GLM-5.3

GLM-5.3 is the GLM-5.2 base model: a 753B-parameter mixture-of-experts (glm_moe_dsa in Transformers) activating about 40B parameters per token, the figure Z.ai published for that base. Z.ai's config.json on Hugging Face shows 78 transformer layers plus one multi-token-prediction layer, the first three layers dense and the rest MoE, 256 routed experts plus one shared expert with 8 routed experts per token, hidden size 6,144, 64 attention heads with low-rank (MLA-style) query and key-value projections, a 1,048,576-position window and a 154,880-token vocabulary. Attention is sparse: a lightweight indexer picks the top 2,048 tokens for each query, and 57 of the 78 layers reuse an indexer computed on an earlier layer (the IndexShare design introduced with GLM-5.2). The main repository ships FP8 (E4M3, 128x128 blocks) weights; zai-org/GLM-5.3-BF16 holds the BF16 checkpoint. Post-training ran on slime, Z.ai's open-source RL framework (Megatron for training, SGLang for rollouts), with SAO and context compaction for long-horizon RL and machine-synthesized task environments whose verifiers are built without the reference solution. Training tokens, compute and knowledge cutoff are undisclosed; the GLM-5 technical report is arXiv 2602.15763.

Scores lean on Z.ai's launch figures, which no independent evaluator had published at verification time, so they are set conservatively.

• Coding (9.4): Terminal-Bench 2.1 88.2 (GPT-5.6 Sol 88.8, Kimi K3 88.3), DeepSWE v1.1 66.9 (Sol 72.7), FrontierSWE 78.1 (Claude Fable 5 88.2, Claude Opus 4.8 66.5), NL2Repo 58.0 (Opus 4.8 69.7). Level with the best open-weights peers and a few points behind Sol and Fable 5 on the hardest rows, on Z.ai's own table.
• Reasoning (9.3): HLE with tools 62.5 against Sol's 64.5 and Fable 5's 63.9. No GPQA, AIME or MMLU-Pro figure was published for GLM-5.3.
• Math (9.5): No math benchmark published for GLM-5.3. Held at GLM-5.2's level because the base model is the same.
• Agentic / tool use (9.4): Top of Z.ai's comparison set on AutomationBench (48.2) and GDPval-AA v2 (1,769 Elo, run by Artificial Analysis), level with Sol on Agents' Last Exam (28.5 vs 28.6), behind Kimi K3 (76.5) and Opus 4.8 (76.2) on Toolathlon Verified (73.0).
• Long context (9.0): 1M-token window on sparse attention built for that length. Z.ai ran several evaluations at 1M context, but published no dedicated long-context benchmark.
• Multilingual (8.5): English and Chinese are the tagged languages; no multilingual benchmark published.
• Vision / document OCR (0 / 0): Text-only. Z.ai's chat template tells the model it cannot process images, video or audio.
• Creative writing (7.5): No data; Z.ai's tuning targets engineering work. Held at GLM-5.2's level.
• Instruction following (8.8): No IFEval figure. Held at GLM-5.2's level.
• Function calling (9.2): Up to 128 functions per request, streamed tool-call arguments (tool_stream), and the Toolathlon gain of 13 points over GLM-5.2.
• Safety calibration (7.0): Z.ai reports strong exploit-development capability in an open-weights model, delayed the weights for safety evaluation and hardening, and published no results from that work. Self-hosters own every safeguard.
• Real-time data (0): No built-in retrieval is listed on the GLM-5.3 model page. Z.ai sells a Web Search tool at $0.01 per use, but does not document it for this model.

All scores below are vendor-reported by Z.ai on its model card and launch blog (2026-08-14). Competitor figures are as published in Z.ai's table. Z.ai's table labels one column "Fable 5 (w/ fallback)"; its blog text refers to the same cyber scores as Mythos 5.

Most agentic and coding rows ran in Claude Code 2.1.207 at max effort, with harness details in Z.ai's footnotes. Z.ai replaced some anti-cheat checks in SWE-Marathon and PostTrainBench with LLM-based inspection, and ExploitGym time budgets were rescaled by each model's throughput. Only Terminal-Bench 2.1 maps to a catalog compare field, on the same version as GLM-5.2's stored 81.0. The HLE result is with tools, while GLM-5.2's stored HLE (40.5) is without tools, so that compare field stays empty rather than mixing modes. Z.ai published no GPQA Diamond, AIME, SWE-bench Verified, LiveCodeBench, MMLU-Pro or LMArena figure for GLM-5.3, and no independent leaderboard result for it could be confirmed at verification time.

No independent throughput or time-to-first-token measurement could be confirmed, so both fields are null. The only figure on Z.ai's pages is in its ExploitGym footnote, which rescales time budgets using 115 output tokens per second for GLM-5.3 (40 for Kimi K3, 47 for Qwen3.8-Max), numbers Z.ai attributes to Artificial Analysis. In practice, reasoning runs on every request and defaults to max, so even simple prompts spend time thinking; low is the fastest setting. Z.ai reports fewer output tokens per task than GLM-5.2 (about 75K against 96K at max effort on its coding benchmark), which shortens long agent runs. Measure on the host you plan to use: Z.ai, Mistral, Alibaba Cloud and six inference providers listed on Hugging Face all serve it.

Z.ai's natively multimodal GLM-5.3-Flash costs $0.15 / $0.50, about a tenth of GLM-5.3.

Faster serving (Prime mode). Alibaba Cloud Model Studio also serves this model as glm-5.3-prime (listed 2026-09-23): the same weights and capabilities with a claimed 1.5–2× output speed, at $2.80 input / $8.80 output / $0.56 cached per 1M tokens (Singapore), about twice Z.ai's own list price. Z.ai does not offer Prime itself. Source: Alibaba Cloud Model Studio, Prime mode (help.aliyun.com/en/model-studio/prime-mode); OpenRouter z-ai/glm-5.3-prime.

• Z.ai API: Z.ai's model page lists three protocols: OpenAI Chat Completions, OpenAI Responses (https://api.z.ai/api/v1) and Anthropic Messages (https://api.z.ai/api/anthropic). The API reference serves chat at https://api.z.ai/api/paas/v4/chat/completions. Accounts that have ever held a GLM Coding Plan, even an expired one, can currently call the API only through the Chat Completions protocol. Zhipu's China platform (bigmodel.cn) has served glm-5.3 since 2026-08-19.
• Open weights: FP8 weights at zai-org/GLM-5.3 and BF16 at zai-org/GLM-5.3-BF16, with recipes for SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth, plus Ascend NPU support through vLLM-Ascend, xLLM and SGLang. At one byte per parameter the FP8 weights alone are about 750 GB before KV cache, so this is a multi-GPU server or cluster model.
• License: The GLM-5.3 License grants MIT-style rights to use, copy, modify, merge, publish, distribute, sublicense and sell the software, and to run, deploy and fine-tune it and build derivatives, on two conditions: keep the copyright and permission notice, and comply with applicable law. The added clause defines "Model as a Service" as giving a third party access to language-model inference or fine-tuning (for example through an API) with meaningful control over the inputs, parameters or training data. End-user products that embed the model only within specific features or harnesses, and services that merely relay requests to models hosted by others, do not count. If the licensee or any affiliate operates a Model as a Service business and the licensee and its affiliates together have more than US$10 billion in total revenue over any consecutive 12 months, the licensee must pass Z.ai's security review before using GLM-5.3 or its derivatives for any commercial purpose, with the review's scope and method set by Z.ai. The trigger is group revenue, not model-as-a-service revenue, and once triggered the review covers all commercial use. For nearly every buyer the terms behave like MIT; hyperscale clouds and the largest platforms need Z.ai's sign-off. Questions go to glmlicense@z.ai.
• Managed hosts: Hugging Face lists Together AI, Fireworks, DeepInfra, Novita, Baseten and Featherless AI as live inference providers. Mistral lists GLM 5.3 as a third-party model on its API pricing page, and its public docs mark it public preview (released on Mistral 2026-09-15) with a 1M context, 128K output, function calling, structured outputs and batching, under the API names zai-glm-5-3, zai-glm-5 and zai-glm-latest. Alibaba Cloud Model Studio serves glm-5.3 with Alibaba as the inference provider, 1,048,576-token input and 131,072-token output.
• Data residency: Z.ai's API data processing addendum says customer data is generally processed in Singapore.

• Framework: Z.ai publishes no named safety framework, capability tiers or refusal metrics for GLM-5.3.
• Dual-use cyber capability: Z.ai's own figures put GLM-5.3 at the top of CyberGym and more than double GLM-5.2 on exploitation benchmarks, while still well behind the closed frontier on ExploitBench and ExploitGym. Z.ai delayed the weights for safety evaluation and hardening but published no results. It runs a Security Disclosure Ledger; at launch it counted 2,436 findings, 53 publicly disclosed and 2,383 under embargo.
• License gate: The GLM-5.3 License requires a Z.ai security review before the largest model-as-a-service groups use it commercially. There are no use-based restrictions for anyone else.
• API data handling: Z.ai's API is operated by Jingsheng Hengxing Technology Pte. Ltd. in Singapore. Its data processing addendum says Z.ai acts as a data processor on the customer's instructions, does not store the content customers input or generate (it is processed in real time), and generally processes customer data in Singapore.
• Compliance: No SOC 2, ISO 27001, HIPAA or GDPR certification is stated.
• Self-hosting: Moderation, refusal behavior and logging are entirely the deployer's responsibility.

Z.ai's official SDKs are zai-sdk for Python and Java, and the OpenAI Chat Completions, OpenAI Responses and Anthropic Messages protocols let standard OpenAI and Anthropic clients connect. Z.ai positions GLM-5.3 for coding agents: it is the flagship model on the GLM Coding Plan and in ZCode, Z.ai's own agent app, and Z.ai benchmarks it inside Claude Code; OpenCode is also supported. For self-hosting, SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth have GLM-5.3 recipes, and Ascend NPUs are supported. Beyond Z.ai, Together AI, Fireworks, DeepInfra, Novita, Baseten, Featherless AI, Mistral and Alibaba Cloud Model Studio serve it. The main Hugging Face repository logged about 1 million downloads in its first month. Popularity is growing, carried by the coding-agent audience GLM-5.2 built.