GLM-5.2 (MIT, 1M Context)

Run GLM-5.2, Z.ai's MIT-licensed 1M-context coding flagship, on a big multi-GPU rig from the Clore.ai marketplace

Status (August 2026): Z.ai released GLM-5.2 on 13 June 2026 under the MIT licence — no regional limits, no revenue clause. Weights: zai-org/GLM-5.2 and zai-org/GLM-5.2-FP8. Sparse-attention MoE, roughly 750B total parameters with ~40B active, a solid 1M-token context, and the strongest open coding scores published so far.

GLM-5.1 was already the open model to beat on software engineering. GLM-5.2 raises Terminal-Bench 2.1 from 63.5 to 81.0 and SWE-bench Pro from 58.4 to 62.1, and it does it with an MIT licence while the rest of the frontier moved toward custom terms with revenue gates.

The catch is size. The FP8 checkpoint is 756GB. This is not a single-card model and never will be. What makes it interesting on Clore.ai is that the quantized builds land inside the ceiling of the biggest rigs on the marketplace.

Key Specs

Property	Value
Parameters	~750B total, ~40B active (MoE)
Architecture	78 layers, 256 routed experts, 8 active per token, IndexShare sparse attention, MTP
Context	1,000,000 tokens
License	MIT
Release Date	13 June 2026
Weights	FP8 756GB · Q2_K_XL GGUF 254GB · IQ1_S GGUF 217GB
Primary Tooling	SGLang ≥ 0.5.13.post1, vLLM ≥ 0.23.0, llama.cpp

What is new versus GLM-5.1

• Solid 1M context — not a headline number but a window that holds up across long-horizon work

• IndexShare — one indexer reused across every four sparse-attention layers, cutting per-token FLOPs by 2.9× at 1M context

• Better MTP — speculative-decoding acceptance length up by as much as 20%

• Flexible thinking effort — trade latency against depth per request

• Still MIT — the weights carry no usage restriction

---

Requirements

This model needs a rig, not a card. FP8 at 756GB exceeds every machine on the Clore.ai marketplace. The largest listed boxes are 4× RTX PRO 6000 Blackwell (380GB) and 10–11× RTX 5090 (310–341GB). Quantized GGUF is the only route here; full-precision serving belongs on bare metal.

Build	Size	Where it runs on Clore.ai
UD-IQ1_S	217GB	8× RTX 5090 (248GB) — 47 servers listed at ≥242GB
UD-IQ2_M	239GB	8× RTX 5090, tight
UD-Q2_K_XL	254GB	10× RTX 5090 (310GB) or 4× RTX PRO 6000 (380GB)
UD-Q3_K_M	343GB	4× RTX PRO 6000 (380GB) only — 2 such servers listed
UD-IQ4_XS	365GB	4× RTX PRO 6000, no headroom
FP8 official	756GB	Not on the marketplace → bare metal

Add 10–15% on top of the weight size for the KV cache and activations at working context lengths, and check system RAM: these rigs need enough to stage the weights during load.

---

Deploy with llama.cpp on a multi-GPU rig

The realistic path on marketplace hardware. Rent a 10× RTX 5090 rig (roughly $5.00/hr at the last snapshot) and split the model across all cards:

# ~254GB — leave the rig some headroom
huggingface-cli download unsloth/GLM-5.2-GGUF \
  --include "*UD-Q2_K_XL*" --local-dir /workspace/glm52

llama-server -m /workspace/glm52/*UD-Q2_K_XL*-00001-of-*.gguf \
  --host 0.0.0.0 --port 8080 \
  -ngl 999 --split-mode layer \
  -c 131072 --flash-attn \
  --no-mmap

--no-mmap matters on rented hardware: the weights are pulled fresh onto a disk you do not keep, and mmap-ing a 254GB file over a slow disk turns first-token latency into minutes. Budget 30–60 minutes for the download itself and pick a server with a fast link.

Deploy with vLLM or SGLang (FP8, dedicated hardware)

If you have the capacity — bare metal, or your own cluster — Z.ai supports both engines directly:

# vLLM >= 0.23.0
vllm serve zai-org/GLM-5.2-FP8 \
  --tensor-parallel-size 8 \
  --max-model-len 1000000 \
  --enable-expert-parallel \
  --gpu-memory-utilization 0.92

# SGLang >= 0.5.13.post1
python3 -m sglang.launch_server \
  --model-path zai-org/GLM-5.2-FP8 \
  --tp 8 --context-length 1000000 \
  --speculative-algorithm EAGLE

Ready-made quantized checkpoints for other stacks: nvidia/GLM-5.2-NVFP4 for Blackwell, cyankiwi/GLM-5.2-AWQ-INT4 for AWQ stacks, amd/GLM-5.2-MXFP4 for Instinct.

---

Clore.ai GPU Recommendations

Setup	Total VRAM	Build	Clore.ai cost
8× RTX 5090	248GB	IQ1_S / IQ2_M	~$2.00–3.50/hr
10× RTX 5090	310GB	Q2_K_XL	~$5.00/hr
4× RTX PRO 6000 Blackwell	380GB	Q2_K_XL or Q3_K_M	~$5.00/hr
Dedicated 8× H200	1,128GB	FP8, full 1M ctx	bare metal

Only a handful of rigs this size exist at any moment — 47 servers at ≥242GB, 9 of them free at the last snapshot. Check the marketplace before planning around one.

---

Benchmarks

From Z.ai's own model card (13 June 2026). Independent reproductions exist for the coding suites; treat the reasoning numbers as vendor-reported.

Benchmark	GLM-5.2	GLM-5.1	DeepSeek-V4-Pro (Preview)	MiniMax M3
Terminal-Bench 2.1 (Terminus-2)	81.0	63.5	64	65
Terminal-Bench 2.1 (best harness)	82.7	69	—	—
SWE-bench Pro	62.1	58.4	55.4	59
NL2Repo	48.9	42.7	35.5	42.1
DeepSWE	46.2	18	8	20
HLE	40.5	31	37.7	37
AIME 2026	99.2	95.3	94.6	—
GPQA-Diamond	91.2	86.2	90.1	93

---

Use Cases

• Autonomous coding agents — the Terminal-Bench and DeepSWE jumps are exactly the long-horizon behaviour agents need

• Whole-repository reasoning — 1M context with sparse attention that stays affordable at length

• Migration and refactor work — NL2Repo measures building a repo from a spec, and GLM-5.2 leads the open field

• Self-hosted alternative to closed frontier APIs — MIT means you can ship it inside a product

• Batch evaluation harnesses — rent a big rig by the minute, run the sweep, close the order

---

Troubleshooting

Issue	Fix
Download takes forever	254GB at Q2 — filter the marketplace for high-bandwidth servers and use --include to fetch one quant only
OOM spread across GPUs	--split-mode layer in llama.cpp; in vLLM raise --tensor-parallel-size to the full GPU count
First token takes minutes	Drop --mmap, keep the model on local NVMe, and pre-warm with a short prompt
vLLM rejects the config	Needs ≥ 0.23.0; SGLang needs ≥ 0.5.13.post1
Quality feels off at IQ1	It is a 1-bit build of a 750B model. Move to Q2_K_XL or higher if the rig allows
Rig disappears mid-download	Someone else rented it. Use on-demand rather than spot for long setup work

---

Next Steps

• Predecessor: GLM-5.1 — the April 744B release

• Fits on one card instead: Qwen3.8-27B — Apache 2.0, 15GB at Q4

• Same weight class: MiniMax M3 · Nemotron 3 Ultra · DeepSeek V4

• Rig sizing: GPU Pricing & Availability · Multi-GPU Setup

Links

• GLM-5.2 on Hugging Face · FP8 · GGUF

• Z.ai blog · GLM-5 GitHub

• vLLM recipe · SGLang cookbook

• Rent a GPU: RTX 5090 · Marketplace · Bare metal