GLM 5.3 Flash

Z.ai announced GLM-5.3-Flash on August 26, 2026, framing it as the first natively multimodal model in the GLM-5 series with 320B total parameters but only 18B active per token. According to the launch blog, the model starts from a newly trained base, introduces a hybrid sparse-plus-linear attention architecture for the On coding and agentic benchmarks, Z.ai reports GLM-5.3-Flash scoring 63.4 on DeepSWE v1.1 (versus 46.2 for GLM-5.2) and 48.8 on AutomationBench (versus 26.2 for GLM-5.2), and reaching 29.0 versus Claude Opus 4.8's 29.5 on the in-house Z.ai Code Bench v1.0 at max effort. The model is positioned at roughly one-tenth the

Yotta Labs' comparison confirms GLM-5.3-Flash launched on August 26, 2026 as a 320B-total / 18B-active MoE with MIT-licensed weights on Hugging Face, native text, image, and video input, and a checkpoint of about 306 GiB. On the one benchmark both models report, DeepSWE v1.1, Flash scores 63.4 while the GLM-5.3 flagship scores 66.9, both per Z.ai. Z.ai lists Flash pricing around $0.15 in and $0.50 out per million tokens. The piece positions Flash for most workloads and recommends the larger GLM-5.3 flagship only for long-running agent and engineering work, noting Flash as text-image-video capable while GLM-5.3 is text-only. It also reports that GLM-5.3 weights were released at the end of August 2026 under a custom glm-5.3 license, requiring about 744 GB in FP8 on an 8x H200 node for self-hosting.

A detailed third-party breakdown confirms GLM-5.3-Flash launched on August 26, 2026, as a 320B mixture-of-experts activating 18B parameters per token, with MIT-licensed weights on Hugging Face from day one. It is the first natively multimodal model in the GLM-5 series and the first built on a new base since GLM-5.2, wi Pricing is set at $0.15 per million input tokens and $0.50 output, compared with $1.40 and $4.40 for the full GLM-5.3 flagship. Notably, GLM-5.3's own 744B open weights remain missing — Z.ai's Hugging Face organization has no GLM-5.3 repository — making GLM-5.3-Flash the only open-weight model developers can access tod

An NVIDIA DGX Spark community thread documents local inference testing of GLM-5.3-Flash across multiple quantization recipes. Users evaluated NVFP4-lab (mixed NVFP4/MXFP8-MTP), RedHatAI W4A4 compressed-tensors, and NVFP4-A16 variants, measuring decode tokens-per-second, time-to-first-token, and quality scores on tool-e The thread reveals a practical vLLM requirement: the framework needs a fix in modelopt.py because MTP layers are stored under language_model.model.layers.45, an issue already addressed in DeepSeek-V4-Flash support. Community members tested RedHat quantizations and discussed ongoing optimizations, with an "update coming

A Medium technical analysis from Greek Ai describes GLM-5.3-Flash as Z.ai's newest open-weight release built around the principle of frontier-level capabilities without frontier-level inference costs. It confirms the 320B total / 18B active parameter architecture, native multimodal support, 1M-token context window, and The piece emphasizes that the GLM-5 family competition is shifting from raw parameter count to intelligence-per-compute, with GLM-5.3-Flash delivering dramatically lower inference costs. The model combines multimodal inputs, agentic capabilities, and a 1M-token context into a single architecture, making it suitable for

An Ampere comparison piece positions GLM-5.3-Flash against the full GLM-5.3 flagship. In independent Artificial Analysis testing, GLM-5.3 scores 60 on the Intelligence Index versus 57 for GLM-5.3-Flash, but Flash costs roughly one-ninth as much at normal API pricing and activates only 18B parameters versus 40B for GLM- Specs confirm GLM-5.3-Flash at 320B total / 18B active parameters with a 1M context window, while GLM-5.3 runs at 753B/40B active. GLM-5.3-Flash pricing is $0.15/M input and $0.50/M output versus $1.40/M and $4.40/M for GLM-5.3. GLM-5.3 is faster at generation (85 tok/s vs 49 tok/s) and wins on complex coding and long-

A merged community PR on the official zai-org/GLM-5.3-Flash Hugging Face repository adds structured evaluation results extracted from the model card's benchmark chart to the .eval_results/ directory. The added numbers are Terminal-Bench 2.1 at 84.3, DeepSWE at 63.4, and Humanity's Last Exam at 55.3, providing an offici The same PR documents three additional model-card claims that could not be registered because they lack a registered eval.yaml on the Hugging Face Hub: Agent's Last Exam at 26.3, AutomationBench v1.0.6 at 48.8, and GDPval-AA v2 at 1773. The contributor also flags that the cited arxiv:2602.15763 paper is the original GL

The official Hugging Face repository for zai-org/GLM-5.3-Flash presents the model as the first natively multimodal release in the GLM-5 series, built on a 320B-total / 18B-active MoE with a redesigned base, hybrid sparse-plus-linear attention, and Manifold-Constrained Hyper-Connections. It was trained on a 30T-token multimodal pre-training corpus and approaches Claude Opus 4.8 on coding and agentic benchmarks at one-tenth the price of GLM-5.2, per Z.ai. The README lists supported deployment frameworks: SGLang, vLLM, TokenSpeed, Transformers, KTransformers, and Unsloth, and documents thinking-budget control through the reasoning_effort parameter with low, high, and max levels (max is the default). It links the GLM-5.3-Flash blog and the GLM-5 Technical Report, and points users to Z.ai's API platform for hosted inference.

StealthModels identifies GLM-5.3-Flash as the revealed identity of the earlier Ox Alpha stealth listing, with Z.AI as developer, MIT license, 1M-token context, 131K maximum output, and 18B active parameters in the GLM-5 family. Its own StealthMark evaluation gives the model an Overall score of 53.7 out of 100, with General Intelligence at 48.5 and Visual Intelligence at 60.2, indicating stronger visual reasoning than general tasks. The page cross-references the Hugging Face repo for license and architecture and the Vercel AI Gateway listing for context and output figures. First availability is recorded as August 26, 2026. The model is positioned for cost-conscious coding and multimodal agent workloads served through APIs or self-hosted on machines with several hundred gigabytes of memory.

Reporting from Linas Beliūnas reconstructs the GLM-5.3-Flash launch: an anonymous listing called Ox Alpha appeared on OpenRouter and OpenCode on August 20, 2026, ran with no owner listed for six days, and processed roughly 23 trillion tokens before Bloomberg confirmed on August 26 that it was built by Zhipu AI / Z.ai. Stripe's CEO Patrick Collison publicly called it impressive, and Zhipu's Hong Kong shares rose more than 12 percent the next day. The piece confirms GLM-5.3-Flash specs of 320 billion total parameters with 18 billion active per token, MIT-licensed weights released on Hugging Face on August 26, and first-in-series native multimodality for text, image, and video. It describes the model as the cheapest at its score on Artificial Analysis's Intelligence Index, and flags it as the largest OpenRouter launch at roughly 2.3x the volume of the next model.

Independent benchmark tracker LLM Stats ranks GLM-5.3-Flash at number 24 on the LLM Stats Score composite, with a blended price near $0.17 per million tokens that places it 13th on cost efficiency. It records a GDPval-AA Elo of 1773 out of 3000 (rank 2), a CharXiv-R score of 0.89 (rank 8), and a Terminal-Bench 2.1 result sourced from Z.AI, signalling strong agentic and multimodal performance for its tier. The tracker also notes a quality-trend improvement of plus 1.82 sigma over baseline, while conversation-depth scores show a dip from about 16.5 at turn 1 to roughly 12.7 at turn 31 and beyond. The Cost Efficiency scatter compares GLM-5.3-Flash against peers like GPT OSS 120B, DeepSeek-V4-Flash-0731, and Claude Haiku 5.5. All scores are credited to the model's official scorecard, paper, or blog posts.

Z.AI's official developer documentation introduces GLM-5.3-Flash as the first natively multimodal model in the GLM-5 series, pairing text, image, video, and file inputs with a 1M-token context and 128K maximum output. It ships with a 320B-total / 18B-active MoE architecture that combines sparse and linear attention, reducing attention compute by about 3.01x and KV cache by 4.44x versus the GLM-5.3 base model. Recommended sampling is temperature 1, top_p 0.95, and reasoning_effort set to max. The page also documents a faster sibling variant, GLM-5.3-FlashX, which is reported at roughly 200 tokens/s for inference, and confirms GLM-5.3-Flash availability on the GLM Coding Plan with three times the usual quota. API access uses the model code glm-5.3-flash or glm-5.3-flashx through Z.AI's Chat Completion endpoint, with multimodal inputs passed as image_url content blocks. The documentation positions Flash for coding loops, GUI/browser control, and Office-style document workflows.