GLM-5.3-Flash: Cheap, Open and Multimodal—Decide From the Artifact
AccessAllGPT desk review of the GLM-5.3-Flash Hugging Face model card, registry record, config.json, weight index and LICENSE at immutable revision 03eb5366286afd40d2221b1d9c63a6dd1ba4832e; the Z.ai release blog; the Z.ai GLM-5.3-Flash developer guide and pricing pages; and the GLM-5 technical-report abstract (arXiv:2602.15763v2), retrieved 2026-09-03. We wrote and ran scripts/audit-glm53-flash-artifact.mjs on Node v24.10.0, Darwin arm64. The script queried the Hugging Face API, paginated the pinned repository tree, reconciled 62 safetensor shards between tree and index, and reported file counts, byte totals, indexed tensor counts, parameter counts by dtype, architecture fields, layer-type counts, quantization exclusions and the license first line without downloading any shard. Vendor capability claims, registry facts, third-party-platform metrics and AccessAllGPT guidance are labeled separately below. This is not an original model run, benchmark reproduction, serving test, API trial, quantization evaluation, license review or hardware assessment.
GLM-5.3-Flash model card at pinned revision 03eb5366
Z.ai on Hugging Face · RetrievedReviewed: Introduction; Serve GLM-5.3-Flash Locally framework list and linked cookbooks; reasoning_effort and clear_thinking notes; evaluation footnotes; citation block; Hugging Face-hosted evaluation results; model tree, downloads and likes
Supports: The pinned card introduces GLM-5.3-Flash as the first natively multimodal GLM-5-series model, states 320B total and 18B active parameters, a newly trained base model, a hybrid sparse-plus-linear attention design, Manifold-Constrained Hyper-Connections and a 30T-token multimodal corpus, links SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth deployment paths, documents reasoning_effort low/high/max with a max default and a clear_thinking chat-template default of false, and publishes per-benchmark run notes including harnesses, budgets, a GPT-5.6-luna judge and an empty Agent’s Last Exam entry. The page also lists Hugging Face-hosted Terminal-Bench 2.1 and ExtractBench rows and shows 517,902 downloads and 2,005 likes in the September 3 snapshot. These are vendor-authored claims and mutable platform metrics, not AccessAllGPT measurements.
Hugging Face registry record for zai-org/GLM-5.3-Flash
Hugging Face · RetrievedReviewed: Repository identity, pinned main revision SHA, createdAt and lastModified timestamps, safetensors parameter totals by dtype, downloads, likes, gating and tags
Supports: The registry records that the repository was created 2026-08-25T06:43:14Z, was last modified 2026-08-31T11:33:06Z, is not gated, and at pinned revision 03eb5366286afd40d2221b1d9c63a6dd1ba4832e declares 321,323,031,390 indexed parameters split across 314,396,639,232 F8_E4M3, 6,926,096,640 BF16 and 295,518 F32 values. These are registry facts about the artifact, not evidence of model quality or capability.
config.json at pinned revision 03eb5366
Z.ai on Hugging Face · RetrievedReviewed: architectures and model_type; text_config layer_types, linear_attn_config, indexer fields, mHC fields, MoE routing, MLA-style dimensions, vocabulary and positions; vision_config; fp8 quantization_config and modules_to_not_convert
Supports: The pinned configuration identifies Glm5NextForConditionalGeneration with glm5_next_text and glm5_next_vision components: 45 text layers typed as 34 linear_attention plus 11 deepseek_sparse_attention at layer indexes 3,7,11,15,19,23,27,31,35,39,43; mhc true with hc_mult 4; 288 routed experts plus 1 shared expert with 8 selected per token; 1,048,576 maximum position embeddings; a 24-layer, 448-pixel, patch-14 vision tower with spatial merge 2; and fp8 quantization excluding 1,509 named modules. These fields describe the shipped artifact only; they do not establish throughput, memory needs or quality.
model.safetensors.index.json and repository tree at pinned revision 03eb5366
Z.ai on Hugging Face · RetrievedReviewed: Weight-map tensor names and shard assignments; metadata total_size; recursive repository tree file list and shard sizes
Supports: AccessAllGPT paginated the pinned tree and reconciled it with the weight index: 72 files, 62 safetensor shards named model-00001-of-00062 through model-00062-of-00062, 76,108 indexed tensors, 328,326,771,576 indexed parameter bytes and 328,337,455,672 bytes of shard payloads. The shard count matches between tree and index. These byte counts size the download; they are not serving-memory requirements.
LICENSE at pinned revision 03eb5366
Z.ai on Hugging Face · RetrievedReviewed: Full 21-line license text
Supports: The pinned license file begins “MIT License” and carries “Copyright (c) 2026 Z.AI Co., Ltd” with the standard MIT permission, notice and warranty text. This establishes the license of the pinned weight artifact only; it does not cover related APIs, services, or other models, and it is not legal advice.
GLM-5.3-Flash release blog
Z.ai · RetrievedReviewed: Announcement and quota statements; Artificial Analysis Pareto claim; six-benchmark comparison table summary; Z.ai Code Bench comparison; efficiency and IndexPool design; base-model evaluation statement; ox-alpha anonymous testing; Chinese-AI-chip serving stack and three-times improvement claim; conclusion
Supports: Z.ai states GLM-5.3-Flash was tested anonymously as ox-alpha on OpenCode and OpenRouter and became the most popular model of the week; scores 57 on the Artificial Analysis Intelligence Index v4.1.1 at $0.045 per task at discounted pricing; outperforms GLM-5.2 across six coding and agentic benchmarks including 63.4 vs 46.2 on DeepSWE v1.1 and 48.8 vs 26.2 on AutomationBench; nearly matches Claude Opus 4.8 at max effort on Z.ai Code Bench v1.0 (29.0 vs 29.5); reduces attention compute 3.0× and KV cache 4.4× versus GLM-5.3 while retaining a larger cache than Kimi-K3 and DeepSeek-V4-Flash; and was served on Chinese AI chips with a claimed 3× end-to-end serving improvement. All of these are vendor claims; AccessAllGPT reproduced none of them.
GLM-5.3-Flash model guide
Z.ai Developer Documentation · RetrievedReviewed: Model overview and hybrid-architecture statements; input and output modality, context and output limits; model code and parameter recommendations; thinking and reasoning_effort constraints; image input construction; Coding Plan quota and off-peak points; capabilities list; best-practice scenarios
Supports: Z.ai documents model code glm-5.3-flash with Video/Image/Text/File input and Text output, a 1M context window and 128K maximum output tokens; recommends temperature 1, top_p 0.95 and reasoning_effort max; states thinking.type supports only enabled and recommends thinking.clear_thinking false with stream and tool_stream enabled; shows image_url content blocks for image input; and states GLM-5.3-Flash is fully available on the GLM Coding Plan with 3× GLM-5.3 quota, a points-based system and 50% point consumption off-peak including weekends. These are vendor documentation statements of the integration contract, not verified behavior.
Z.ai pricing
Z.ai Developer Documentation · RetrievedReviewed: Latest-models price table, per-million-token units, discount framing and promotion end date, cached-input storage note
Supports: The published table lists GLM-5.3-Flash at $0.075 input, $0.015 cached input and $0.25 output per million tokens, shown against struck-through list prices of $0.15, $0.03 and $0.50, with cached-input storage limited-time free; the 50% promotion ends 24:00 on September 9, 2026 UTC+8. GLM-5.3 and GLM-5.2 list $1.4 input, $0.26 cached input and $4.4 output. These are posted list prices on one platform; they are not total cost of ownership and are subject to change.
GLM-5: from Vibe Coding to Agentic Engineering (arXiv:2602.15763v2)
GLM-5 Team on arXiv · RetrievedReviewed: Version history; abstract; subject classification; license badge
Supports: The linked technical report (v1 February 17, 2026; v2 February 24, 2026; CC BY 4.0) describes the GLM-5 generation’s adoption of DSA to reduce training and inference cost while maintaining long-context fidelity, an asynchronous reinforcement-learning infrastructure, and asynchronous agent-RL algorithms. The report is the citation the GLM-5.3-Flash card supplies, but its abstract describes GLM-5; it does not isolate GLM-5.3-Flash’s results, and AccessAllGPT reviewed the abstract and metadata only.