Zixuan Li@ZixuanLi_
The official GLM-5.2 NVFP4 from NVIDIA is now available. Curious how it compares to other quantizations. https://huggingface.co/nvidia/GLM-5.2-NVFP4
11:17 PM · Jun 25, 2026 · 40K Views
NVIDIA releases an NVFP4-quantized checkpoint of the 744B GLM-5.2 MoE model with day-zero SGLang support
The model uses 40 billion active parameters for coding.
295821914554.3K
Sentiment
Positive users hail NVIDIA's GLM-5.2 NVFP4 quantized model as amazing and AGI-like for slashing compute costs, while negative users criticize it for hurting multilingual performance and precision due to outliers.
Pos
33.4%
Neg
66.6%
7 comments with sentiment.
Cluster Engagement
Digg Deeper
No Digg Deeper questions have been answered for this story yet.
Related links
Posts from X
Most Activity
Most Activity
VIEWS901
Abhi Prajapati@abhip05
@ZixuanLi_ Are GLM models currently under too much overload as I'm getting 429 since morning ?
13hViews 901Likes 1
BOOKMARKS1
Jonathan Washburn@JonWashburn
We run the FP8 version on an 8×B200 box for our autonomous Lean/proof fleet, so we tested whether to switch.
The cutover was flawless: live in ~6 min, correct FP4 kernel, +57% KV cache, ~300GB HBM freed, quality within noise of FP8 on NVIDIA's own evals. On paper, a clear upgrade.
We rolled it back. Our fleet runs ~16 concurrent with 0 queue: it's decode-bound, not memory-bound. NVFP4 has no MTP speculative decoding, so single-stream fell ~24%. The freed memory bought nothing we use. Lesson: profile your real bottleneck, not the spec sheet.
7hViews 34Likes 1Bookmarks 1
LIKES11REPLIES1
Zixuan Li@ZixuanLi_
@abhip05 We're seeing increased load and are working on scaling up to reduce 429 errors. Thanks for your patience.
13hViews 825Likes 11
RETWEETS7
LMSYS Org@lmsysorg
🎉 NVIDIA just released an NVFP4 checkpoint of GLM-5.2 from @Zai_org, a 744B MoE (40B active) for reasoning & coding. Day-0 support is live in SGLang! 🤝 @nvidia
> NVFP4 quantization via NVIDIA Model Optimizer: frontier-class reasoning at a fraction of the memory > Sparse attention with IndexShare indexer for efficient long-context > Ready to serve on Blackwell / Grace Blackwell, run it now with SGLang!
13hViews 7KLikes 71Bookmarks 28
LMSYS Org@lmsysorg
Cookbook: https://docs.sglang.io/cookbook/autoregressive/GLM/GLM-5.2
13hViews 481Likes 8Bookmarks 1
Niko(la)@nibzard
@ZixuanLi_ glm 5.2 in ultracode mode is agi
12hViews 454Likes 1Bookmarks 1
sunny madra@sundeep
https://huggingface.co/nvidia/GLM-5.2-NVFP4
Barathwaj Anandan@BarathAnandan7
Good news. We cooked!
@NVIDIAAI GLM 5.2 NVFP4 is out for anyone who's been waiting on a quality quant.
Size ~465GB.
Link below.
Blackwell go brrr
13hViews 7.6KLikes 24Bookmarks 2
Mehdi Ataei@AtaeiMe
@ZixuanLi_ 💚 amazing model
14hViews 616Likes 2
The AI Therapist ⚡@TheAIShrink
@ZixuanLi_ Fp4 keeps extreme values int4 can't. costs precision. glm-5.2 activations have outliers. test on your data, not the paper.
13hViews 415Likes 1
Rasmus@synquid
@ZixuanLi_ Their quantisations usually really hurt multilingual performance in my experience. You could do it better :)
14hViews 622
Learner@Learner81105683
@ZixuanLi_ Is this PTQ or QAT?
13hViews 573
Abhi Prajapati@abhip05
@ZixuanLi_ Thank you for the quick response 🙌
13hViews 202
pablo (nvfp4)@Pablogomez3
@lmsysorg @Zai_org @nvidia I’ll miss Lukealonso image 🥺
6hViews 77
Moshe Malawach@MosheMalawach
@ZixuanLi_ Testing it right now... attention is handled better than in others. Some outputs are weird though, need to look into why.
9hViews 65
Adel Bucetta@adelbucetta
@ZixuanLi_ nvfp4 doesn't just compress models, it rewrites the physics of compute costs
13hViews 62
kerr@hepin1989
@ZixuanLi_ 为世界贡献中国方案、中国智慧。
9hViews 52
Moshe Malawach@MosheMalawach
@TheAIShrink @ZixuanLi_ Nvidia's NVFP4 keeps all attention + shared-experts in high precision.
9hViews 10