GLM-4.5: Agentic Foundation Model from Z.ai

TL;DR

Z.ai’s open release of GLM-4.5 (355B total / 32B active) and GLM-4.5-Air (106B / 12B), a hybrid-reasoning MoE family with thinking / non-thinking modes, MIT-licensed weights, and FP8 variants. The model card markets the series as foundation models designed for intelligent agents — unifying reasoning, coding, and tool-use into a single hybrid model — and reports a 63.2 average across 12 benchmarks, placed third behind two unnamed proprietary frontiers. The release ships base, hybrid-reasoning, and FP8 variants plus reference recipes for vLLM, SGLang, Transformers, Llama-Factory and Swift, along with a published technical report (arXiv:2508.06471) and a Z.ai technical blog that — per the sharer — emphasises a “mid-training” stage stuffed with agentic data.

Key claims

GLM-4.5 has 355B total parameters with 32B active, and GLM-4.5-Air uses a more compact 106B / 12B MoE design [§Model Introduction].
Both models are hybrid-reasoning models that expose two modes — a thinking mode for complex reasoning and tool use and a non-thinking mode for immediate responses — toggleable at request time via enable_thinking [§Model Introduction, §Request Parameter Instructions].
The full family — base, hybrid-reasoning, and FP8 variants for both 355B and 106B sizes — is released under the MIT license with commercial-use rights [§Model Introduction, §Model Downloads].
Across 12 industry-standard benchmarks the 355B model averages 63.2 and the 106B-Air averages 59.8; Z.ai self-reports the 355B as placing third behind two unnamed proprietary frontiers [§Model Introduction].
A working FP8 deployment of the 355B fits on 8× H100 (or 4× H200) for inference at typical batch sizes, with full 128K context requiring 16× H100 / 8× H200 in FP8 [§Inference, tables].
The release ships reference recipes for Transformers, vLLM (with --tool-call-parser glm45 --reasoning-parser glm45), and SGLang with EAGLE speculative decoding (--speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4) [§vLLM, §SGLang].
The underlying GLM-4.5 technical report (arXiv:2508.06471) and a separate Z.ai technical blog are listed alongside the model card as the authoritative descriptions; the sharer’s read of that blog is that more agentic data is added in a “mid-training” stage [pointers.note, §Model Introduction].

Method

The model card itself does not describe training mechanics in depth — it points to the technical report (arXiv:2508.06471) and the Z.ai blog for that — but it does fix the architectural and release shape. GLM-4.5 is a fine-grained MoE (32B active out of 355B; 12B out of 106B for Air) trained as a hybrid reasoning model where the two modes share weights and are selected by a chat-template flag (enable_thinking). The reference inference recipe pairs the model with EAGLE-style multi-token-prediction speculative decoding (the SGLang command uses --speculative-algorithm EAGLE with a 3-step / top-1 / 4-draft configuration), and vLLM ships a custom glm45 tool-call parser and reasoning parser — suggesting the post-training emits a structured reasoning + tool-call format that needs dedicated parsing rather than the OpenAI default.

The sharer’s note flags the more interesting part: per the Z.ai blog (not fetched here directly — only the model card is the artifact filed), GLM-4.5 puts a larger share of agentic training data into a “mid-training” stage between base pretraining and post-training. This mirrors the Luma omni memo’s argument for interleaved data — pushing capability acquisition that would normally land in SFT/RL into a pre-RL stage with longer-form, mixed-objective sequences. The exact mid-training recipe is not in this artifact.

Results

The model card reports a single headline number: a 63.2 average across 12 benchmarks for the 355B model and 59.8 for the 106B-Air. No per-benchmark breakdown, no comparison table, and no description of the benchmark suite live in the card itself — the full evaluation is deferred to the technical report and blog. The placement claim (“3rd among all proprietary and open-source models”) is similarly unsubstantiated in the card.

What is substantiated in the card is the deployment envelope: FP8 inference of the 355B fits on 8× H100 (4× H200), BF16 fits on 16× H100 / 8× H200. The 128K-context envelope doubles those budgets. Fine-tuning numbers are also given: LoRA fits the 355B on 16× H100; full SFT or RL on the 355B requires 128× H20-96GiB.

Why it’s interesting

This is the third meaningful Z.ai datapoint in the wiki — alongside the GLM-5 vibe-coding tech report (GLM-5: from Vibe Coding to Agentic Engineering §1) and the GLM-5V-Turbo and GLM-5.1 announcements (GLM-5V-Turbo — Vision Coding Model from Z.ai (announcement), GLM-5.1 — Next Level of Open Source (Z.ai announcement)) — and the earliest of them by release date (Aug 2025). It anchors the GLM-4.5 → GLM-5 lineage with a concrete architecture-and-license commitment: 355B/32B MoE, MIT-licensed, FP8 from day-0, hybrid-reasoning dual mode. Per the Open foundation-model releases framing, this is another single-model open release with deep variant packaging (base + hybrid-reasoning + FP8, two sizes) plus published technical report — same release-shape grammar as DeepSeek and Qwen. The MoE active-fraction (~9% for the 355B, ~11% for Air) is in the same band as DeepSeek-V3 and Kimi K2, consistent with the MoE Routing Design observation that fine-grained-MoE-with-high-sparsity is the modal 2025–2026 frontier-open recipe. The sharer’s “mid-training with more agentic data” reading also touches the Reasoning RL thread: if the agentic capability is mostly baked in pre-RL through interleaved mid-training data rather than learned through outcome-reward RL post-training, that recipe differs from the OpenThoughts (OpenThoughts: Data Recipes for Reasoning Models) / DeepSeek-R1 line and is closer to what the Luma omni memo argues.