What GPU do I need to run zai-org/GLM-4.7-FP8?
358.5B parameters, published in F8_E4M3. View on Hugging Face
Set up GLM-4.7-FP8GLM-4.7-FP8 is published by zai-org on Hugging Face, with 20,743 downloads and 125 likes to date. It's a Glm4MoeForCausalLM model built for text-generation, published natively in F8_E4M3.
VRAM required & cheapest live GPU fit
Required VRAM = weight size at each precision times a flat 1.2 overhead for activations and fragmentation. The KV-cache grows with context and is not in that factor; it is listed per model below where the architecture is published. Full formula and assumptions: methodology.
| Precision | Weight size | Required VRAM | Cheapest live fit | GPUs needed | Est. $/hr (full fit) |
|---|---|---|---|---|---|
| FP8 (native) | 333.8 GB | 400.6 GB | RTX PRO 6000 | 5 | $7.00/hr |
| INT4 (quantized) | 166.9 GB | 200.3 GB | RTX A6000 | 5 | $1.81/hr |
A GPU is only matched to a row if its hardware supports that precision, and the primary recommendation is always a single-GPU fit when one exists.
INT4 requires a quantized checkpoint actually published for this model, check its Hugging Face page before relying on this row.
Cheapest way to run GLM-4.7-FP8 at its published (F8_E4M3) precision: 5× RTX PRO 6000, at $1.40/hr per GPU ($7.00/hr total). Quantizing to FP8 or INT4 (rows above) can cost less, but requires a compatible quantized checkpoint to exist for this model.
GLM-4.7-FP8: KV cache by context length
The KV cache is the memory the attention layers hold for every token of context, on top of the weights. It is not part of the flat 1.2x overhead in the table above, grows with context length and with every concurrent request, and is why a long-context deployment needs more VRAM than the table shows.
Grouped-query attention: every layer caches keys and values for a few shared KV heads. Cached per token: 376,832 bytes at 16-bit.
| Context | KV cache, one sequence |
|---|---|
| 4K | 1.4 GB |
| 32K | 11.5 GB |
| 128K | 46.0 GB |
| 198K (model maximum) | 71.2 GB |
Computed from the layer, head and window counts in the model's own configuration (zai-org/GLM-4.7-FP8), at 16 bits per cached value. Each concurrent sequence needs its own cache, on top of the weights.
GLM-4.7-FP8: common questions
Can GLM-4.7-FP8 run on a single GPU?
No. At FP8 (native) it needs 400.6 GB of VRAM, more than a single 32 GB desktop card holds. The cheapest capable card in the live feed is a 96.0 GB RTX PRO 6000, and it takes 5 of them.
Is GLM-4.7-FP8 already quantized?
Yes. It is published in FP8, one byte per parameter, so the 400.6 GB figure is already a quantized footprint rather than a full-precision one. Only INT4 goes below it, at 200.3 GB. A GPU without FP8 tensor cores cannot run it as published, which is why cards here are matched on precision support and not on VRAM alone.
How many GPUs do I need to run GLM-4.7-FP8?
5 at FP8 (native). It needs 400.6 GB of VRAM and the cheapest capable live offer is a 96.0 GB RTX PRO 6000, so 5 of them come to $7.00/hr in total.
Does quantizing GLM-4.7-FP8 lower the GPU bill?
Yes. At FP8 (native) the cheapest live fit is 5 RTX PRO 6000 cards at $7.00/hr. At INT4 (quantized) it drops to 5 RTX A6000 cards at $1.81/hr, provided a quantized checkpoint exists for it.
Weight-to-VRAM math, the fit rules, and how live prices are normalized: full methodology.
More GLM-4.5 models
- GLM-4.5 (358.3B, BF16)
- GLM-4.7 (358.3B, BF16)
- GLM-4.6 (356.8B, BF16)
- GLM-4.5-Air-FP8 (110.5B, F8_E4M3)
- GLM-4.5-Air (110.5B, BF16)
Alternatives at this size
Other models for text-generation within about a third of GLM-4.7-FP8's 358.5B parameters, from other model lines.
- MiMo-V2.5 (310.8B, F8_E4M3)
- Ornith-1.5-397B (403.4B, BF16)
- Llama-3.1-405B (405.9B, BF16)
- Hy3-preview (298.8B, BF16)
- Qwen3-Coder-480B-A35B-Instruct (480.2B, BF16)
More on GLM-4.7-FP8
Fits on a 288 GB GPU at INT4: every model that fits in 288 GB.
Related reading: RTX PRO 6000 pricing and specs, How much VRAM you need for LLMs, Serving LLMs with vLLM, vLLM vs TensorRT-LLM vs SGLang, and Best GPU for LLM inference.
