Separate the Two Similarly Named Models First
Coverage over the past fortnight repeatedly conflates GLM-5.3 and GLM-5.3-Flash. They are two models: Start with GLM-5.3: released August 14, text in and text out with no vision or audio path, reusing the GLM-5.2 base with all gains from post-training, priced at $1.40 input and $4.40 output per million tokens, offering low/high/max effort levels with max as default and no option to disable thinking. Then GLM-5.3-Flash: released August 26, natively multimodal on a newly trained base — hybrid sparse and linear attention, Manifold-Constrained Hyper-Connections, a 30T-token multimodal pretraining corpus — which Z.ai says uses roughly 3.0× less attention compute and a 4.4× smaller KV cache than the text GLM-5.3. The prices differ by nearly 10× and the capability axes differ too, so do not substitute one for the other.
The Cheap and Open Parts Are Real
An MIT license at this parameter scale is uncommon: download, fine-tune and deploy commercially with no additional restrictions. On price, list is $0.15 per million input tokens, $0.03 cached input and $0.50 output, with a 50% promotion ($0.075/$0.015/$0.25) running through September 9; OpenRouter shows 22 providers serving it. Z.ai's self-reported numbers are DeepSWE 63.4 (against GLM-5.2's 46.2) and AutomationBench 48.8 (against 26.2), with a claim of landing within half a point of Claude Opus 4.8 on its internal coding benchmark at roughly a tenth of the price. All vendor-reported, none independently reproduced. The release mechanics are worth noting: it ran anonymously as "Ox Alpha" on OpenCode and OpenRouter for a week before the reveal, reportedly served on domestically produced Chinese AI chips. Running anonymously first is effectively a round of blind community evaluation before anyone knows whose model it is.
Three Limits to Know Before Selecting It
First, the 1M context is an input ceiling, not a quality depth. Architecturally max_position_embeddings is 1,048,576 and OpenRouter lists 1,310,720 tokens served, but the model card's published long-context evaluations only run to 300,000 tokens. There is no public evidence for behavior past 300K, so do not design a pipeline around a million. Second, self-hosting is not light. FP8 weights run about 306 GiB, requiring multi-GPU serving through SGLang, vLLM or TokenSpeed. Unsloth's dynamic GGUFs run on a 128 GB machine at 1–2 bit, but the glm5_next architecture is not yet in mainline llama.cpp, so Ollama and LM Studio cannot serve it locally — anyone hoping to run this on a laptop should set that aside. Third, the promotional price expires. It reverts to list on September 9, so cost models should use $0.15/$0.50 rather than the promo rate. For teams selecting a low-cost multimodal model, the reason to shortlist this one is the combination of MIT weights, native multimodality and that price point. If your use case depends on real quality at very long context, or on single-machine local deployment, it is not there yet.
via: MarkTechPost: GLM-5.3-Flash, Z.AI developer docs, OpenRouter model page, VentureBeat on GLM-5.3