GLM-5.3-Flashreview, pricing and limits

GLM-5.3-Flash is Z.ai's open-weight sibling to GLM-5.3, offering a 1,048,576-token context window and MIT-licensed weights published on Hugging Face for immediate self-hosting. It replaces heavier closed multimodal models for teams that need self-hosted vision-and-video reasoning at 320B total parameters with only 18B active per token.

GLM-5.3-Flash is Z.ai's first natively multimodal GLM-5 model, released August 26, 2026 with 320 billion total parameters and 18 billion active per token. It scores 84.3 on Terminal-Bench 2.1, within a point of Claude Opus 4.8's 85.0, and ships as MIT-licensed open weights.

Provider: Z.ai · Family: GLM-5.3

More about Z.ai on HokAI

Context window: 1,048,576 tokens · Max output: 48,000

Input modalities: text, image, video, pdf, tool-calls · Output: text, tool-calls

About GLM-5.3-Flash

GLM-5.3-Flash is a multimodal large language model built by Z.ai (formerly Zhipu AI), a Beijing-based AI lab that completed a Hong Kong IPO in January 2026. Z.ai confirmed the model's identity on August 26, 2026, after it had circulated for six days on OpenRouter, OpenCode and Cline under the codename "Ox Alpha." It is a Mixture-of-Experts transformer with 320 billion total parameters and 18 billion active per token, built from 45 layers that mix KDA linear attention with NoPE sparse MLA attention and route each token through 8 of 288 experts. It is the first natively multimodal release in the GLM-5 line, sitting alongside the larger, text-only GLM-5.3 (released August 14, 2026) as the faster, cheaper sibling. On agentic and coding evaluations, GLM-5.3-Flash scores 84.3 on Terminal-Bench 2.1, within a point of Claude Opus 4.8's 85.0 and close to GPT-5.6 Terra's 87.4. On Z.ai's own Code Bench v1.0 it scores 29.0 against Opus 4.8's 29.5. On DeepSWE v1.1 it jumps to 63.4, up from 46.2 for the previous GLM-5.2, and AutomationBench nearly doubles to 48.8 from 26.2. Artificial Analysis puts its composite Intelligence Index at 57, ranking third of 109 tracked models at launch. Standard academic benchmarks such as GPQA Diamond, MMLU-Pro and SWE-bench Verified had not been separately published for the Flash variant as of this writing; Z.ai's disclosures focus on agentic and vision tasks instead. The model ships with a 1,048,576-token context window, built on an IndexPool mechanism that compresses indexer key vectors, cutting attention compute roughly 3x and KV-cache size 4.4x compared with the full GLM-5.3. Maximum completion length is 48,000 tokens on the hosted OpenRouter endpoint. Z.ai has not published a separate long-context recall score specific to this model. GLM-5.3-Flash accepts text, images, video and files as input and returns text, including structured JSON output. Z.ai's documentation describes tool-calling support for external tool integration, and the model runs in an always-on thinking mode: reasoning cannot be disabled per the vendor's own parameter guidance. A vision encoder handles image and video frames, letting the model read screenshots, interfaces, charts and documents inside an agent loop, then inspect its own rendered output and iterate. Standard API metering runs at a fraction of the full-size GLM-5.3 rate, with a limited-time launch discount cutting that further through early September 2026 (exact figures in the pricing tab). The model is bundled into every tier of the GLM Coding Plan, from the entry Lite plan to the top Max plan, at three times the usable quota of GLM-5.3. The primary channel is Z.ai's own OpenAI-compatible developer API at docs.z.ai. Open weights are published as zai-org/GLM-5.3-Flash on Hugging Face under the MIT license, with a native FP8 safetensors checkpoint (about 306 GiB) and a BF16 variant at roughly double that size; community GGUF and NVFP4 quantizations followed within days. Self-hosting the FP8 checkpoint needs an 8-GPU node of NVIDIA Hopper-class or newer hardware, or AMD Instinct gfx950 via ROCm, with vLLM, SGLang, KTransformers and TokenSpeed all shipping day-one recipes. OpenRouter, Cloudflare Workers AI and Vercel's AI Gateway all added the model the same week as the official announcement. As of this writing, none of the big three hyperscaler catalogs (AWS Bedrock, Google Vertex AI, Azure AI Foundry) list GLM-5.3-Flash directly, though they carry earlier GLM-5 releases. Z.ai's public governance disclosures describe supervised fine-tuning and RLHF alignment plus a hosted-API content filter, with no formal responsible-scaling-policy analog to Anthropic's RSP. For the wider GLM-5.3 release, Z.ai delayed shipping open weights to complete a safety evaluation after finding the model's code-vulnerability discovery ability had increased sharply; no equivalent delay has been reported for the Flash variant specifically. The MIT license lets outside researchers audit and red-team the weights directly, which is the company's main stated safety lever rather than a published red-team partner list. Teams building agentic coding tools, browser or computer-use agents, or document and screenshot-reading pipelines get near-frontier agentic scores at open-weight prices, and can self-host for data residency reasons closed frontier labs cannot match. Teams needing verified academic reasoning scores (GPQA, MMLU-Pro, formal SWE-bench Verified) should look elsewhere until Z.ai publishes those numbers for this specific checkpoint. Raw generation speed, 48.7 to 50.2 tokens per second across measured providers, trails the roughly 67 t/s median for comparable open-weight models, so latency-sensitive, high-throughput chat products may prefer GLM-5.2 or a smaller model instead. Z.ai trained the underlying architecture on a 30-trillion-token multimodal corpus; the company has not disclosed a specific training data cutoff date for the Flash checkpoint. Z.ai has not published SOC 2, HIPAA or GDPR certifications for its hosted API, and data residency for the hosted service defaults to China-based infrastructure plus whatever regions third-party API resellers offer. The MIT license carries no data-retention terms of its own; retention and training-on-inputs policy for the hosted API is set by Z.ai's terms of service rather than the model card. GLM-5.3-Flash follows GLM-5.2 (open weights, released in mid-June 2026) and the full-size GLM-5.3 (released August 14, 2026), arriving twelve days later as the multimodal, cheaper sibling. It had already been serving production traffic anonymously as "Ox Alpha" on OpenRouter, OpenCode, Cline and Nous Research's portal since August 20, 2026, before Z.ai confirmed the identity match on August 26 and published weights the same day.

Pricing

List price is $0.15 per 1M input tokens, $0.50 per 1M output, $0.03 per 1M cached input via Z.ai's official API. A launch promotion halves all three (to $0.075/$0.25/$0.015) through September 9, 2026, 16:00 UTC. Open weights can also be self-hosted for the cost of compute only.

Key Features

• Native Multimodal Input: Accepts text, images, video and files in a single request, the first GLM-5 release built multimodal from the ground up rather than bolted on.
• Million-Token-Class Context Window: A vision-capable long-context window built on Z.ai's IndexPool mechanism, which trims attention compute and KV-cache size well below the full GLM-5.3's footprint.
• MIT-Licensed Open Weights: Full-precision and quantized checkpoints published on Hugging Face under MIT, permitting unrestricted commercial use, fine-tuning and self-hosting.
• Sharp Agentic-Coding Gains Over GLM-5.2: DeepSWE v1.1 climbed to 63.4 from GLM-5.2's 46.2, and AutomationBench nearly doubled to 48.8, the biggest jump of any metric Z.ai disclosed at launch.
• Steep Discount Off the Flagship Rate: Priced far under what Z.ai charges for the full-size GLM-5.3, and a launch-week promo trims the bill even more into early autumn 2026.

Pros

• Artificial Analysis ranks its Intelligence Index third of 109 tracked models at launch, an unusually strong showing for an open-weight release.
• A vision-capable, million-token-class context window with a meaningfully lighter attention and KV-cache footprint than the full GLM-5.3.
• MIT license allows self-hosting, fine-tuning, and independent audit of the weights, unlike closed frontier models.

Cons

• No published GPQA Diamond, MMLU-Pro, or SWE-bench Verified score for this checkpoint as of launch.
• Raw output speed sits below the median for its open-weight peer group, so it is not the fastest choice for high-throughput chat.
• Self-hosting the FP8 checkpoint needs an 8-GPU Hopper-class node (about 306 GiB VRAM), out of reach for smaller teams.

Benchmarks

• mvbench: 77.8
• nl2repo: 56.3
• babyvision: 53.4
• toolathlon: 78.4
• chartography: 78
• officeqa pro: 62.4
• deep swe v1 1: 63.4
• hle with tools: 55.3
• agents last exam: 26.3
• automation bench: 48.8
• gdpval aa v2 elo: 1773
• charxiv reasoning: 89.4
• zai code bench v1: 29
• terminal bench 2 1: 84.3
• artificial analysis intelligence index: 57
• artificial analysis price blended per m: 0.1
• artificial analysis speed tokens per sec: 48.7

Visit GLM-5.3-Flash Official Page