Holo-3.1-35B-A3B

30 of 40 layers use linear attention, which keeps a fixed-size recurrent state instead of a per-token cache. Those layers do not grow with context at all — treating them as ordinary attention, as a flat formula does, overstates this model's cache by roughly 4.0× at long context.

A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 18.39 GiB. The real file is 19.71 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

How much VRAM does Holo-3.1-35B-A3B need?

Q4_K_M is exactly 21,166,757,664 bytes (19.71 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.

How large is Holo-3.1-35B-A3B's KV cache?

0.63 GiB at 32K context with an f16 cache, computed per layer. Quantizing the cache to q8_0 roughly halves it, which is often the difference between a context length fitting and not.

Is Holo-3.1-35B-A3B a mixture-of-experts model?

Yes — 256 experts, 8 routed per token. Every expert must be resident, but only the routed ones are read per token, which is why its memory requirement and its speed behave very differently.

Which quantization of Holo-3.1-35B-A3B should I use?

Q4_K_M is the usual default. Pick the largest quantization that fits your card at the context you actually need — the table above gives exact sizes for every one published.