GLM-5.3-Flash

Linas's Newsletter reports on Z.ai's GLM-5.3-Flash, the model that launched anonymously on OpenRouter and OpenCode on August 20, 2026 as "Ox Alpha" before becoming the most-used model on the platform at about 23 trillion tokens processed. Z.ai's identity was confirmed by Bloomberg on August 26, 2026, with weights landi The report explicitly attributes GLM-5.3-Flash to Z.ai (the Beijing lab that operates internationally as Z.ai / Zhipu AI) and cites the model as 320 billion parameters with 18 billion active per token, marking the first natively multimodal release in the GLM-5 series. Stripe CEO Patrick Collison called the anonymous pr

FelloAI's coverage of GLM-5.3-Flash details Z.ai's August 26, 2026 release of a 320B-parameter mixture-of-experts model that activates 18B per token, built on a new base rather than post-trained on GLM-5.2. MIT-licensed weights landed on Hugging Face on day one, and the post confirms that GLM-5.3-Flash is the model pre The post contrasts pricing — $0.15 per million input tokens and $0.50 output for Flash versus $1.40 and $4.40 for the flagship GLM 5.3 — and notes that GLM-5.3's promised 744B weights have not yet appeared, with no GLM-5.3 repository currently on Z.ai's Hugging Face. Architecture is summarized as a hybrid sparse and li

AI Profit Boardroom's guide confirms Z.ai released GLM 5.3 Flash on August 26, 2026 as a 320B-parameter model with 18B active per token, natively multimodal with a one-million-token context window, and released under the MIT license. The article documents that an anonymous model appeared on a third-party API platform o Technical specifics cited from Z.ai's announcement and the OpenRouter listing describe a mixture-of-experts model with hybrid sparse and linear attention, text/image/video input with text output, a listed context window of 1,310,720 tokens with up to 131,072 completion tokens, plus tool calling and JSON-formatted outpu

Greek Ai's Medium write-up presents Z.ai's GLM-5.3-Flash, released on August 26, 2026, as the newest member of the GLM-5 family and the first natively multimodal model in the series. The model has 320B total parameters with about 18B active, supports multimodal inputs and a context window of up to one million tokens, a Framing the release as competition around intelligence-per-compute rather than raw scale, the post attributes GLM-5.3-Flash to Z.ai (the company formerly known as Zhipu AI) and lists its capabilities as text understanding, image understanding, and video understanding. Z.ai's preview strategy — releasing Ox Alpha throug

GLM-5.3-Flash is a sparse mixture-of-experts model from Z.ai with 320B total parameters and 18B active per token, released on August 26, 2026 in FP8 and BF16 checkpoints under an MIT license, with weights published on Hugging Face and API pricing set at $0.15 per million input tokens. According to the LumaDock explaine On Z.ai's launch table, GLM-5.3-Flash posts 63.4 Pass@1 on DeepSWE v1.1 (versus 46.2 for GLM-5.2), 84.3 on Terminal-Bench 2.1 (Claude Opus 4.8 at 85.0 and GPT-5.6 Terra at 87.4), and 55.3 on Humanity's Last Exam with tools, and Z.ai claims a hybrid linear-plus-sparse attention design that delivers roughly 3x less atten

Ampere.sh compares GLM 5.3 and GLM 5.3 Flash, attributing both to Z.ai and reporting Artificial Analysis Intelligence Index scores of 60 for GLM 5.3 and 57 for GLM 5.3 Flash. GLM 5.3 Flash is described as dramatically cheaper, natively multimodal, and delivering close performance at much lower compute, while GLM 5.3 is The spec table pins GLM 5.3 Flash at 320B total / 18B active parameters, 1M context, native image input, native multimodal capability, tool use, $0.15/M input and $0.50/M output pricing, and 3× usable quota in Z.ai's Coding Plan versus 1× for the flagship. GLM 5.3 is listed at 753B/40B, $1.40/$4.40 pricing, no native i

Local AI Zone's technical deep dive opens with Z.ai's quiet August 26, 2026 release of GLM-5.3-Flash, a 320B-parameter mixture-of-experts model that activates 18B per token, runs natively in FP8, and supports a 1,048,576-token context window. The post confirms the Ox Alpha identity — for about a week the anonymous prev The deep dive details the hybrid sparse and linear attention design, benchmark deltas versus GLM-4.6 / GLM-5.2 / GLM-5.3 (beating GLM-5.2 across six coding/agentic benchmarks at roughly one-tenth the price and approaching Claude Opus 4.8 on long-horizon agent tasks), training pipeline notes, inference characteristics,

This Hugging Face discussion on the official zai-org/GLM-5.3-Flash model repository documents a merged community PR that adds evaluation results extracted from the model's published benchmark chart. The PR lists Terminal-Bench 2.1 at 84.3, DeepSWE at 63.4, and Humanity's Last Exam at 55.3 as benchmarks successfully reg As the closest item to an official channel in the candidate set (hosted on Z.ai's own Hugging Face organization page), this PR provides concrete, traceable benchmark provenance directly tied to the GLM-5.3-Flash model card. The benchmark scores were read visually from an embedded chart rather than re-run via HF Jobs wi