DEV.co
Open-Source LLM · zai-org

GLM-4.7-FP8

GLM-4.7-FP8 is a 358B-parameter open-source language model from zai-org optimized for coding, tool use, and complex reasoning tasks. It is quantized to 8-bit floating point for reduced memory footprint and is available under the MIT license without access restrictions. The model shows competitive performance on coding benchmarks (SWE-bench, LiveCodeBench) and reasoning tasks (HLE, AIME), with claimed improvements in multilingual agentic coding and UI generation.

Source: HuggingFace — huggingface.co/zai-org/GLM-4.7-FP8
358.5B
Parameters
mit
License (OSI-approved)
Unknown
Context (tokens)
72.2k
Downloads (30d)

Key facts

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

FieldValue
Developerzai-org
Parameters358.5B
Context windowUnknown
Licensemit — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads72.2k
Likes124
Last updated2025-12-23
Sourcezai-org/GLM-4.7-FP8

What GLM-4.7-FP8 is

GLM-4.7-FP8 is a 358.5B parameter mixture-of-experts (MoE) transformer trained on multilingual data (English, Chinese). It supports interleaved, preserved, and turn-level thinking modes for agentic workflows. The FP8 quantization reduces precision from standard float32/bfloat16 while maintaining usability via inference frameworks (vLLM, SGLang). Context length is not specified. Default inference uses temperature 1.0, top-p 0.95, and up to 131k tokens max output; task-specific settings vary.

Quickstart

Run GLM-4.7-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-4.7-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

Autonomous code generation and bug fixing

73.8% on SWE-bench Verified. Suitable for CI/CD integration, code review automation, and developer-copilot scenarios requiring multi-file context and terminal command execution.

Complex reasoning and mathematical problem solving

42.8% on HLE (with tools), 95.7% on AIME 2025. Applicable to educational platforms, competitive programming assistance, and research problem decomposition.

Multi-step tool-using agents

67.5% on BrowseComp (with context management), supports web browsing and terminal tasks. Ideal for automation workflows that require planning, tool sequencing, and state tracking.

Running & fine-tuning it

ESTIMATE (not officially stated): FP8 quantization of 358B parameters requires approximately 180–220 GB VRAM for inference (vs. ~720 GB for bfloat16). Multi-GPU setup (8× H100 or A100 80GB) or tensor parallelism recommended. Single-GPU inference not feasible; batch serving scales with available VRAM.

Model card does not mention LoRA, QLoRA, or fine-tuning procedures. 358B scale suggests full-parameter training is prohibitive for most teams. Feasibility of LoRA depends on internal architecture documentation (not provided). Recommend contacting zai-org or checking GitHub repo for community-contributed fine-tuning recipes.

When to avoid it — and what to weigh

  • Latency-critical or real-time inference scenarios — 358B parameters at FP8 still requires substantial VRAM and compute. Preserved thinking mode adds latency overhead for multi-turn workflows. If sub-100ms response times are required, consider smaller models.
  • Resource-constrained edge or embedded deployments — FP8 quantization helps but 358B model is not suitable for mobile, IoT, or single-GPU setups without clustering. Quantization to lower precision (int4, int8) would require retraining or additional optimization work.
  • Production deployments without vLLM/SGLang validation — Model card states vLLM and SGLang require their nightly/dev branches for support. Transformers library integration at version 4.57.3+ may not be stable. Requires thorough testing in your inference stack before committing to production.
  • Specialized vertical domains without evaluation data — Benchmarks focus on coding, reasoning, and tool use. No evaluation provided for legal, medical, financial, or domain-specific fine-tuning. Unclear generalization to proprietary use cases.

License & commercial use

MIT License. MIT is a permissive OSI-approved license allowing commercial use, redistribution, and modification with attribution and liability disclaimer.

MIT license permits commercial use, including proprietary applications and closed-source deployment. No restrictions on commercial services, API endpoints, or monetization. No gating or access controls enforced. Suitable for building commercial products without additional licensing obligations, though users remain responsible for compliance with underlying training data sources and any local regulations.

DEV.co evaluation signals

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

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

No explicit security audit, threat model, or adversarial robustness evaluation provided. Large 358B model trained on multilingual data; typical LLM risks apply (prompt injection, jailbreaking, hallucination, toxic outputs). Preserved thinking mode may increase context window attack surface in multi-turn scenarios. FP8 quantization does not inherently reduce or increase security. Recommend standard hardening: input validation, output filtering, rate limiting, and monitoring if deployed in customer-facing systems. No mention of safety classifiers or content filtering.

Alternatives to consider

DeepSeek-V3.2

Comparable 358B+ parameter model; 73.1% on SWE-bench Verified (vs. GLM-4.7's 73.8%). Better on some reasoning benchmarks (HLE w/ tools: 40.8% vs. 42.8%). Considered a strong regional alternative with different training provenance.

Claude 3.5 Sonnet

77.2% on SWE-bench Verified (higher than GLM-4.7). Closed-source, hosted API only. Smaller effective parameter count, lower latency, battle-tested in production. Trade-off: no local deployment, per-token pricing, data privacy constraints.

Llama 3.1-405B (or smaller quantized variants)

Permissive Llama 2 Community License for commercial use. 405B base model or smaller quantized versions (70B) offer flexibility. Smaller variants suit resource-constrained environments. Ecosystem mature (ollama, llama.cpp support). Coding performance lower on SWE-bench but sufficient for many agentic workflows.

Software development agency

Ship GLM-4.7-FP8 with senior software developers

Explore managed inference options, benchmark on your dataset, or review the GitHub repository for local deployment via vLLM/SGLang. Contact zai-org for production support and custom integrations.

Talk to DEV.co

Related open-source tools

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

Related on DEV.co

Explore the category and the services that help you build with it.

GLM-4.7-FP8 FAQ

Can I use GLM-4.7-FP8 in a commercial product?
Yes. MIT license permits commercial use without additional licensing fees or restrictions. You may build closed-source applications, APIs, or SaaS offerings. Ensure compliance with any restrictions on training data sources (not specified in card) and local regulations. Always attribute per MIT terms.
What are the realistic hardware and cost requirements for self-hosting?
Expect 180–220 GB VRAM (FP8 quantization). Typical multi-GPU setup: 8× H100/A100 80GB GPUs (~$500k–1M initial hardware + energy). Monthly TCO: $50k–150k depending on utilization and provider. Alternatively, use managed inference (Z.ai, Together AI, Replicate) to avoid upfront capital; per-token cost (~$0.5–2 per 1M tokens) may be higher at scale.
What is the context length?
Not specified in the model card. Default max new tokens output is 131,072; input context window size unknown. Check GitHub repo or contact zai-org for exact limit. Relevant for long-document coding and multi-turn agentic tasks.
Is this model suitable for my specific domain (legal, medical, etc.)?
Benchmarks focus on coding, math, and tool use. No domain-specific fine-tuning or evaluation provided. Recommend testing on representative data from your domain before production deployment. Fine-tuning procedures not documented in card.

Work with a software development agency

Need help beyond evaluating GLM-4.7-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-4.7-FP8?

Explore managed inference options, benchmark on your dataset, or review the GitHub repository for local deployment via vLLM/SGLang. Contact zai-org for production support and custom integrations.