Blog / Model guides
GLM-5.3-Flash (Ox Alpha) hardware requirements and license
Z.ai has identified itself as the developer of Ox Alpha, the anonymous model listed on OpenRouter since 20 August. On 26 August the company told Bloomberg that Ox Alpha was a new model in its GLM series and that the weights would be released that evening. The weights were published the same day as GLM-5.3-Flash on Hugging Face under the MIT license. Z.ai's documentation states: "Before release, we tested GLM-5.3-Flash anonymously as ox-alpha on OpenCode and OpenRouter to gather user feedback." OpenRouter's stealth listing now names Z.ai as the developer.
Bloomberg reports that the anonymous model topped OpenRouter's usage leaderboard and more than doubled DeepSeek's traffic there, and that Stripe's Patrick Collison called the release "very impressive". The code name comes from the Chinese film Niu Lai, "Ox Comes". Z.ai went public in January and reports its first results as a listed company on Monday.
What GLM-5.3-Flash is
The figures below come from the repository's config.json and Z.ai's vLLM recipe.
| Property | Value |
|---|---|
| Parameters | 320B total, 18B active per token |
| Layout | 45 layers, 288 routed experts plus one shared, 8 experts per token |
| Attention | Hybrid sparse and linear attention, with multi-token prediction |
| Context | 1,048,576 tokens |
| Input | Text, images and video (a 448 px vision encoder is in the config) |
| Checkpoint | Native FP8, about 306 GiB; a BF16 variant, GLM-5.3-Flash-BF16, is roughly twice that |
| License | MIT |
| API price | $0.15 in, $0.50 out per million tokens; halved until 9 September |
What it takes to run
The FP8 weights alone, 306 GiB, exceed the memory of any single GPU. The vLLM recipe specifies Hopper or newer, tensor parallelism of at least 4 and vLLM 0.29. From those figures, 4x H200 (564 GB) holds the weights with about 250 GB left for KV cache, and an 8x H100 80 GB node (640 GB) also fits. The BF16 variant needs 8x H200. On Hopper the KV cache must stay in BF16, since FP8 KV cache is unsupported there, which raises memory use at long context. Z.ai lists SGLang, vLLM, TokenSpeed and KTransformers as serving stacks. A community NVFP4 quantisation exists; no measurements of it have been published.
Estimated cost to run privately
We have not deployed this model yet. The figures below are estimates from published rates. The H200 rate is the $7 to $14 per hour observed for 2x H200 on RunPod in August, scaled to four GPUs; the AWS figure is the p5en.48xlarge list price. Months are 730 hours, business hours 176.
| Deployment | Per hour | Around the clock, per month | Business hours only |
|---|---|---|---|
| GLM-5.3-Flash, 4x H200 on RunPod | $14 to $28 | $10,200 to $20,400 | $2,460 to $4,930 |
| GLM-5.3-Flash, 8x H200 on AWS (p5en.48xlarge, list) | $63.30 | $46,200 | $11,100 |
| DeepSeek V4 Flash, 2x H200 on RunPod (observed) | $7 to $14 | $5,100 to $10,200 | $1,230 to $2,460 |
Against the APIs: Z.ai lists GLM-5.3-Flash at $0.15 in and $0.50 out per million tokens, halved until 9 September; DeepSeek V4 Flash is $0.44 and $1.32 at peak, $0.22 and $0.66 off-peak. The break-even between a private node and the API depends on sustained token volume; the cost tool computes it from these rates.
The DeepSeek comparison
Bloomberg's headline says the model rivals DeepSeek. The numbers behind that are Z.ai's own, comparing GLM-5.3-Flash with DeepSeek-V4-Flash-Vision-Exp: 84.3 to 83.9 on Terminal-Bench 2.1, 63.4 to 59.3 on DeepSWE v1.1, 48.8 to 38.8 on AutomationBench, and 26.3 to 27.3 on Agents' Last Exam, where DeepSeek scores higher. No independent reproduction has been published. On hardware, DeepSeek V4 Flash runs on 2x H200 in our measurements; GLM-5.3-Flash needs at least 4x H200.
GLM-5.3, the flagship, is a separate release
GLM-5.3-Flash is a different model from GLM-5.3. The flagship keeps the roughly 753B GLM-5.2 base, and its Hugging Face page lists 28 August as the release date. Its hardware requirements are on the GLM-5.3 page.
Questions people ask
Is GLM-5.3-Flash the same model as Ox Alpha?
Yes. Z.ai's documentation says it tested GLM-5.3-Flash anonymously as ox-alpha on OpenCode and OpenRouter before release, and OpenRouter's stealth listing now says the model was developed and operated by Z.ai. The weights went to Hugging Face on 26 August 2026 under MIT.
What GPUs does GLM-5.3-Flash need?
The native FP8 checkpoint is about 306 GiB, so no single GPU can hold it. Z.ai's vLLM recipe calls for Hopper or newer, tensor parallelism of at least 4 and vLLM 0.29 or later. Derived from that, 4x H200 or 8x H100 80 GB fits the FP8 weights with room for KV cache; the BF16 variant needs an 8x H200 node.
Is GLM-5.3-Flash the same as GLM-5.3?
No. Flash is 320B parameters with 18B active. GLM-5.3 is the flagship on the roughly 753B GLM-5.2 base, and its weights are listed on Hugging Face for 28 August 2026.