What GPU do you need for Qwen3.8, Qwen3.6, Gemma 4 and gpt-oss locally? (measured, Oct 2026)

RigBase — Sat, 24 Oct 2026 19:12:56 +0000

Every time a new open model comes out, people ask the same thing: will it run on my card, and how fast? Spec-sheet calculators can tell you whether the weights fit. They can't tell you what speed you'll actually get. So here are measured generation speeds (decode tokens/sec) for the models people are running in October 2026, collected from public benchmarks. Every number is linked to the place it came from.

How to read the numbers

Speeds are generation (decode) tokens/sec at short context unless noted. Prompt processing is a different number and is usually much faster.
Sources use different engines, quants and llama.cpp builds. Treat gaps under ~10% as noise.
"llamabench community" rows are aggregated llama-bench tg128 runs from llamabench.ai. Open the source for the exact quant and build.

TL;DR

If you have…	What runs well (measured)
12 GB (RTX 3060)	Gemma 4 12B at 38.5 tok/s, gpt-oss 20B at 84 tok/s
16 GB (RTX 5060 Ti / 5070 Ti / 5080)	gpt-oss 20B at 112 / 189 / 205 tok/s
24 GB (RTX 3090 / 4090)	Qwen3.6 35B-A3B at 140–153 tok/s on a 3090; Qwen3.8 27B at 40.4 tok/s (72.5 with MTP) on a 3090; Gemma 4 31B at 43.8 tok/s on a 4090
32 GB (RTX 5090)	Qwen3.6 35B-A3B at 258, Qwen3.8 27B at 75.2, Gemma 4 31B at 70 tok/s
96 GB+ (RTX Pro 6000, Macs)	gpt-oss 120B at 196 tok/s (Pro 6000) / 79.7 tok/s (M2 Ultra)

Details and sources below.

1. gpt-oss 20B: the one model everyone has benchmarked

gpt-oss 20B runs in its native MXFP4 format, so these numbers compare well across cards. These are llama-bench tg128 results from the official llama.cpp gpt-oss guide (discussion #15396), plus Apple results:

Hardware	Memory	gpt-oss 20B tok/s	Source
RTX Pro 6000 (Blackwell)	96 GB	286.91	llama.cpp #15396
RTX 5090	32 GB	282.51	llama.cpp #15396
RTX 5080	16 GB	204.85	llama.cpp #15396
RTX 5070 Ti	16 GB	189.45	llama.cpp #15396
RTX 3090	24 GB	161.77	llama.cpp #15396
Apple M2 Ultra (76-core GPU)	192 GB	116.08	llama.cpp #4167
Apple M3 Ultra	512 GB	115.52	llama.cpp #15396
RTX 5060 Ti 16GB	16 GB	111.51	llama.cpp #15396
Apple M4 Max (MacBook)	36 GB	92.36	llama.cpp #15396
Apple M1 Max	64 GB	75.15	llama.cpp #15396

Takeaway: a 16 GB card is enough for gpt-oss 20B. Even the 5060 Ti does over 100 tok/s, which is far faster than you can read.

Speed and power: whole-system measurements

Jeff Geerling's ai-benchmarks records peak whole-system power during each run. That data is rare. (His summary table doesn't state the quant.)

System	gpt-oss 20B tok/s	Peak system power	tok/s per W (arithmetic)	Source
Core Ultra 265K + RTX 4090	279.65	509.1 W	0.55	#43
Core Ultra 265K + Radeon AI Pro R9700 (32 GB)	163.26	332.9 W	0.49	#35
Core Ultra 265K + RTX 4070 Ti	163.14	298.7 W	0.55	#45
Mac Studio M3 Ultra 512 GB	115.29	227 W	0.51	#42
Core Ultra 265K + RTX A4000	93.55	199.4 W	0.47	#36
Dell Pro Max GB10 (DGX Spark)	87.97	122.8 W	0.72	#34
Core Ultra 265K + RTX 3060	83.96	229.3 W	0.37	#40

The DGX Spark is the slowest box here on raw speed, but it gets the most tokens per watt. The 4090 is the fastest, at about 4× the Spark's peak draw.

2. gpt-oss 120B: you need ~96 GB of fast memory

Hardware	gpt-oss 120B tok/s	Source
RTX Pro 6000 (Blackwell, 96 GB)	196.31	llama.cpp #15396
Apple M2 Ultra 192 GB	79.68	llama.cpp #4167

3. Qwen3.6 35B-A3B: the 24 GB sweet spot

It's a mixture-of-experts model with only ~3B active parameters, so it runs very fast once it fits.

Hardware	Setup	tok/s	Source
RTX 5090 (32 GB)	llama-bench tg128	258	llamabench.ai
4× RTX 3090 (using 2)	vLLM, AWQ-INT4, TP=2 + MTP-3 speculative decoding	225–231	tfriedel/qwen3.6-rtx3090-lab
4× RTX 3090 (using 2)	vLLM, AWQ-INT4, TP=2	149	same
RTX 3090 (24 GB)	llama-bench tg128	146–153	llamabench.ai
RTX 3090 (24 GB)	Q4 GGUF, all layers on GPU, ~90K context	139.6	Giles Thomas
RTX 3090 (24 GB)	llama.cpp, IQ4_XS, 1 GPU	115–133	tfriedel lab
RTX 4090 (24 GB)	llama.cpp, Q4_K_M	160.6	llm-speed
Apple M5 Max (36 GB)	oMLX, 4-bit MLX	145.8	Rapid-MLX compare
Intel Arc Pro B70 (32 GB)	vLLM XPU, GPTQ Int4, no MTP	120	riceharvest cookbook
DGX Spark (GB10)	vLLM, NVFP4, 128K ctx, no spec. decoding	78.1	Cristian Tala
Strix Halo (Radeon 8060S, 96 GB)	llama-bench tg128	50.2–62.4	llamabench.ai

Two things stand out:

One used RTX 3090 is enough. Giles Thomas got 139.6 tok/s while keeping ~90K tokens of context on the GPU. At the full 262K context, with 12 layers of FFNs moved to the CPU, he still got about 66 tok/s (source).
A second 3090 helps through vLLM, not llama.cpp. Tensor parallel with AWQ went from 115–133 to 149 tok/s, and adding MTP speculative decoding pushed it to 225–231.

4. Qwen3.8 27B: the new default dense model

Qwen3.8 27B replaced Qwen3.6 27B and is now the most-benchmarked model in RigBase (50 results). It's dense, so every token reads all ~27B weights, and speed follows memory bandwidth. These are plain decode numbers, without speculative decoding:

Hardware	Setup	tok/s	Source
RTX 5090 (32 GB)	llama.cpp, Q4_K_M	75.2	llm-speed
RTX 3090 (24 GB)	OpenJet megakernel, Q4_K_M	42.9	OpenJet
RTX 3090 (24 GB)	llama.cpp, Q4_K_M	40.4	OpenJet
Apple M5 Max (128 GB)	MLX 4-bit	31	Baobab Tech
Radeon AI Pro R9700 (32 GB)	Ollama, Q4_K_M (includes hidden reasoning tokens)	30	vaditaslim
RTX 5060 Ti 16GB	llama.cpp, UD-IQ4_XS, 8K ctx	28.0	JESVS
Intel Arc Pro B70 (32 GB)	llama.cpp SYCL, Q4_K_M	27.8	b70-optimization-lab
Strix Halo (Radeon 8060S)	llama.cpp ROCm, Q4_K_M	12.6	AkitaOnRails
DGX Spark (GB10)	vLLM, 128K ctx	11.0	Cristian Tala
RTX 3060 12GB	llama-bench (doesn't fit, offloads)	5.2–6.7	llamabench.ai

Turn on MTP. Qwen3.8 supports multi-token-prediction (MTP) drafting, and in same-hardware tests it helps a lot. On the RTX 3090 it took llama.cpp from 40.4 to 72.5 tok/s on a code prompt (50.9 on a reasoning prompt) (OpenJet). On Strix Halo it took Vulkan Q6_K from 9.59 to 18.01 (Sleeping Robots), and on the DGX Spark it took vLLM from 11.0 to 19.0 (Cristian Tala).

5. Gemma 4 and Qwen3.6 27B: dense models, where bandwidth wins

Dense models read every weight for every token, so speed follows memory bandwidth. Here are llamabench community tg128 results:

Model	RTX 5090	RTX 4090	5090 ÷ 4090 (arithmetic)
Gemma 4 31B	70.0	43.8	1.60×
Qwen3.6 27B (legacy, now Qwen3.8 27B)	75.7	47.0	1.61×
Gemma 4 26B-A4B (MoE)	235	164	1.43×
Gemma 4 12B	142	98.1	1.45×
gpt-oss 20B	313	203	1.54×

Sources: RTX 5090, RTX 4090.

On dense ~30B models, the 5090 is about 1.6× faster than the 4090. On the MoE models the gap is smaller.

For the RTX 3090, llamabench lists 40.2 tok/s on Qwen3.6 27B (source). That row includes some runs with speculative decoding, so don't compare it directly with the table above.

Gemma 4 12B on smaller cards (llamabench community):

Card	Gemma 4 12B tok/s	Source
RTX 5070 Ti (16 GB)	87.6	llamabench.ai
RTX 3060 (12 GB)	38.5	llamabench.ai

So what should you buy?

Based only on the numbers above:

gpt-oss 20B / Gemma 4 12B: any 16 GB card works. A 12 GB RTX 3060 runs both at usable speeds.
Qwen3.6 35B-A3B: a single 24 GB card. A used RTX 3090 does 140–153 tok/s.
Dense ~30B (Qwen3.8 27B, Gemma 4 31B): 24 GB fits them at Q4. A used 3090 does ~40 tok/s on Qwen3.8 27B (≈72 with MTP); the RTX 5090 does 75.2 and is ~1.6× faster than a 4090 on Gemma 4 31B.
gpt-oss 120B: a 96 GB GPU or a big-memory Mac.
Lowest power: the DGX Spark had the best tokens-per-watt on gpt-oss 20B in Geerling's tests, though it was the slowest box.

Missing your card?

The full dataset has 117 builds and 298 measured results, with parts lists, power draw and a link to every source: RigBase. Browse by GPU or by VRAM size, check fit on the Can it run? calculator, or download it all free as CSV/JSON (CC BY 4.0).

If you've run llama-bench on a current model, submit it. It needs a proof link, and that's the only way a number gets in.

Data snapshot: RigBase builds.json, first published 19 Oct 2026, updated 24 Oct 2026. Corrections welcome. Every row links to its source, so check there first.

DEV Community: RigBase