Z.ai · Open Weights · MIT License
GLM-5.2: The Biggest Open Model You Can Actually Self-Host
GLM-5.2 (Z.ai, June 2026) is a 753B-parameter Mixture-of-Experts model — roughly 40B active per token, per reported figures — with a 1M-token context window, released under a plain MIT license. It was reportedly ranked the #1 open model on Artificial Analysis at launch, and unlike Kimi K3, it has a real local path today: the official Hugging Face model card confirms local-inference support, and 112 quantized versions are available for llama.cpp, Ollama, and LM Studio. The honest caveat: quantized, it still needs a big-RAM rig — Mac Studio 512GB class, an EPYC server with KTransformers, or a multi-GPU stack. This is not an 8GB-VRAM model.
Key takeaways
- →753B total MoE, ~40B active (reported) — per-token compute is modest; memory is the constraint.
- →Plain MIT license — confirmed on the Hugging Face card. No custom riders.
- →Reported #1 open model at launch on Artificial Analysis.
- →112 ready-made quants for llama.cpp / Ollama / LM Studio — the widest ready-made quant ecosystem of the June-July open wave.
- →Hardware floor: big-RAM rigs — ~380-420 GB at ~4-bit (estimate). Mac Studio 512GB, EPYC + KTransformers, or multi-GPU.
- →1M-token context with Z.ai's new IndexShare sparse attention.
Quick verdict
GLM-5.2 is the model to pick if you want the biggest open model that can genuinely live on hardware you own. Most frontier-size open releases are "open" in the sense that a datacenter could run them. GLM-5.2 is different because the local tooling shipped with it: vLLM, SGLang, Transformers, KTransformers, and Unsloth support are confirmed on the Hugging Face card, and 112 quantized versions are already available covering llama.cpp, Ollama, and LM Studio.
Who it's not for: anyone on consumer GPUs. If your machine tops out at 24GB of VRAM or 64GB of RAM, look at smaller models — our Ollama RAM/VRAM table maps what fits where. And if you don't want to self-host at all, Z.ai's API runs $1.40/$4.40 per million input/output tokens (secondary-source pricing).
Specs at a glance
| Vendor | Z.ai (Zhipu AI) |
| Released | Hugging Face card dated June 17, 2026 (launch window June 13-16 per secondary sources) |
| Architecture | Mixture-of-Experts with IndexShare sparse attention |
| Total parameters | 753 billion |
| Active parameters | ~40 billion per token (reported by secondary sources) |
| Context window | 1,000,000 tokens |
| License | MIT (confirmed on HF card) |
| Local inference | vLLM, SGLang, Transformers, KTransformers, Unsloth; 112 quants for llama.cpp / Ollama / LM Studio |
| Hugging Face | zai-org/GLM-5.2 |
| API pricing | $1.40 / $4.40 per Mtok in/out (secondary sources) |
Can you actually run it locally?
Yes — with the caveat that "locally" here means a serious machine, not a gaming PC. What makes GLM-5.2 notable is that the local path is official and already turnkey. The Hugging Face card lists five supported inference stacks (vLLM, SGLang, Transformers, KTransformers, Unsloth), and 112 quantized versions already exist for llama.cpp, Ollama, and LM Studio. Compare that to Kimi K3 — 2.8T parameters, weights due around July 27 — which no home rig will ever hold.
The ~40B-active-per-token design (reported) is what makes this work. Once the weights fit in memory, each token only touches a fraction of them, so per-token compute stays modest. The whole problem collapses to one question: can your machine hold a few hundred gigabytes of weights with decent memory bandwidth? That's the question the next section answers.
Hardware requirements
Unquantized, 753B parameters is multi-GPU server territory — for scale, the predecessor GLM-5 needed roughly 8× H200 for FP8. Quantized, three realistic self-host targets emerge:
1. Mac Studio, 512GB unified memory
At ~4-bit the weights are roughly 380-420 GB (estimate) — inside a 512GB unified-memory envelope. The cleanest single-box path, using the ready-made llama.cpp / LM Studio quants.
2. EPYC / Threadripper server + KTransformers
384-512GB of system RAM with KTransformers (officially supported on the HF card). At 2-3-bit dynamic quants, weights shrink to roughly 200-300 GB (estimate), which is why the 384GB tier becomes viable.
3. Multi-GPU stack with large aggregate VRAM
vLLM or SGLang across enough cards to hold a quantized checkpoint. Highest bandwidth of the three options, and the highest cost.
What we won't do is quote you exact GGUF file sizes or tokens-per-second — those vary by quant and rig, and we haven't benchmarked this one ourselves yet. The sizes above are arithmetic estimates from the parameter count. If your budget is smaller than any of these tiers, the RAM/VRAM sizing table covers models that fit ordinary hardware.
GLM-5.2 vs the June-July open wave
GLM-5.2 landed in the middle of an unusually crowded open-weights stretch. How the wave stacks up:
| Model | Size | Released | License | Local story |
|---|---|---|---|---|
| GLM-5.2 | 753B / ~40B active (reported) | June 2026 | MIT | 112 quants; llama.cpp / Ollama / LM Studio confirmed |
| MiniMax M3 | 428B / ~23B active | June 1 | — | Official MXFP8 + Unsloth GGUF |
| Nemotron 3 Ultra | ~550B / 55B active | June 4 | OpenMDW | — |
| Tencent Hy3 | 295B / 21B active | July 6 | Apache 2.0 | — |
| Thinking Machines Inkling | 975B | July 15 | — | — |
| Kimi K3 | 2.8T | Weights due ~July 27 | — | Not home-runnable |
Two things distinguish GLM-5.2 in this crowd: it's the one reportedly ranked #1 open model on Artificial Analysis at launch, and it has the widest ready-made quant ecosystem — the only frontier-size release in the wave with confirmed Ollama and LM Studio coverage. Against DeepSeek V4, the other frontier open family, GLM-5.2's pitch is the same: plain MIT plus a turnkey local path. For the full landscape, see our open-source LLM rankings.
FAQ
What is GLM-5.2?
Can I run GLM-5.2 locally with Ollama?
What hardware do I need for GLM-5.2?
GLM-5.2 vs DeepSeek V4: which should I pick?
GLM-5.2 vs Kimi K3: what's the difference for local use?
Is GLM-5.2 free for commercial use?
Related models
- → GLM-5 — predecessor; the ~8× H200 FP8 baseline this page compares against
- → GLM-4.6 — earlier GLM generation
- → DeepSeek V4 — the other frontier open-weight family
- → MiniMax M3 — smaller wave-mate with official MXFP8 + Unsloth GGUF
- → Nemotron 3 — Nvidia's ~550B entry in the same wave
- → Kimi K3 — 2.8T; the model GLM-5.2 beats on self-hostability
- → Best open-source LLMs 2026 — pillar comparison
Go from reading about AI to building with AI
20 structured courses. Hands-on projects. Runs on your machine. Start free.
Written by the Local AI Master Team
The team behind Local AI Master
We build Local AI Master around practical, testable local AI workflows: model selection, hardware planning, RAG systems, agents, and MLOps. The goal is to turn scattered tutorials into a structured learning path you can follow on your own hardware.