GLM-5.3 Flash Abliterated

Abliterated is Abliteration's hosted version of Z.ai's open-weight GLM-5.3 with its refusal behavior removed through abliteration, which finds the activation directions that produce refusals and removes them from the weights. It keeps GLM-5.3's coding and agentic strengths, is text-only, and has a 1M-token context window. Reasoning always runs and cannot be turned off (effort: low, high, or max). It supports tool calling, JSON mode and JSON Schema output, streaming, and web search, through OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages request shapes. It is US-hosted in FP8 with zero prompt retention. Image and video input are rejected.

GLM-5.3-Flash is Z.ai’s flagship natively multimodal GLM-5-series model with a 1M-token context window, accepting text plus images and video and producing text outputs. It targets fast, high-quality general and coding assistance with strong reasoning at “flash” latency and is released as open weights under the MIT license.

GLM-5.3-FlashX is Z.ai's high-speed serving variant of GLM-5.3-Flash, delivering inference speeds of up to 200 tokens/s for faster responses. It inherits Flash's 320B-total/18B-active MoE design, native multimodality, hybrid sparse and linear attention architecture, and 1M-token context window. FlashX runs on a purpose-built serving stack (SGLang-based inference engine, W8A8 quantization, mixed INT8/FP8/BF16 cache quantization, and encode-prefill-decode disaggregation), reporting roughly 3x end-to-end throughput over Z.ai's initial baseline on the same hardware. It is suited for interactive coding agents, browser/computer-use loops, visual coding, and other multi-step workflows where latency compounds.