English | 简体中文
This repository is dedicated to collecting, organizing, and continuously updating the list of all API providers, cloud acceleration solutions, aggregation gateways, and local self-hosting deployment guides for GLM-5.2 (Z.ai / Zhipu AI's flagship open-weight LLM). Contributions via PRs are welcome!
Note
GLM-5.2 (Open-weight version released in mid-June 2026) is the world's first frontier MoE model trained completely without Nvidia hardware dependencies, built entirely on Huawei Ascend hardware. It features a total parameter count of 744B~753B, dynamically activating 40B parameters per token, and supports up to 1M (1,048,576 tokens) context window.
- 📊 API Providers Matrix
- 💡 Deep Dive: Core Providers & Features
- 🤖 Agentic Workflows & "Execution Tax" Evaluation
- 💻 Self-Hosting & Hardware Compatibility Guide
- 🎯 Business Decision Matrix
- 🤝 Contributing
- 📄 License
The following table aggregates the technical and economic metrics of official channels, third-party Serverless platforms, and API gateways for GLM-5.2 (Data verified as of June 25, 2026):
| Provider | Classification | Website & Docs | Model ID / Identifier | Context Limit | Pricing / Billing Model (per 1M Tokens) | Typical TPS / Performance |
|---|---|---|---|---|---|---|
| Z.ai Global (Official) | Official Global Portal | Website / Docs | glm-5.2 |
1,048,576 | Input: $1.40 / Output: $4.40 Cache Hit: $0.26 |
~106 tok/s (Peak flow speed) |
| BigModel.cn (Official) | Official China Portal | Website | glm-5-2 |
1,048,576 | Input: ¥8.00 / Output: ¥28.00 Cache Hit: ¥2.00 |
~19.71 tok/s (Standard speed) |
| DeepInfra | Serverless Provider | Website | zai-org/GLM-5.2 |
1,048,576 | Input: $0.95 / Output: $3.00 Cache Hit: $0.18 |
~60 tok/s (Best value) |
| SiliconFlow | Serverless Provider | Website | zai-org/GLM-5.2 |
1,048,576 | Input: ¥6.00 / Output: ¥28.00 Cache Hit: ¥2.00 |
~41 tok/s (Low latency in China) |
| Fireworks AI | Serverless Provider | Website | accounts/fireworks/models/glm-5p2 |
1,040,000 | Input: $1.40 / Output: $4.40 Cache Hit: $0.26 (Fire Pass available) |
~318 tok/s (Fastest overall) |
| Together AI | Serverless Provider | Website | zai-org/GLM-5.2 |
262,144 (Restricted) | Input: $1.40 / Output: $4.40 Cache Hit: $0.26 |
~115 tok/s (FP4 Quantized) |
| Novita AI | Serverless Provider | Website | zai-org/glm-5.2 |
1,048,576 | Input: $1.40 / Output: $4.40 Cache Hit: $0.26 |
~48 tok/s |
| Baseten | Dedicated Instance / PTU | Website | zai-org/GLM-5.2 |
1,048,576 | Custom PTU Enterprise Billing | ~242 tok/s (Dedicated instance) |
| Alibaba Cloud (DashScope) | Cloud Provider | Website | glm-5-2 |
1,048,576 | Input: ¥8.00 / Output: ¥28.00 (TokenPlan credits apply) | N/A (Enterprise concurrency SLA) |
| Cloudflare Workers AI | Edge Serverless | Website | @cf/zai-org/glm-5.2 |
262,144 (Restricted) | Input: $1.40 / Output: $4.40 Cache Hit: $0.26 |
N/A (Dynamic edge load) |
| OpenRouter | Aggregation Gateway | Website | z-ai/glm-5.2 |
1,048,576 | Weighted Avg Input: ~$0.50 / Output: ~$4.22 (After prompt caching) |
Depends on chosen backend |
| ZenMux.ai | Aggregation Gateway | Website | z-ai/glm-5.2 |
1,048,576 | List Price: $1.40 / $4.40 (Subscription tiers available) | Depends on chosen backend |
| Vercel AI Gateway | Developer Gateway | Website | zai/glm-5.2 |
1,048,576 | Passthrough billing of backend | Depends on chosen backend |
| Neuralwatt | Energy-Metered Platform | Reddit Community | zai-org/GLM-5.2 |
1,048,576 | GPU Energy Metered: $0.06 ~ $0.12 | ~60 - 100 tok/s |
| OpenCode Go | Subscription Service | Website | opencode/glm-5.2 |
1,048,576 | $5 first month, then $10/mo (Rolling credits: $12/5h, $30/wk, $60/mo) |
Depends on backend provider |
Warning
Some third-party platforms (like Together AI and Cloudflare Workers AI) physically limit the maximum context window of GLM-5.2 to 262,144 tokens. For 1M long-context tasks, choose the official platform or other Serverless providers without context window caps.
API providers have heavily optimized the GLM-5.2 infrastructure layer, offering distinct hardware setups, compiler improvements, and commercial models.
- Flat-Rate Long Context Pricing (BigModel.cn / Z.ai): Official pricing has removed length-based tiers. GLM-5.2 charges a flat ¥8.00 / $1.40 per million input tokens, completely mitigating the risk of billing spikes during full-codebase audits or massive RAG search tasks.
- GLM Coding Plan (Unlimited Monthly Subscription): Official subscriptions (Lite: 49元/mo, Pro: 149元/mo, Max: 469元/mo; Enterprise Team plans: $12.60 to $112) allow developers to bind GLM-5.2 into Zed, Cursor, Claude Code, or Z.ai's official Z Code desktop terminal, facilitating high-intensity "Vibe Coding" without per-token constraints.
- Alibaba Cloud DashScope: DashScope offers TokenPlan and Savings Plan models. Unified enterprise credits can be shared across models, letting enterprises seamlessly balance Qwen3 Max and GLM-5.2 calls.
Baseten utilizes dedicated NVIDIA Blackwell GPU clusters to achieve high-throughput GLM-5.2 API performance, clocking in at 242 TPS for streaming output (Fireworks AI currently leads overall serverless throughput at 318 TPS).
- Blackwell NVFP4 Native Quantization: Using native 4-bit floating-point (FP4) hardware execution on Blackwell GPUs to compress the 750B model weights without degrading coding generation quality.
- Dynamo Prefill-Decode Disaggregation (PD separation): Decouples the computationally heavy Prompt Prefill (establishing KV Cache) from the streaming Decode phase by running them on distinct physical Blackwell nodes.
- Cache-Sensitive Routing: Tracks and preserves KV Caches across interactive turns. In long-horizon coding tasks (e.g., using Claude Code), this avoids repeated calculation of up to 1MB prompts.
- Fireworks AI Fire Pass: A subscription model ($10/month) that unlocks a developer key for GLM-5.2. Personal projects using Claude Code, Cline, or Kilo Code can call the model with zero per-token cost (restricted to non-production usage).
- Neuralwatt (Energy-based Billing): Users pay based on real GPU power consumption. Running GLM-5.2 in
Max Thinkingmode during off-peak night hours yields an effective pricing of $0.06 ~ $0.12 per million output tokens. - OpenCode Go Subscription: A low-cost developer bundle subscription costing $5 for the first month, then $10/month. Instead of raw call counts, it operates on a rolling dollar-equivalent credit system ($12 per 5 hours, $30 per week, and $60 per month). It provides a unified API key to access GLM-5.2 alongside other open-weight models, offering excellent value for active agent development.
- OpenRouter Multi-Provider Routing & Caching: OpenRouter serves as a dynamic aggregator routing to 8 underlying providers (including NovitaAI, Z.ai, StreamLake, Fireworks, Friendli, GMICloud, Together, and AtlasCloud). Because it supports robust prompt caching, the rolling 30-day weighted average effective price drops to $0.500 per 1M input tokens and $4.22 per 1M output tokens (saving 60-80% on repeat context input), making it one of the most cost-effective global routing methods. It also provides millisecond-level failover during peak traffic periods (GLM-5.2 upstream uptime average is ~95.12%).
- ZenMux.ai "Hallucination & Latency Insurance": An integrated probe monitors upstream API performance. If it detects severe latency spikes, upstream silent downgrades, or invalid JSON structures, ZenMux credits the user's account automatically as compensation.
In complex, multi-step Agent workflows (such as browser control, tool calls, and codebase refactoring), if a model outputs invalid JSON or gets stuck in tool-calling loops, the framework must resend the entire interaction history. This overhead is defined as the "Agent Execution Tax".
Based on system-wide stress tests on the WebVoyager suite (60 tasks on 15 commercial websites like Amazon and GitHub, repeated 180 times):
| Model Selection (on Fireworks) | Total LLM Calls | Parse/JSON Retries | Retry Rate | Avg. Calls per Task | "Agent Execution Tax" (Overhead) |
|---|---|---|---|---|---|
| Moonshot Kimi K2.5 | 852 | 0 | 0.0% | 10.2 | 0.0% (Zero overhead) |
| Z.ai GLM-5 (Flagship) | 884 | 5 | 0.6% | 10.3 | 0.6% (Near-perfect run) |
| Baichuan MiniMax M2.5 | 828 | 13 | 1.6% | 9.8 | 1.6% (Negligible overhead) |
| Google Gemini 2.5 Flash | 886 | 165 | 18.6% | 14.7 | 22.9% (Severe financial/latency tax) |
While some low-cost models seem cheaper on paper, their 18.6% error rate means 22.9% of budget is wasted on retries. Due to dedicated fine-tuning for Tool Calling, MCP, and structured outputs, GLM-5 flagship experiences a tax of only 0.6%, offering superior cost-efficiency at scale.
In auditing Insecure Direct Object Reference (IDOR) vulnerabilities on Flask backends, GLM-5.2 achieved a 39% F1-score using raw prompts ("bare prompt") without prompt engineering or scaffolds. In comparison, Claude Code (utilizing its full agent toolset) achieved a 32% F1-score. Based on GLM-5.2's API cost, the scanning system averaged just $0.17 per confirmed vulnerability, making enterprise-wide code auditing highly viable.
Unsloth provides day-zero optimized dynamic GGUF quantization tables:
- 2-bit Dynamic Quantization (UD-IQ2_M): Only requires 239GB disk space (84% reduction from 1.51TB FP16 weights). By dynamically preserving linear attention weights, it retains 82% of baseline intelligence.
- System Requirements: Runs natively on an Apple Mac Studio (256GB Unified Memory). On personal PCs, it requires a single 24GB VRAM GPU (e.g., RTX 3090/4090) for Prefill acceleration paired with 256GB system RAM, utilizing off-line MoE offloading.
- Ollama Boot Command (Ollama v0.30+):
# Enable Bash auto-execution (YOLO mode) ollama run frob/glm-5.2 --experimental --experimental-yolo # Disable thinking process for simpler tasks to speed up inference ollama run frob/glm-5.2 --think=false
Utilizing Tsinghua's open-source KTransformers (KT-Kernel) SGLang integration allows running the model across CPU RAM and GPU VRAM. The GPU handles MLA attention layers and high-frequency experts, offloading the remaining 240 sparse experts to DDR5 system memory.
Run SGLang server with FP8 quantized GLM-5.2:
export PYTORCH_ALLOC_CONF=expandable_segments:True
export SGLANG_ENABLE_JIT_DEEPGEMM=0
python -m sglang.launch_server \
--model-path /path/to/GLM-5.2-FP8 \
--kt-weight-path /path/to/GLM-5.2-FP8 \
--kt-cpuinfer 96 \
--kt-threadpool-count 2 \
--kt-num-gpu-experts 30 \
--kt-method FP8 \
--max-total-tokens 4096 \
--trust-remote-codeNote: --kt-cpuinfer 96 designates 96 CPU cores; --kt-num-gpu-experts 30 keeps 30 experts cached in VRAM to boost throughput.
GLM-5.2 uses DeepSeek-Sparse-Attention (DSA) operators, which native vLLM and SGLang compilers bind to Nvidia Hopper (H100/H200) and Blackwell (B200) specific WGMMA instructions. Running it on consumer Ada GPUs (RTX 4090 / sm_89) triggers runtime crashes.
The community patch ada_dsa.py resolves this:
- Principle: Uses custom Triton operators and non-WGMMA tilelang paths to rewrite Indexer GEMM, Top-K, and Page-Mapping operators, enabling execution on sm_89 hardware.
- Deployment: Distributing the FP8 model across 24x RTX 4090 GPUs (spread across 3 nodes with TP=8, PP=3 configuration) yields a stream throughput of ~10 tokens/sec at full FP8 accuracy.
When serving GLM-5.2 via ROCm + vLLM on AMD Instinct MI300X (192 GiB) nodes, work around these issues:
- PP Garbled Output during Tool Calls: When Pipeline Parallelism is enabled (
PP>=2), vLLM's internalcondense()function experiences token pointer alignment shifts on long XML-heavy tool prompts. Fix: Run in single-node Tensor Parallelism mode (PP1,TP8). - HIP Memory Crashes (OOM): You must append
--disable-custom-all-reduceto prevent HIP custom all-reduce allocator buffer initialization failures. Additionally, enable--enable-chunked-prefillto prevent initial RAG context windows from saturating the MI300X VRAM.
| Team Profile / Scenario | Resource Budget | Recommended Path | Decision Rationale |
|---|---|---|---|
| Independent Dev / MVP Startup | Extremely low budget, validating MVPs | DeepInfra (US) / SiliconFlow (CN) | Utilize pay-per-token options on DeepInfra ($0.95/1M input) or SiliconFlow (¥6.00/1M input) to start MVP testing with minimal early capital expenses. |
| Active Developer / Vibe Coding | Continuous daily calls, deep Cursor/Claude Code binding | Z.ai Official Coding Plan | The flat-rate monthly subscriptions (Pro at ¥149/mo or Max at ¥469/mo) shield developers from heavy billing run-ups caused by continuous codebase scanning. |
| High Concurrency SaaS Platform | Large scale production workloads, highly latency-sensitive | Baseten Dedicated Blackwell | Latency defines retention: Baseten's Blackwell native FP4 path outputs at 242 TPS (with serverless providers like Fireworks AI reaching 318 TPS), eliminating UI blocking during multi-step Agent calls. |
| Regulated/Financial/Auditing Corp | Hard data-compliance guidelines, strict physical privacy | RTX 4090 Cluster (with ada_dsa patch) | By bypassing export-restricted H100s, enterprises can build secure, air-gapped FP8 instances on affordable, consumer-grade hardware. |
We welcome community contributions! You can:
- Submit Price Updates: Cloud provider pricing changes rapidly. Feel free to open a PR to adjust tables.
- Share Self-Hosting Guides: If you successfully run GLM-5.2 on alternative AI accelerators (e.g., Ascend, MI300X native configurations, Moore Threads), submit an issue or PR with instructions.
- Add New Providers: Add new verified API channels along with corresponding latency/concurrency metrics.
This repository is licensed under the MIT License.
Disclaimer: All provider details, pricing, and benchmark metrics are compiled from public community resources and documentation. Commercial offerings may change; check vendor portals before purchasing.