 if you want the lowest open-weight price and a slightly shorter 160K context.
No. GLM 5 is not uncensored; it includes safety alignment. For fully uncensored open-weight models, browse Venice's uncensored collection.
Yes. The open weights are downloadable, but self-hosting at full precision requires massive GPU infrastructure. Most users will find Venice's private API far more practical.
Venice supports a 198K-token context window and up to 32K tokens of output per request.
Run GLM 5 privately
No prompt logging. No data used for training.