Computed

gpt-oss 20B on an Apple M4 Max (128GB)

Yes — the unquantized weights fit with room for 128k tokens of context.

Change the numbers →gpt-oss 20B on every card Apple M4 Max (128GB) across every model

Verdict

Fits at FP16

96.0 GiB usable of 128 GB

Weights (Q4_K_M)

10.8 GiB

12.8 GiB at FP16 · 21.5B params · 6 of 6 quant rows are measured files

KV cache

48.0 KiB per token at FP16

24 layers × 8 KV heads × 64 dimensions

Speed ceiling

184 tok/s

546 GB/s ÷ bytes read per token

Every quant, against this card

Weight bytes are the size of the real published file wherever one exists — 6 of these6 rows are measured from unsloth/gpt-oss-20b-GGUF, the rest computed from the parameter count. Max context is what the KV cache can grow to in whatever memory the weights leave behind, capped at the 128k tokens this model was trained to address.

Bars are the weight bytes at each precision; the dashed line is the usable memory of an Apple M4 Max (128GB). Drawn from the computed byte counts, not sketched.

Precision	Bits/weight	Weights	Source	Fits	Max context	Ceiling
As released	5.12	12.8 GiB	measured file	yes	131,072 (model cap)	184 tok/s
Q8_0	4.50	11.3 GiB	measured file	yes	131,072 (model cap)	208 tok/s
Q6_K	4.48	11.2 GiB	measured file	yes	131,072 (model cap)	209 tok/s
Q5_K_M	4.36	10.9 GiB	measured file	yes	131,072 (model cap)	214 tok/s
Q4_K_M	4.32	10.8 GiB	measured file	yes	131,072 (model cap)	216 tok/s
Q3_K_M	4.28	10.7 GiB	measured file	yes	131,072 (model cap)	218 tok/s

What the context actually costs

The KV cache is 2 · layers · kv_heads · head_dim bytes per token per element — 2 · 24 · 8 · 64 · 2 B = 48.0 KiB. It depends on the key/value head count, not the attention-head count and not the parameter count. This model shares 8 KV heads across 64 query heads, which divides the cache by 8 against multi-head attention.

Sliding-window attention. 12 of this model's 24 layers attend to a 128-token window and stop growing there; only the remaining 12 keep scaling with context. That is why the cache figures below flatten out — and why a calculator that ignores the layer pattern over-states this model's long-context footprint several times over.

Context	KV cache (FP16)	KV cache (8-bit)	Plus Q4_K_M weights
4,096	99 MiB	50 MiB	10.9 GiB
8,192	195 MiB	98 MiB	11.0 GiB
32,768	771 MiB	386 MiB	11.6 GiB
131,072	3.0 GiB	1.5 GiB	13.8 GiB

The speed ceiling, and where it comes from

Generating one token reads every active weight from memory once. At FP16 / BF16 that is 2.6 GiB — only 20% of the checkpoint, because this is a mixture of experts and each token is routed to 4 of 32, plus a pass over the KV cache. An Apple M4 Max (128GB) moves 546 GB/s, so the arithmetic ceiling is 184 tok/s. Treat it as a bound, not an estimate: attention overhead, kernel launches and imperfect memory access keep real runtimes at roughly 60–80% of it, and nothing pushes past it.

Where these numbers come from

The model

Parameters: 21,511,953,984
Layers: 24
Attention / KV heads: 64 / 8
Head dimension: 64
Trained context: 131,072
Checkpoint as published: 12.8 GiB

Read from openai/gpt-oss-20b. The parameter count is the Hub's own total over the tensor shapes, not a figure taken from the model's name. This release is already quantized — its published checkpoint is 12.8 GiB, far below the 40.1 GiB that two bytes per parameter would imply, so the FP16 row above describes a file nobody ships.

The accelerator

Memory: 128 GB LPDDR5X unified
Bandwidth: 546 GB/s
Assumed usable: 75% → 96.0 GiB

Capacity and bandwidth from the vendor's specification. The usable fraction is an assumption, not a spec: unified memory is shared with the OS and the display, and the GPU working-set cap is raisable on Apple silicon with `sudo sysctl iogpu.wired_limit_mb`.

Questions this pairing answers

How much VRAM does gpt-oss 20B need?

12.8 GiB for the weights at FP16 — 21.5B parameters at two bytes each — and 10.8 GiB at Q4_K_M. The KV cache is on top of that and is not a fixed number: this model spends 48.0 KiB per token of context on its full-attention layers, so 8,192 tokens costs a further 195 MiB. An Apple M4 Max (128GB) makes 96.0 GiB of its 128 GB available on the assumption below.

Can an Apple M4 Max (128GB) run gpt-oss 20B?

Yes — the unquantized weights fit with room for 128k tokens of context. That is the weights and the KV cache together, against 96.0 GiB of usable memory.

How fast will gpt-oss 20B run on an Apple M4 Max (128GB)?

No faster than 184 tokens/second at FP16 / BF16, and in practice below it. Decoding is memory-bound: every token reads the 20% of weights this MoE routes to 2.6 GiB of weights plus the cache, and this card moves 546 GB/s. That division is the ceiling — no kernel, runtime or driver beats it, and a real runtime typically reaches 60–80% of it.

Why does the context length change how much memory gpt-oss 20B needs?

Because the KV cache holds one key and one value vector per token, per layer, for the whole conversation, and it is allocated separately from the weights. This model has 24 layers and 8 key/value heads of 64 dimensions, shared across 64 query heads — grouped-query attention, which divides the cache by 8. That works out at 48.0 KiB per token, except that 12 of the 24 layers use a 128-token sliding window and stop growing there. Parameter count tells you nothing about this number.

The same model on a different card

A different model on the same card

All 62 models on Apple M4 Max (128GB) →

Method and limits. Weight bytes are the byte size of the real published file wherever one exists, and the model's exact parameter count times the published llama.cpp bits-per-weight where it does not. That distinction is on every row above and it matters at both ends: a sub-1B model's Q4_K_M file runs a third larger than the nominal figure because k-quants keep its embedding tables at higher precision, and an already-4-bit release cannot be quantized upward at all. The KV cache is 2 · layers · kv_heads · head_dim · bytes per token, summed over layers with each sliding-window layer capped at its window. The speed figure is a roofline bound, not a benchmark: bandwidth divided by bytes read per token, which no runtime exceeds and every runtime falls short of. The usable fraction of card memory is an assumption: 75% here, which is what this lane assumes for unified memory, where the OS and window server share the same pool — the other class assumes 92%, so it is not one number applied to every device. That is the only assumed input on this page; every other figure is computed from the model config and the card's published specification. Nothing on this page is written by a language model. Architecture from the model's published config, fetched 2026-08-07.