Zhipu Open-Sources GLM-5.3 as Unsloth Ships Quantized Build

Episode 1 · Unsloth Releases GGUF Quantized GLM-5.3 Flash (2026-08-27, 2 posts)

Unsloth released a GGUF quantized version of GLM-5.3-Flash under the MIT license, enabling local inference on consumer hardware via llama.cpp and compatibility with transformers pipelines.

• Unsloth releases GGUF quantization of GLM-5.3-Flash model — unsloth · 2026-08-27
• GLM-5.3 Flash Unsloth GGUF Version Now Available — ElementNumber6 · 2026-08-27

Episode 2 · Z.ai Open-Sources GLM-5.3 and GLM-5.3-Flash (2026-08-27, 28 posts)

Around August 28, Z.ai open-sourced GLM-5.3, positioned as the strongest open model for agentic coding and cyber defense. Built on the same base as GLM-5.2, all gains come from scaled post-training: 50% improvement on the internal Z.ai Code Bench and a jump from 4.6% to 28.3% on Terminal-Bench 3.0, reaching open-source SOTA. The company simultaneously released GLM-5.3-Flash (formerly Ox Alpha), a MIT-licensed 320B-parameter MoE with 18B active parameters, the first natively multimodal model in the GLM-5 family, supporting 1M-token context and hybrid attention.

Confirmed

• GLM-5.3 is a MoE with 743B total parameters (some posts say 744B) and 39B active (some say 40B), MIT-licensed, with a 1M-token context window and 128K max output
• Cybersecurity attack-and-defense capabilities emerged from post-training scaling; Ethan Mollick praised the model while urging model cards and red-teaming for open releases
• GLM-5.3-Flash matches Luna on DeepSWE while completing over twice the work at the same budget; its predecessor Ox Alpha runs fully on Chinese AI chips
• Day-0 ecosystem support: vLLM, SGLang (over 500 tok/s on multi-turn agentic tasks across Blackwell/Hopper and AMD MI300), Ollama cloud, Together AI, Baseten, SayGm (backed by KubeTEEAI), Nebius Token Factory (zero data retention)
• Unsloth offers GGUF quants and local guides: 1-bit needs 100GB RAM (71% quality retained), 3-bit needs 128GB (87% retained), plus a 2-bit dynamic scheme; incoai provides the DFlash 2 speculative-decoding drafter (4.4x FP8 throughput), NVFP4 checkpoints, and a TokenRouter deployment

Why it matters

• A 50% coding gain and emergent security skills from post-training alone is strong evidence for the post-training scaling路线
• MIT-licensed weights plus quantization down to 1-bit dramatically lower the barrier to local deployment of a frontier-class model
• Per Together AI, GLM-5.3 approaches Fable 5 on long-horizon tasks within its 1M-token context at far lower cost, offering a new open option for agentic workloads

• Zai open-sources GLM-5.3-Flash: 320B-parameter model running on Chinese chips — ccerrato147 · 2026-08-27
• Run GLM-5.3-Flash locally: 3-bit on 128GB RAM via Unsloth — StefanoGogioso · 2026-08-27
• Zai releases GLM-5.3 Flash: 320B params, 2x efficiency on DeepSWE — togethercompute · 2026-08-28
• ZAI Launches GLM-5.3 Flash: 320B Native Multimodal Model — togethercompute · 2026-08-28
• GLM-5.3-Flash Weights Released: 320B MoE Model with MIT License — arankomatsuzaki · 2026-08-28
• Zhipu's GLM-5.3 goes open-weight, tuned for agentic coding and cyber defense — scaling01 · 2026-08-28
• Ollama adds Z.ai's GLM-5.3-Flash: 18B active params, 1M context, near Opus 4.8 — ollama · 2026-08-28
• zai releases GLM-5.3 open-weight model for agentic coding and defense — zai-org · 2026-08-28
• GLM-5.3 Released: 50% Coding Gain, Emergent Cyber Capabilities — jacek2023 · 2026-08-28
• Zhipu releases open-source GLM-5.3 model alongside Flash variant — 赛博禅心 · 2026-08-28
• Inferact releases GLM-5.3 NVFP4 version, cutting memory to 465GB — vllm_project · 2026-08-28
• GLM-5.3 released as open-weight model for agentic coding — shaunralston · 2026-08-28
• GLM-5.3 Open Weights: SGLang Reports Over 500 tok/s on Agentic Workloads — BanghuaZ · 2026-08-28
• Zai Releases Open-Weight GLM-5.3 Coding Model Amid Safety Criticism — dhadfieldmenell · 2026-08-28
• GLM 5.3 open weights arrive; DFlash 2 speculative decoding hits 4.4x FP8 throughput — gan_chuang · 2026-08-29
• GLM 5.3 open weights released with NVFP4 checkpoint achieving 4.4x throughput — songhan_mit · 2026-08-29
• GLM-5.3 Released: MIT Licensed, Focused on Agent Coding and Security — baseten · 2026-08-29
• GLM-5.3 open weights released; Unsloth enables local run with 83% size reduction — danielhanchen · 2026-08-29
• GLM-5.3 open-weight release: 50% coding boost, emergent cyber capabilities — gharik · 2026-08-29
• GLM-5.3 open-sourced with 1M context window, vLLM integration available — AccBalanced · 2026-08-29

8 more related posts →