DEV.co
Open-Source LLM · zai-org

GLM-4.5-Air-FP8

GLM-4.5-Air-FP8 is a 106B-parameter mixture-of-experts (MoE) model with 12B active parameters, quantized to FP8 precision for reduced memory footprint. It supports reasoning, code generation, tool use, and conversational tasks in English and Chinese. MIT-licensed and ungated, it is available for download and commercial use. Inference requires 2× H100 or 1× H200 GPUs for full-featured operation; fine-tuning is supported via LoRA on 4× H100 or equivalent.

Source: HuggingFace — huggingface.co/zai-org/GLM-4.5-Air-FP8
110.5B
Parameters
mit
License (OSI-approved)
Unknown
Context (tokens)
37.9k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developerzai-org
Parameters110.5B
Context windowUnknown
Licensemit — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads37.9k
Likes80
Last updated2025-08-12
Sourcezai-org/GLM-4.5-Air-FP8

What GLM-4.5-Air-FP8 is

GLM-4.5-Air-FP8 is a hybrid reasoning MoE model released by zai-org under MIT license. Architecture includes MTP (multi-token prediction) layers, speculative decoding support, and dual-mode reasoning (thinking and non-thinking). Trained on 128K context length (stated in system requirements table). FP8 quantization reduces weight and cache size compared to BF16 baseline. Model code is integrated in HuggingFace Transformers, vLLM, and SGLang. Last updated 2025-08-12.

Quickstart

Run GLM-4.5-Air-FP8 locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="zai-org/GLM-4.5-Air-FP8")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Cost-efficient agent deployment

FP8 quantization and MoE architecture allow running intelligent agents on 2× H100 GPUs instead of 8×, reducing infrastructure cost while maintaining performance (59.8 benchmark score vs 63.2 for full GLM-4.5). Suitable for production systems where efficiency matters.

Code-centric applications

Tagged for code-generation tasks with reasoning capabilities. Hybrid thinking mode enables debugging and multi-step problem-solving. Practical for CI/CD automation, code review tools, and developer-facing AI features.

Bilingual agent systems

Supports English and Chinese natively with agentic capabilities. Suitable for customer support agents, internal tooling, and regional enterprise deployments where multilingual reasoning is required.

Running & fine-tuning it

Minimum inference: 2× H100 or 1× H200 GPU with ≥1TB server memory, batch size ≤8, FP8 support native to hardware. Full 128K context: 4× H100 or 2× H200. Fine-tuning LoRA: 4× H100 with 96GB VRAM per GPU (H20 example given). Estimated VRAM per GPU (FP8, batch size 1): ~50–70 GB for inference; higher for context handling.

LoRA fine-tuning is supported via Llama Factory and Swift frameworks. Llama Factory requires 4× H100 with batch size 1 per GPU. Swift offers more flexible scaling (H20 96GB examples: 4 GPUs for LoRA, 32+ for SFT/RL). Full fine-tuning (SFT/RL) requires significantly more GPUs (32–128× H20). QLoRA is not explicitly mentioned; assume LoRA is primary lightweight option.

When to avoid it — and what to weigh

  • Extreme latency sensitivity — FP8 quantization and MoE routing add computational overhead. Recommended batch size ≤8; real-time, single-token-per-request use cases may see higher latency than smaller, non-quantized models. Check benchmarks in technical report before committing.
  • Limited GPU infrastructure — Minimum 2× H100 or 1× H200 required for full-featured inference; 4× H100 needed for fine-tuning. Not suitable for edge devices, mobile, or single-GPU development environments. Requires 1T+ server memory.
  • Strict output determinism — Mixture-of-experts models may exhibit variance in routing decisions, affecting reproducibility. If exact-match token sequences are required, evaluate variation under your use case before production deployment.
  • Proprietary dependency concerns — Model weights and reasoning/tool parsers depend on HuggingFace Transformers, vLLM, and SGLang updates. Any breaking changes in those libraries could affect inference. Monitor upstream compatibility regularly.

License & commercial use

MIT license. Explicitly permits commercial use and secondary development according to model card.

MIT is a permissive OSI-approved license that explicitly allows commercial use, modification, and distribution with attribution. No restrictions stated in model card. However, commercial use of any LLM depends on compliance with data usage terms, output liability, and deployment context. Verify with legal counsel regarding your specific commercial application (e.g., SaaS liability, data residency, GDPR).

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityHigh
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Model weights and code are publicly available on HuggingFace and GitHub (ungated). Inference frameworks (vLLM, SGLang) are community-maintained; evaluate their security posture independently. FP8 quantization does not introduce known security vulnerabilities, but inference still requires sandboxing and access controls for production use. No security audit or threat model in provided data. Treat as with any large language model: validate outputs, monitor for prompt injection, and isolate inference endpoints.

Alternatives to consider

Llama 3.1 70B (or 405B)

Open-source, MIT-licensed, larger community, but higher VRAM requirements. No native reasoning mode. Better if you prioritize ecosystem size over efficient reasoning.

Qwen2.5 72B

Competitive benchmark performance, Apache 2.0 license, lower quantized VRAM footprint. No MoE or reasoning mode. Better if bilingual support is secondary and you want simpler inference.

Mixtral 8x7B (or 8x22B)

MoE architecture with lower total parameter count, easier to run on smaller GPU clusters. Apache 2.0 license. Lacks reasoning and tool-use specialization; consider if agent features are not critical.

Software development agency

Ship GLM-4.5-Air-FP8 with senior software developers

Start with our infrastructure consulting to size your GPU cluster, validate inference performance on your workload, and set up vLLM or SGLang pipelines. We'll also review commercial use compliance and fine-tuning costs.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

GLM-4.5-Air-FP8 FAQ

Can I use this model commercially without paying for a license?
Yes. MIT license permits commercial use, modification, and redistribution with attribution. However, commercial deployment of LLMs involves liability, data residency, and compliance considerations beyond the license itself. Consult legal counsel for your use case.
What hardware do I need to run this model?
Minimum 2× H100 or 1× H200 GPU with ≥1TB server memory for baseline inference (batch ≤8, no full context). For full 128K context, you need 4× H100 or 2× H200. FP8 precision is critical for memory efficiency; BF16 requires 8× the VRAM.
How does FP8 quantization affect output quality?
Model card reports benchmark score of 59.8 (competitive among open models) vs 63.2 for full GLM-4.5 BF16. No detailed quality degradation study in provided data. Evaluate on your own use cases before production.
Can I fine-tune this model on my own data?
Yes, LoRA fine-tuning is supported via Llama Factory (4× H100) or Swift (4× H100 for LoRA, 32+ for full SFT). QLoRA is not explicitly mentioned. Plan for significant compute resources; contact framework maintainers for sparse or low-rank alternatives.

Software developers & web developers for hire

Need help beyond evaluating GLM-4.5-Air-FP8? DEV.co is a software development agency offering software development services and web development for teams of every size. Our software developers and web developers build custom software, web applications, APIs, and open-source llms integrations — and maintain them long-term.

Ready to deploy GLM-4.5-Air-FP8?

Start with our infrastructure consulting to size your GPU cluster, validate inference performance on your workload, and set up vLLM or SGLang pipelines. We'll also review commercial use compliance and fine-tuning costs.