GLM-5.2-FP8
Z.ai's next-generation flagship for agentic software engineering. Significantly stronger coding than GLM-5.1, with a 1M-token context window for sustained repository-scale workflows.
Quantization
FP8
GLM MoE
Context Window
1M
Tokens
Intelligence
51
AA Index
License
Open
Open Weights
A Major Step Up From GLM-5.1
GLM-5.2-FP8 is Z.ai's latest flagship open model for long-horizon agentic work, coding, and complex engineering. It pairs stronger real-world coding performance with a 1M-token context window — ideal for repository-scale workflows, extended terminal sessions, debugging, tool orchestration, and ambiguous multi-step projects.
Under the hood, GLM-5.2 uses Z.ai's improved GLM MoE architecture with FP8 quantization, IndexShare sparse-attention optimization, and enhanced speculative decoding via improved MTP — reducing long-context compute while improving throughput. Thinking mode is on by default; disable with {"chat_template_kwargs": {"enable_thinking": false}}.
Flagship Agentic Engineering Model
Best for sustained engineering agents
Agentic Engineering
Strong performance on SWE-bench Pro, NL2Repo, DeepSWE, FrontierSWE and SWE-Marathon — built for multi-file edits and repository-scale work.
Terminal & Tool Use
Tuned for Terminal Bench 2.1 workflows — sustained tool calls, shell execution, and iterative debugging across long sessions.
Long-Horizon Reasoning
1M-token context plus IndexShare sparse attention keeps the model productive over extended planning, experimentation and iteration loops.
Repository Generation
Top-tier NL2Repo performance — turn specs, tickets and prompts into working codebases without losing the thread mid-build.
Artificial Analysis Intelligence Index
AA Intelligence Index v4.0 scores. GLM-5.2 holds frontier-tier reasoning while taking a clear step up on agentic engineering benchmarks over GLM-5.1.
Intelligence Index
Better than 92% of models
GPQA Diamond
Better than 93% of models
τ²-Bench Telecom
Better than 95% of models
| Category | Benchmark | Score | Description |
|---|---|---|---|
| Reasoning | GPQA Diamond | 85% | Graduate-level scientific reasoning |
| Reasoning | Humanity's Last Exam | 29% | Frontier reasoning evaluation |
| Reasoning | τ²-Bench Telecom | 98% | AI agents in dual-control scenarios |
| Reasoning | AA-LCR | 64% | Long context reasoning evaluation |
| Reasoning | IFBench | 77% | Instruction-following accuracy |
| Reasoning | GDPval-AA | 54% | Agentic performance on real-world work tasks |
| Coding | SWE-bench Pro | 47% | Real-world software engineering tasks |
| Coding | Terminal Bench 2.1 | 48% | Agentic coding & terminal use |
| Coding | NL2Repo | 52% | Natural-language to repository generation |
| Coding | SciCode | 46% | Python for scientific computing |
| Knowledge | AA-Omniscience Accuracy | 28% | Proportion of correctly answered questions |
| Knowledge | AA-Omniscience Non-Hallucination | 73% | Confidently answered questions that are correct |
Metrics sourced from Artificial Analysis and Z.ai's published evaluations. Reasoning (thinking) mode enabled.
Flexible Pricing Tiers
Choose the optimal balance of speed and cost for your workflow. Prices are per 1M tokens.
| Tier | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| Batch | $0.47 | $1.50 |
| Async | $0.70 | $2.25 |
| Realtime | $0.93 | $3.00 |
Context window natively supported up to 1,048,576 tokens.
Start Building in Minutes
GLM-5.2-FP8 is accessible via OpenAI-compatible endpoints. Default sampling: temperature=1.0, top_p=0.95.
from openai import OpenAI
client = OpenAI(
api_key="your-api-key-here",
base_url="https://api.doubleword.ai/v1"
)
# Long-horizon agentic coding task (thinking enabled by default)
response = client.chat.completions.create(
model="zai-org/GLM-5.2-FP8",
messages=[
{"role": "user", "content": "Refactor this repo to use async I/O end-to-end."}
],
temperature=1.0,
top_p=0.95,
# To disable step-by-step reasoning:
# extra_body={"chat_template_kwargs": {"enable_thinking": False}},
)
print(response.choices[0].message.content)💡 Pro Tip
GLM-5.2 is built for long-horizon agentic work — let it sustain reasoning across planning, tool use, experiments, and iterative debugging. Keep thinking mode on for ambiguous tasks. Disable with "chat_template_kwargs": {"enable_thinking": false} for latency-sensitive endpoints.