GLM-5.2 introduces a robust 1M-token context and advanced, multi-effort coding capabilities to significantly enhance performance on long-horizon tasks. Its new IndexShare architecture and improved MTP layer simultaneously boost efficiency by reducing per-token FLOPs and increasing speculative decoding lengths.
ServerlessDocs | GLM 5.2 is available via Fireworks' serverless API, where you pay per token. There are several ways to call the Fireworks API, including Fireworks' Python client, the REST API, or OpenAI's Python client. |
On-demand DeploymentDocs | On-demand deployments allow you to use GLM 5.2 on dedicated GPUs with Fireworks' high-performance serving stack with high reliability and no rate limits. |
Run queries immediately, pay only for usage
A Mixture-of-Experts model from Z.ai, released June 16, 2026 as the successor to GLM-5.1. It is positioned as a flagship model for long-horizon tasks with a 1M-token context.
Coding-agent scenarios: large-scale implementation, automated research, performance optimization, and complex debugging. Documented cases include project-level codebase analysis, cross-file refactoring, and on-device debugging through ADB and logcat.
1,048,576 tokens, set as max_position_embeddings in config.json. Z.ai and the model card both describe this as a 1M-token context.
Fireworks lists a 1040k-token context window, available on both the serverless endpoint and on-demand deployments.
Fireworks serves the unquantized zai-org/GLM-5.2 repository at BF16, and publishes an FP8 checkpoint separately as glm-5p2-fp8, which is on-demand only.
Fireworks applies the model's Hugging Face generation_config.json when sampling parameters are not set. For GLM-5.2 that is temperature 1.0 with top_p 0.95.
max_tokens defaults to 2,048 and accepts values up to the full context window; hitting the limit returns a finish_reason of length. Z.ai caps output at 128K on its own API.
Z.ai documents reward-hacking risks in coding-task training and evaluation, including shortcut behaviors that can corrupt the training signal, and describes an anti-hack module that detects and blocks such actions during training and evaluation.
Both are supported. Streaming runs through the stream parameter on chat completions, and the model page lists function calling as supported.
GLM-5.2 has 743B parameters in a Mixture-of-Experts architecture.
Yes, LoRA with managed SFT and DPO plus a full-parameter shape, at a 204,736-token training context. Training targets the GLM 5.2 FP8 variant on an on-demand deployment.
Fireworks returns prompt_tokens, completion_tokens, and total_tokens in the usage object. Serverless billing prices input, cached input, and output at $1.40, $0.14, and $4.40 per 1M tokens.
Serverless ceilings for this model's size tier default to 43.2M total prompt TPM, 10.8M uncached prompt TPM, and 432k generated TPM, adaptive per account and model. On-demand deployments carry no rate limits.
Fireworks announces serverless model deprecations in advance, following its serverless model lifecycle policy. For long-term version stability it recommends on-demand deployments.
Released under the MIT license, which permits commercial use provided the copyright and license notice are retained. Z.ai states there are no regional limits.
No. Fireworks operates zero data retention by default and does not log prompt or generation data for open models without opt-in. The Response API is the exception, retaining data 30 days unless store is false.
