GLM 5.3 Flash: Z.ai’s cheaper, multimodal sibling matches larger models but slows on max reasoning
Z.ai's GLM 5.3 Flash is a 320-billion-parameter multimodal model that costs about one-ninth the output price of the flagship GLM 5.3. Independent tests show strong function-calling and agentic behavior, but max‑reasoning mode slows throughput to 50–60 tokens per second, and agentic fallbacks to the larger GLM are expensive, pushing users toward other frontier models for critical tasks.
Listen to this dispatch
Narrated by an AI-generated voice.
GLM 5.3 Flash, reviewed here, is a smaller, cheaper sibling to Z.ai's GLM 5.3 flagship. The two models share a name, but Flash is a separate pre-trained model rather than a distilled version of the larger one. The full GLM 5.3 is a 744-billion-parameter mixture-of-experts model with 40 billion active parameters and text-only input. Flash has 320 billion total parameters with 18 billion active, and adds image and video support.
Flash was pre-trained on more data — over 30 trillion multimodal tokens, versus the 28.5 trillion text tokens behind the flagship. Its architecture differs as well: sparse attention and linear attention are interleaved across 45 layers (the full model has 80), with eight of 288 experts active per token and a multi-token-predictor draft layer for speculative decoding.
The reviewer's own tests found strong function-calling and agentic behavior, including graceful handling of deliberate tool failures and resistance to tool-baiting prompts. Low reasoning mode often uses under 50 thinking tokens while still completing tasks. SVG generation is visually strong on the max setting but consumes large numbers of tokens. The reviewer observed signs of overfitting: a red dragon with a very similar prompt to the pelican example produced notably weaker results. RAG tasks were efficient; output design work was rudimentary at low token budgets.
The pricing difference is pronounced. At full price, Flash costs $0.50 per million output tokens and $0.15 per million input tokens, against $4.40 per million output tokens for the full-size model — about a ninth of the cost. Both models expose a reasoning effort parameter (low, high, max), and neither allows thinking to be disabled.
Z.ai's published benchmarks show Flash outperforming GLM 5.2 despite being roughly half its size, and performing comparably to models like Opus 4.8, GPT-5.6 Tera, and Gemini Flash. The full 5.3 still wins on some tasks — the humanities last exam, for instance, where Flash scores 55.3 against the larger model's 62.5. On an aggregate intelligence score cited in the video, Flash lands at 57 versus 60 for the full model, while competing against much larger open-weight systems like Qwen 3.8 and Kimi K3, both well over 2 trillion parameters. The presenter notes Flash's agentic index trails Anthropic models but sits ahead of Kimi K3.
Speed is the practical caveat. In max mode the model thinks for a long time, and the presenter measured API throughput around 50-60 tokens per second. That creates a workflow constraint: a failing task might succeed after retries, but each attempt is slow, and the fallback to the full-size GLM 5.3 costs nine times more per output token. The presenter suggests the realistic fallback for agentic failures may be frontier models from OpenAI or Anthropic rather than the larger GLM.
The reviewer plans a follow-up on local execution speeds. Whether Flash's hybrid attention architecture is a test bed for Z.ai's next flagship, or a permanent tier, remains an open question.