GLM 5.5 might come in versions with sizes you can run at home!!! A leak shows Zhipu training 32B, 122B and 256B models to merge into the 5.5 base. At 4bit the 32B is ~16GB. One GPU🤞

GLM-5.3 is now available in the United States on Alibaba Cloud Model Studio. Take on complex challenges by harnessing advanced reasoning, coding, and agentic capabilities. Build with GLM-5.3 on Model Studio: #AlibabaCloud# #ModelStudio# #GLM# #AIModels#

GLM just helped uncover 23 Windows vulnerabilities, all acknowledged by Microsoft MSRC. Beyond open-source projects, commercial software security is also being improved through GLM-powered independent vulnerability discovery and responsible disclosure.

GLM 5.3 Flash on 4 DGX Sparks with TensorFold 🔥 vs my vLLM TP4 stack, same harness: ⚡ Code 127 vs 120 tok/s ⚡ JSON 133 vs 120 🛠️ Tool calls 93 vs 68 (+37%) 📈 Faster at every concurrency, C1 to C6 🚀 500K-token prompt in 191s (vLLM 202s) 🧠 8 streams, 500K context each, vision on

GLM-5.3-Flash running faster on my 3090 thanks to Project Maya from @PeasantSmith ! After some tweaking (PR coming soon) i'm able to run my Aj-IQ2_XXS mixed quant with Maya. - ~20t/s avg - low 200t/s prefill. Without MTP btw. I'm working on some pr to improve speed, at this rate even GLM-5.3-flash will be usable in my 3090 before end of this month. Checkout Maya: Get my quant:

GLM-5.3-Flash on one DGX Spark at 33 tok/s. Plus CYBER-FROST at 55 tok/s on code. Two impressive community one-box recipes from this week. I re-ran both with the authors' own benchmark method. Both hold up. RED-SNOW 5.3 Flash, 2.49 bpw: • long outputs (1000+ tokens): 32.95 tok/s median. README says 34.4. • prose: 25.4 tok/s. README says ~24. • MTP acceptance 89.5%. CYBER-FROST 3.8, 3.87 bpw, with the repo's own bench script: • code: 54.7 tok/s. DEPLOY.md says 53.2. • devops: 49.0 tok/s. DEPLOY.md says 45.8. • prose: 46.9 tok/s. DEPLOY.md says 46.5. So both land within 7% of what the repos say. Honest limits: • My box runs with a 2200 MHz clock cap. That can cost a few percent against an uncapped box. • Engine pins differ too: I ran exllamav3 fork 24f0dec + TabbyAPI main; the repo and yume ran 047ce72 + TabbyAPI 7a524a9. Part of the gap below may come from that. • @yume_arasaki got 60.7 tok/s on the code suite. I got 54.7, best single sample 60.6. • 2.49 bpw is a tight squeeze: the card reports 90% top-1 agreement with BF16. • 3 runs each. Both are de-refused security finetunes (RED-SNOW is an abliterated GLM-5.3-Flash). Recipes @WescheNex1q + @ViC305, finetunes @Blackfrost_AI.

GLM-5.3-Flash vs Qwen3.8 Flash-Next: Both on 2× DGX Spark, same 80-test Spark-Bench: -GLM 84.8 -Qwen 79.4 (vLLM) / 77.9 (TensorFold) GLM wins most categories. The biggest gap is long context. Where each one wins • Long context: 70.8 vs 21.6–27.9. GLM found the planted fact on the hard needle test, Qwen missed it on both engines. • Code: 90.2 vs 85.7–87.1 • Visual: 79.8 vs 71.4–75.0 • Tool use: 78.2 vs 58.2–75.6 • Agentic: 78.5 vs 67.1–75.1 Takeaway: Qwen follows instructions and formats more precisely. GLM is the stronger generalist: it codes, builds and retrieves better. Same 80 tests, uncapped, 2 repeats.