GLM-5.3-Flash

The Substack guide reports that on August 20, 2026 an unattributed listing called "Ox Alpha" appeared on OpenRouter and OpenCode, offering a one-million-token context window with text, image, and video input. Within six days it became the most-used model on OpenRouter, processing roughly 23 trillion tokens at about 2.3 According to the article, GLM-5.3-Flash ships with 320B total / 18B active parameters, is the first natively multimodal model in the GLM-5 series, and is the cheapest model at its Artificial Analysis Intelligence Index score. Weights landed on Hugging Face under an MIT license the same evening, and Zhipu's Hong Kong sh

A Hacker News thread (805 points, 281 comments) discussed Z.ai's GLM-5.3 open-weight release, linking to the Hugging Face repository and z.ai's official blog at z.ai/blog/glm-5.3. Commenters positioned GLM-5.3 as a competitive open-weights option relative to DeepSeek Flash and Kimi, noting it was less restrictive on cy Community commentary covered local inference economics, with users debating whether buying high-end hardware (e.g., a Mac M5 Ultra with 512GB unified memory for 4-bit quantized inference) is justified given cloud pricing and rapidly evolving model landscapes. Several commenters concluded that cloud deployment beats loc

Z.ai's GLM-5.3-Flash launched on August 26, 2026, as an open-weight model combining multimodal AI, a 1M-token context window, agentic coding capabilities, and dramatically lower inference costs than frontier alternatives. The model has 320 billion total parameters but only about 18 billion active parameters per token, The article frames GLM-5.3-Flash within the broader shift in AI model competition, where the race is no longer about building the largest model but about maximizing intelligence per unit of compute actually used. The Flash variant is described as offering multimodal inputs, a 1-million-token context window, and agentic

Z.ai's GLM-5.3-Flash is described as the cost-optimized sibling to the flagship GLM-5.3, released under the codename "Ox Alpha" before its public reveal. The model uses a 320B-aggregate / 18B-active MoE configuration, supports a 1M-token context window, and is natively multimodal. DataCamp reports a Terminal-Bench 2.1 The article frames GLM-5.3-Flash within the broader budget-tier landscape, noting that Qwen3.8-Flash-Next launched the same day, intensifying competition in the cost-optimized segment. Artificial Analysis rates the Flash variant at an intelligence index of 57 versus 60 for GLM-5.3, confirming the trade-off between capa

A direct head-to-head comparison from Ampere.sh confirms GLM 5.3 Flash as a Z.ai model with 320B total and 18B active parameters, a 1M-token context window, native image input, native multimodality, reasoning support, and tool use, with public weights available under MIT at launch. GLM 5.3, by contrast, uses a 753B-tot On benchmarks and economics, Artificial Analysis gave GLM 5.3 Flash an Intelligence Index of 57 versus 60 for GLM 5.3, with output speed of 49 tokens/second (Flash) versus 85 tokens/second (GLM 5.3), and time-to-first-token of 1.46s versus 1.57s. API pricing for Flash ($0.15/M input, $0.50/M output) compared to GLM 5.3

The LLM Stats benchmark aggregator ranks GLM-5.3-Flash at number 17 overall with a composite LLM Stats Score of 50.3 at a blended price of approximately $0.17 per million tokens. In per-category capability tiers, the model ranks in the top 10% for Tool Calling at 7 of 193, holds an average position in Reasoning at 16 o Specific benchmark results include a GDPval-AA Elo score of 1773 out of 3000 from Artificial Analysis (ranked 2nd), a CharXiv-R score of 0.89 on the reasoning component from z.ai, and a Terminal-Bench 2.1 score also visible in the excerpt. The Quality Tracker shows a +0.44σ improvement over the community voting window.

This technical deep dive describes GLM-5.3-Flash as a 320B-total / 18B-active mixture-of-experts model running natively in FP8 with a 1,048,576-token context window, the first natively multimodal model in the GLM-5 series and the first built on a new base since GLM-5.2. The article details the model's hybrid sparse and On benchmarks, GLM-5.3-Flash is reported to beat GLM-5.2 across six coding and agentic benchmarks at roughly one-tenth the price and approach Claude Opus 4.8 on long-horizon agent tasks. The piece provides a changelog versus GLM-4.6, GLM-5.2, and GLM-5.3, covers the training pipeline, and walks through the developer-fa

Z.ai released GLM 5.3 Flash on August 26, 2026, as a 320-billion-parameter mixture-of-experts model that activates only 18 billion parameters per token, built on a newly trained base rather than post-trained on GLM 5.2. It is the first natively multimodal model in the GLM-5 series, combining text, image, and video unde On pricing, Z.ai set GLM 5.3 Flash at $0.15 per million input tokens and $0.50 per million output tokens, roughly one-ninth the cost of GLM 5.3's $1.40 input / $4.40 output rates, with API pricing positioned as a leading capability-to-cost ratio on Artificial Analysis's Intelligence Index. Independent Artificial Analys