glm47-flash-blockwise-fp8 by RadixArk: Open-Weight Model

Runs On

What it takes to serve glm47-flash-blockwise-fp8 (31.2B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.

Precision	Weights	Memory needed	Cheapest setup	Per hour	Also fits
16-bit	62.4 GB	74.9 GB	1x MI300X (192 GB) Vultr	$1.85	1x H100 $1.99 · 1x MI325X $2.00
8-bit	31.2 GB	37.5 GB	1x MI300X (192 GB) Vultr	$1.85	1x H100 $1.99 · 1x MI325X $2.00
4-bit	15.6 GB	18.7 GB	1x MI300X (192 GB) Vultr	$1.85	1x H100 $1.99 · 1x MI325X $2.00

Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 16, 2026.

glm47-flash-blockwise-fp8 on every accelerator the SAVRN Index prices, at every precision

Model Card

By RadixArk, published under mit, revision 4b16c299a3a6.

Blockwise FP8 quantization of zai-org/GLM-4.7-Flash: e4m3 weights with one scale per 128×128 block and dynamic activation scaling, about half the size of the BF16 original (30 GiB vs 58 GiB). See the original model card for model details. gsm8k (5-shot, full test set) on SGLang with MTP speculative decoding: 0.809 for FP8 vs 0.819 for BF16; average accept length 2.41 vs 2.43. Attention tensor parallelism must be 1 or 2: at 4, each rank's kvbproj shard (2240 rows) is not a multiple of the 128-row block. For more GPUs, use DP attention with expert parallelism, for example --tp-size 4 --dp-size 4 --enable-dp-attention --ep-size 4 --moe-a2a-backend deepep --cuda-graph-max-bs-decode 128. MIT…

Read RadixArk's full model card

Blockwise FP8 quantization of zai-org/GLM-4.7-Flash: e4m3 weights with one scale per 128×128 block and dynamic activation scaling, about half the size of the BF16 original (30 GiB vs 58 GiB). See the original model card for model details.

Quantization

FP8: the linear layers of attention, the dense MLP, and all routed and shared experts, including the MTP layer.
BF16: embeddings, lm_head, norms, and the MoE router.

Evaluation

gsm8k (5-shot, full test set) on SGLang with MTP speculative decoding: 0.809 for FP8 vs 0.819 for BF16; average accept length 2.41 vs 2.43.

Usage (SGLang)

python -m sglang.launch_server --model-path RadixArk/glm47-flash-blockwise-fp8 --tp-size 2 \
  --speculative-algorithm EAGLE --speculative-num-steps 2 \
  --speculative-eagle-topk 1 --speculative-num-draft-tokens 3

Attention tensor parallelism must be 1 or 2: at 4, each rank's kv_b_proj shard (2240 rows) is not a multiple of the 128-row block. For more GPUs, use DP attention with expert parallelism, for example --tp-size 4 --dp-size 4 --enable-dp-attention --ep-size 4 --moe-a2a-backend deepep --cuda-graph-max-bs-decode 128.

License

MIT, same as the original model by Z.ai.

Configuration

Architecture: Glm4MoeLiteForCausalLM
Context length (tokens): 202,752
Layers: 47
Hidden size: 2,048
Feed-forward size: 10,240
Attention heads: 20
Key/value heads: 20
Vocabulary size: 154,880
Routed experts: 64
Experts active per token: 4
RoPE base: 1,000,000
Model type: glm4_moe_lite
Quantization: fp8

Identity and Version

Repository: RadixArk/glm47-flash-blockwise-fp8
Publisher: RadixArk
Task: Text generation
Modality: Text
Library: transformers
Parameters: 31.2B parameters
Languages: en, zh
Revision: 4b16c299a3a61346e6160be4bc967d11726b6a12
First published: 2026-09-28
Last updated: 2026-09-28

Files and Weights

56 files, 32.5 GB in total. The weights are 48 files totalling 32.5 GB in safetensors.

Weights48 files · 32.5 GB

Configuration3 files · 1.8 MB

Tokenizer2 files · 20.2 MB

Documentation1 file · 1.5 KB

Other1 file · 3.1 KB

Repository1 file · 1.6 KB

Every file

File	Type	Size	SHA-256
model-00001-of-00048.safetensors	Weights	1.4 GB	cc7a92dcd465
model-00002-of-00048.safetensors	Weights	635.7 MB	2a6674a47990
model-00003-of-00048.safetensors	Weights	635.7 MB	ae2bfbe0efc6
model-00004-of-00048.safetensors	Weights	635.7 MB	cbe9713b54d5
model-00005-of-00048.safetensors	Weights	635.7 MB	3627cf93d63d
model-00006-of-00048.safetensors	Weights	635.7 MB	3b830c2c63c1
model-00007-of-00048.safetensors	Weights	635.7 MB	4b0ee4ff5ce5
model-00008-of-00048.safetensors	Weights	635.7 MB	ab61a8535815
model-00009-of-00048.safetensors	Weights	635.7 MB	9d99939e0e2a
model-00010-of-00048.safetensors	Weights	635.7 MB	757903c8f8ba
model-00011-of-00048.safetensors	Weights	635.7 MB	00aadfad9fe5
model-00012-of-00048.safetensors	Weights	635.7 MB	837101f11a69
model-00013-of-00048.safetensors	Weights	635.7 MB	ccf9e8df80e5
model-00014-of-00048.safetensors	Weights	635.7 MB	77798f441b04
model-00015-of-00048.safetensors	Weights	635.7 MB	f87ff3d99430
model-00016-of-00048.safetensors	Weights	635.7 MB	cf1360f45aed
model-00017-of-00048.safetensors	Weights	635.7 MB	0edbac1564cb
model-00018-of-00048.safetensors	Weights	635.7 MB	7e82a9fcd3fc
model-00019-of-00048.safetensors	Weights	635.7 MB	7236ce32bc61
model-00020-of-00048.safetensors	Weights	635.7 MB	d9013b1a21ca
model-00021-of-00048.safetensors	Weights	635.7 MB	b15d369551fe
model-00022-of-00048.safetensors	Weights	635.7 MB	c21bb5702e56
model-00023-of-00048.safetensors	Weights	635.7 MB	2d9501fd0b40
model-00024-of-00048.safetensors	Weights	635.7 MB	e19c0a6ed91d
model-00025-of-00048.safetensors	Weights	635.7 MB	9eb51158eed3
model-00026-of-00048.safetensors	Weights	635.7 MB	1eb377d4bdfd
model-00027-of-00048.safetensors	Weights	635.7 MB	30e8f02c3a05
model-00028-of-00048.safetensors	Weights	635.7 MB	d608ddc12a24
model-00029-of-00048.safetensors	Weights	635.7 MB	c6c52b46082a
model-00030-of-00048.safetensors	Weights	635.7 MB	b85f4ad118d6
model-00031-of-00048.safetensors	Weights	635.7 MB	2f78f31cda2e
model-00032-of-00048.safetensors	Weights	635.7 MB	ccb627140550
model-00033-of-00048.safetensors	Weights	635.7 MB	d7bf7a6abc1c
model-00034-of-00048.safetensors	Weights	635.7 MB	7ab6eb985f51
model-00035-of-00048.safetensors	Weights	635.7 MB	fd0e25a4078f
model-00036-of-00048.safetensors	Weights	635.7 MB	cf0e9a5daf29
model-00037-of-00048.safetensors	Weights	635.7 MB	aa6044df8a6a
model-00038-of-00048.safetensors	Weights	635.7 MB	f99e5a8fcc15
model-00039-of-00048.safetensors	Weights	635.7 MB	991f9bac8fbc
model-00040-of-00048.safetensors	Weights	635.7 MB	2226516bcea8
model-00041-of-00048.safetensors	Weights	635.7 MB	39c5e6b14ef8
model-00042-of-00048.safetensors	Weights	635.7 MB	e2d1c6bb31bf
model-00043-of-00048.safetensors	Weights	635.7 MB	d52e111a2b78
model-00044-of-00048.safetensors	Weights	635.7 MB	e4836ad58fc1
model-00045-of-00048.safetensors	Weights	635.7 MB	4ad8f120374c
model-00046-of-00048.safetensors	Weights	635.7 MB	87c54b303e39
model-00047-of-00048.safetensors	Weights	1.9 GB	7ef4222dda88
model-00048-of-00048.safetensors	Weights	652.4 MB	5fd5c83d19c1
config.json	Configuration	14.9 KB	—
generation_config.json	Configuration	181 B	—
model.safetensors.index.json	Configuration	1.8 MB	—
README.md	Documentation	1.5 KB	—
chat_template.jinja	Other	3.1 KB	—
.gitattributes	Repository	1.6 KB	—
tokenizer.json	Tokenizer	20.2 MB	19e773648cb4
tokenizer_config.json	Tokenizer	7.2 KB	—

License and Download

License: mit
Access: Open weights, no gate
Download size: 32.5 GB

Download from RadixArk

Released by RadixArk through its official repository on Hugging Face. Read the license.

Built From

Derived from zai-org/GLM-4.7-Flash
Quantized from zai-org/GLM-4.7-Flash

Memory Requirements

Precision	Weights in memory
As published	32.5 GB
16-bit	62.4 GB
8-bit	31.2 GB
4-bit	15.6 GB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Questions About glm47-flash-blockwise-fp8

How much GPU memory does glm47-flash-blockwise-fp8 need?

About 74.9 GB at 16-bit and 18.7 GB at 4-bit: the weights (31.2B parameters) plus a working margin. A long context needs more.

What is the cheapest GPU to run glm47-flash-blockwise-fp8 on?

At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.

Can I use glm47-flash-blockwise-fp8 commercially?

Yes. glm47-flash-blockwise-fp8 is released under MIT License. The MIT License is a short permissive license. It permits commercial use, modification and redistribution, provided the copyright notice and permission notice are included.

What is glm47-flash-blockwise-fp8's context length?

202,752 tokens, from the maximum position embeddings in its published configuration.

Similar Models

Model · Text generation

GLM-4.7-Flash

Z.ai

Join our Discord community. Check out the GLM-4.7 technical blog, technical report(GLM-4.5). Use GLM-4.7-Flash API services on Z.ai API Platform. One click to GLM-4.7. GLM-4.7-Flash is a 30B-A3B MoE model. As the strongest model in the 30B class, GLM-4.7-Flash offers a new option for lightweight deployment that balances performance and efficiency. Default Settings (Most Tasks) For multi-turn agentic tasks (τ²-Bench and Terminal Bench 2), please turn on Preserved Thinking mode. Terminal Bench, SWE Bench Verified τ^2-Bench For τ^2-Bench evaluation, we added an additional prompt to the Retail and Telecom user interaction to avoid failure modes caused by users ending the interaction…

Open weights mit 31.2B parameters 202,752 tokens transformers

View model

Model · Text generation

OTel-2.0-LLM-31B-IT

Farbod Tavakkoli

OTel-2.0-LLM-31B-IT is a telecom-specialized instruction model post-trained from Gemma 4 31B-IT on approximately 440 billion telecom training tokens. It is the first release in the OTel 2.0 family and is designed to support telco-grade AI workflows across network operations, standards interpretation, product development, network configuration assistance, RAG, and telecom-specific question answering. OTel 2.0 extends the original OTel effort from a RAG-oriented telecom fine-tuning release into a larger domain-adapted training program. The model was trained from a much larger standards and telecom corpus, with new data preparation coverage for direct telecom QnA, abstention, RAG…

Open weights apache-2.0 31.3B parameters 262,144 tokens transformers

View model

Model · Text generation

Gemma-4-Writers-31B

Ateron

"Less Gemma - more human" I had a lot of fresh and creative experience with both ConicCat tunes, but they had some instability issues, that I hopefully fixed there.

Open weights apache-2.0 31.3B parameters 262,144 tokens

View model

Model · Text generation

Qwen3-VL-30B-A3B-Instruct-AWQ

QuantTrio

As of 2025-10-08, create a fresh Python environment and run: For more details, refer to vLLM Official Qwen3-VL Guide Meet Qwen3-VL — the most powerful vision-language model in the Qwen series to date. This generation delivers comprehensive upgrades across the board: superior text understanding & generation, deeper visual perception & reasoning, extended context length, enhanced spatial and video dynamics comprehension, and stronger agent interaction capabilities. Available in Dense and MoE architectures that scale from edge to cloud, with Instruct and reasoning‑enhanced Thinking editions for flexible, on‑demand deployment. Text Understanding on par with pure LLMs: Seamless text–vision…

Open weights apache-2.0 31.1B parameters 262,144 tokens transformers

View model

Model · Text generation

nemotron-30b-a3b-gptq-component-jang-g64-lora-mtp

Roman Romenskyi

Open weights 31.6B parameters 262,144 tokens mlx

View model

Model · Text generation

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16

NVIDIA

September 2025 \- December 2025 The post-training data has a cutoff date of November 28, 2025\. The pre-training data has a cutoff date of June 25, 2025\. Nemotron-3-Nano-30B-A3B-BF16 is a large language model (LLM) trained from scratch by NVIDIA, and designed as a unified model for both reasoning and non-reasoning tasks. It responds to user queries and tasks by first generating a reasoning trace and then concluding with a final response. The model's reasoning capabilities can be configured through a flag in the chat template. If the user prefers the model to provide its final answer without intermediate reasoning traces, it can be configured to do so, albeit with a slight decrease in…

Open weights other 31.6B parameters 262,144 tokens transformers

View model