Qwen3 30B-A3B (2507) hardware requirements

What is Qwen3 30B-A3B (2507)?

This is Qwen3-30B-A3B-Instruct-2507, the July 2025 update from Alibaba Cloud's Qwen team to the non-thinking mode of Qwen3 30B-A3B. It answers directly with no thinking step, and its model card reports better instruction following, coding, tool use and understanding of long inputs up to its native 256k context. It is a mixture-of-experts model with about 30B parameters, of which about 3B are active per token.

Model card

What to know before running Qwen3 30B-A3B (2507)

Qwen3 30B-A3B is a mixture-of-experts model: about 30B parameters in total, but only about 3B are used for each token. At Q4_K_M it needs 19.8 GB with an 8k context, so a 24GB card holds it completely and runs it great; see our page for Qwen3 30B-A3B (2507) on the GeForce RTX 3090. On smaller cards, the experts that do not fit stay in system RAM, and because each token only touches a few of them, CanRun still expects it to run well on 8GB to 16GB cards such as the GeForce RTX 4060 and the GeForce RTX 5060 Ti 16GB. Those are theoretical estimates (±30%) that depend on your system RAM speed.

This is the 2507 Instruct release, which has no thinking mode: it answers directly, as the model card says. CanRun therefore uses the normal speed bar for it, not the stricter one for thinking models. On a Mac, the 48GB Apple M4 Pro runs it great. Even without a graphics card, a dual-channel DDR5 desktop with 32GB runs it. CanRun caps every CPU-only setup at Runs slowly whatever its speed, though its calibrated estimate there clears the Runs well line.

About the model

Parameters: 30.5B total · 3.3B active (MoE)
Attention: Full attention
Max context: 262,144 tokens
Released: 2025-07-31
Model card: Qwen3 30B-A3B (2507) on Hugging Face

Other sizes in this series

Graphics card memory each model needs at its baseline quant with 8k context, and its verdict on common graphics cards, smallest model first
Model	Parameters	Baseline quant	Memory needed (8k)	GeForce RTX 4060	GeForce RTX 5060 Ti 16GB	GeForce RTX 3090
Qwen3 8B	8.2B	Q4_K_M	6.5 GB	Runs great	Runs great	Runs great
Qwen3 14B	14.8B	Q4_K_M	10.5 GB	Runs slowly	Runs great	Runs great
Qwen3 30B-A3B (2507)this model	30.5B total · 3.3B active (MoE)	Q4_K_M	19.8 GB	Runs well	Runs well	Runs great
Qwen3 32B	32.8B	Q4_K_M	22.0 GB	Won't run	Runs slowly	Runs well

GGUF files

Tracked quantizations of Qwen3 30B-A3B (2507), smallest first
Quant	File size	Quality	Published by
Q2_K	11.26 GB	Heavy loss	Community GGUF · unsloth
IQ4_XS	16.38 GB	Good balance	Community GGUF · unsloth
Q4_K_Mbaseline	18.56 GB	Good balance	Community GGUF · unsloth
Q8_0	32.48 GB	Near-lossless	Community GGUF · unsloth

Some repositories publish only Unsloth Dynamic (UD) variants; sizes are those of the file you would download.

Memory needed by quant and context

Graphics card memory in GB: weights loaded on the GPU + f16 KV cache + compute buffer. Add about 0.6 GB if the GPU also drives your display on Windows. Part of the file stays off the graphics card: llama.cpp keeps the input embedding table in system RAM. A Mac counts the whole file, so it needs a little more.
Quant	4k	8k	16k	32k	64k	128k	256k
Q2_K	12.1	12.5	13.4	15.2	18.7	25.7	39.8
IQ4_XS	17.1	17.6	18.5	20.2	23.7	30.8	44.9
Q4_K_M	19.3	19.8	20.6	22.4	25.9	33.0	47.0
Q8_0	33.1	33.5	34.4	36.2	39.7	46.7	60.8

Which GPUs can run Qwen3 30B-A3B (2507)?

How to read the verdicts

Runs great: Fully on the GPU at 20 tok/s or more
Runs well: Fully on the GPU at 8–20 tok/s; MoE experts in system RAM at 20 tok/s or more; or at least 90% on the GPU at 8 tok/s or more
Runs slowly: 2–8 tok/s; CPU-only; less than 90% on the GPU; or MoE experts in system RAM below 20 tok/s
Won't run: Does not fit, or under 2 tok/s

Sort

Filter this list by budget, type and context in the calculator

Scroll sideways to see every column.

58 GPUs, Macs and CPU setups at Q4_K_M with 8k context, sorted by verdict and speed
Hardware	Verdict	Speed	Quant	Memory	Runs as	Context	Notes	Price
GeForce RTX 5090	Runs greatQwen3 30B-A3B (2507) verdict on the GeForce RTX 5090	est. 195.0 tok/s156.0–234.1calibrated estimate ±20%	Q4_K_M	20.4 / 32.0 GB	Full GPU	up to 64k	—	$4,200 street (as of 2026-08-09)
NVIDIA RTX PRO 6000 Blackwell	Runs great	est. 195.0 tok/s156.0–234.1calibrated estimate ±20%	Q4_K_M	20.4 / 96.0 GB	Full GPU	up to 256k	—	—
NVIDIA RTX PRO 6000 Blackwell Max-Q	Runs great	est. 195.0 tok/s156.0–234.1calibrated estimate ±20%	Q4_K_M	20.4 / 96.0 GB	Full GPU	up to 256k	—	—
GeForce RTX 3090 Ti	Runs great	est. 146.7 tok/s117.4–176.1calibrated estimate ±20%	Q4_K_M	20.4 / 24.0 GB	Full GPU	up to 32k	—	$1,999 launch MSRP
GeForce RTX 4090	Runs greatQwen3 30B-A3B (2507) verdict on the GeForce RTX 4090	est. 146.7 tok/s117.4–176.1calibrated estimate ±20%	Q4_K_M	20.4 / 24.0 GB	Full GPU	up to 32k	—	$2,755 street (as of 2026-08-09)
GeForce RTX 3090	Runs greatQwen3 30B-A3B (2507) verdict on the GeForce RTX 3090	est. 140.6 tok/s112.5–168.7calibrated estimate ±20%	Q4_K_M	20.4 / 24.0 GB	Full GPU	up to 32k	—	$1,050 street (as of 2026-08-09)
GeForce RTX 5090 Laptop	Runs great	est. 137.0 tok/s109.6–164.5calibrated estimate ±20%	Q4_K_M	20.4 / 24.0 GB	Full GPU	up to 32k	—	—
Apple M5 Ultra · 96 GB	Runs great	est. 106.6 tok/s85.3–127.9calibrated estimate ±20%	Q4_K_M	19.9 / 67.2 GB	Unified memory	up to 256k	—	—
Apple M5 Max (40-core GPU) · 48 GB	Runs great	est. 104.2 tok/s83.4–125.0calibrated estimate ±20%	Q4_K_M	19.9 / 33.6 GB	Unified memory	up to 128k	—	—
Radeon RX 7900 XTX	Runs great	est. 103.3 tok/s82.6–123.9calibrated estimate ±20%	Q4_K_M	20.4 / 24.0 GB	Full GPU	up to 32k	—	$999 launch MSRP
Apple M4 Max (40-core GPU) · 48 GB	Runs great	est. 96.9 tok/s77.5–116.2calibrated estimate ±20%	Q4_K_M	19.9 / 33.6 GB	Unified memory	up to 128k	—	—
Apple M3 Ultra · 96 GB	Runs great	est. 88.3 tok/s70.7–106.0calibrated estimate ±20%	Q4_K_M	19.9 / 67.2 GB	Unified memory	up to 128k	—	—
Apple M2 Ultra · 64 GB	Runs great	est. 87.3 tok/s69.8–104.7calibrated estimate ±20%	Q4_K_M	19.9 / 44.8 GB	Unified memory	up to 128k	—	—
Apple M5 Max (32-core GPU) · 36 GB	Runs great	est. 86.6 tok/s69.3–103.9calibrated estimate ±20%	Q4_K_M	19.9 / 25.2 GB	Unified memory	up to 32k	—	—
Apple M4 Max (32-core GPU) · 36 GB	Runs great	est. 80.0 tok/s64.0–96.0calibrated estimate ±20%	Q4_K_M	19.9 / 25.2 GB	Unified memory	up to 32k	—	—
Apple M5 Pro · 48 GB	Runs great	est. 64.8 tok/s51.8–77.8calibrated estimate ±20%	Q4_K_M	19.9 / 33.6 GB	Unified memory	up to 64k	—	—
NVIDIA DGX Spark · 128 GB	Runs great	est. 60.7 tok/s48.6–72.9calibrated estimate ±20%	Q4_K_M	19.9 / 116.9 GB	Unified memory	up to 64k	—	$3,999 launch MSRP (128 GB)
Apple M4 Pro · 48 GB	Runs greatQwen3 30B-A3B (2507) verdict on the Apple M4 Pro	est. 59.2 tok/s47.4–71.1calibrated estimate ±20%	Q4_K_M	19.9 / 33.6 GB	Unified memory	up to 64k	—	—
Ryzen AI Max+ 395 (Strix Halo) · 32 GB	Runs great	est. 54.2 tok/s43.3–65.0calibrated estimate ±20%	Q4_K_M	19.9 / 22.4 GB	Unified memory	up to 16k	—	$1,999 launch MSRP (128 GB)
Apple M6 · 32 GB	Runs great	est. 40.4 tok/s32.3–48.5calibrated estimate ±20%	Q4_K_M	19.9 / 22.4 GB	Unified memory	up to 16k	—	—
Apple M5 · 32 GB	Runs great	est. 36.9 tok/s29.5–44.3calibrated estimate ±20%	Q4_K_M	19.9 / 22.4 GB	Unified memory	up to 16k	—	—
Apple M4 · 32 GB	Runs greatQwen3 30B-A3B (2507) verdict on the Apple M4	est. 29.8 tok/s23.8–35.7calibrated estimate ±20%	Q4_K_M	19.9 / 22.4 GB	Unified memory	up to 16k	—	—
Radeon RX 7900 XT	Runs well	est. 87.9 tok/s61.5–114.3theoretical estimate ±30%	Q4_K_M	19.6 / 20.0 GB + 0.9 GB RAM	MoE experts in RAM	up to 128k	Not fully on the GPU — Runs great needs the whole model in GPU memory Try IQ4_XS: Runs great	$899 launch MSRP
GeForce RTX 5080	Runs well	est. 75.8 tok/s53.0–98.5theoretical estimate ±30%	Q4_K_M	15.6 / 16.0 GB + 4.9 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory Q2_K (heavy quality loss): Runs great	$1,256 street (as of 2026-08-09)
GeForce RTX 5070 Ti	Runs well	est. 74.3 tok/s52.0–96.6theoretical estimate ±30%	Q4_K_M	15.6 / 16.0 GB + 4.9 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory Q2_K (heavy quality loss): Runs great	$949 street (as of 2026-08-09)
GeForce RTX 5080 Laptop	Runs well	est. 74.3 tok/s52.0–96.6theoretical estimate ±30%	Q4_K_M	15.6 / 16.0 GB + 4.9 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory Q2_K (heavy quality loss): Runs great	—
GeForce RTX 4080 Super	Runs well	est. 70.0 tok/s49.0–90.9theoretical estimate ±30%	Q4_K_M	15.6 / 16.0 GB + 4.9 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory Q2_K (heavy quality loss): Runs great	$999 launch MSRP
GeForce RTX 4080	Runs well	est. 69.3 tok/s48.5–90.2theoretical estimate ±30%	Q4_K_M	15.6 / 16.0 GB + 4.9 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory Q2_K (heavy quality loss): Runs great	$1,199 launch MSRP
GeForce RTX 4070 Ti Super	Runs well	est. 67.8 tok/s47.5–88.2theoretical estimate ±30%	Q4_K_M	15.6 / 16.0 GB + 4.9 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory Q2_K (heavy quality loss): Runs great	$799 launch MSRP
GeForce RTX 4090 Laptop	Runs well	est. 64.1 tok/s44.9–83.3theoretical estimate ±30%	Q4_K_M	15.6 / 16.0 GB + 4.9 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory Q2_K (heavy quality loss): Runs great	—
GeForce RTX 5060 Ti 16GB	Runs wellQwen3 30B-A3B (2507) verdict on the GeForce RTX 5060 Ti 16GB	est. 57.7 tok/s40.4–75.1theoretical estimate ±30%	Q4_K_M	15.6 / 16.0 GB + 4.9 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory Q2_K (heavy quality loss): Runs great	$569 street (as of 2026-08-09)
Radeon RX 9070	Runs well	est. 57.0 tok/s39.9–74.2theoretical estimate ±30%	Q4_K_M	15.6 / 16.0 GB + 4.9 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory Q2_K (heavy quality loss): Runs great	$639 street (as of 2026-08-09)
Radeon RX 9070 XT	Runs well	est. 57.0 tok/s39.9–74.2theoretical estimate ±30%	Q4_K_M	15.6 / 16.0 GB + 4.9 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory Q2_K (heavy quality loss): Runs great	$689 street (as of 2026-08-09)
Radeon RX 7800 XT	Runs well	est. 56.5 tok/s39.6–73.5theoretical estimate ±30%	Q4_K_M	15.6 / 16.0 GB + 4.9 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory Q2_K (heavy quality loss): Runs great	$499 launch MSRP
GeForce RTX 3080 12GB	Runs well	est. 53.8 tok/s37.6–69.9theoretical estimate ±30%	Q4_K_M	11.6 / 12.0 GB + 9.0 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$799 launch MSRP
GeForce RTX 5070	Runs wellQwen3 30B-A3B (2507) verdict on the GeForce RTX 5070	est. 50.5 tok/s35.3–65.6theoretical estimate ±30%	Q4_K_M	11.6 / 12.0 GB + 9.0 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$629 street (as of 2026-08-09)
GeForce RTX 5070 Ti Laptop	Runs well	est. 50.5 tok/s35.3–65.6theoretical estimate ±30%	Q4_K_M	11.6 / 12.0 GB + 9.0 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory	—
GeForce RTX 4070	Runs wellQwen3 30B-A3B (2507) verdict on the GeForce RTX 4070	est. 46.8 tok/s32.8–60.9theoretical estimate ±30%	Q4_K_M	11.6 / 12.0 GB + 9.0 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$599 launch MSRP
GeForce RTX 4070 Super	Runs well	est. 46.8 tok/s32.8–60.9theoretical estimate ±30%	Q4_K_M	11.6 / 12.0 GB + 9.0 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$599 launch MSRP
GeForce RTX 2080 Ti	Runs well	est. 46.3 tok/s32.4–60.2theoretical estimate ±30%	Q4_K_M	10.5 / 11.0 GB + 10.1 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$999 launch MSRP
GeForce RTX 4060 Ti 16GB	Runs well	est. 46.3 tok/s32.4–60.1theoretical estimate ±30%	Q4_K_M	15.6 / 16.0 GB + 4.9 GB RAM	MoE experts in RAM	up to 32k	Not fully on the GPU — Runs great needs the whole model in GPU memory Q2_K (heavy quality loss): Runs great	$499 launch MSRP
GeForce RTX 3080 10GB	Runs well	est. 45.3 tok/s31.7–58.9theoretical estimate ±30%	Q4_K_M	9.4 / 10.0 GB + 11.2 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$699 launch MSRP
GeForce RTX 4080 Laptop	Runs well	est. 44.7 tok/s31.3–58.1theoretical estimate ±30%	Q4_K_M	11.6 / 12.0 GB + 9.0 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory	—
Radeon RX 9060 XT 16GB	Runs well	est. 44.5 tok/s31.1–57.8theoretical estimate ±30%	Q4_K_M	15.6 / 16.0 GB + 4.9 GB RAM	MoE experts in RAM	up to 32k	Not fully on the GPU — Runs great needs the whole model in GPU memory Q2_K (heavy quality loss): Runs great	$449 street (as of 2026-08-09)
GeForce GTX 1080 Ti	Runs well	est. 43.6 tok/s30.5–56.7theoretical estimate ±30%	Q4_K_M	10.5 / 11.0 GB + 10.1 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$699 launch MSRP
GeForce RTX 3060 12GB	Runs wellQwen3 30B-A3B (2507) verdict on the GeForce RTX 3060 12GB	est. 42.0 tok/s29.4–54.6theoretical estimate ±30%	Q4_K_M	11.6 / 12.0 GB + 9.0 GB RAM	MoE experts in RAM	up to 32k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$250 street (as of 2026-08-09)
Intel Arc B580	Runs well	est. 38.1 tok/s26.6–49.5theoretical estimate ±30%	Q4_K_M	11.6 / 12.0 GB + 9.0 GB RAM	MoE experts in RAM	up to 32k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$290 street (as of 2026-08-09)
GeForce RTX 3070	Runs well	est. 37.1 tok/s26.0–48.3theoretical estimate ±30%	Q4_K_M	7.6 / 8.0 GB + 13.0 GB RAM	MoE experts in RAM	up to 32k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$499 launch MSRP
GeForce RTX 5060	Runs well	est. 37.1 tok/s26.0–48.3theoretical estimate ±30%	Q4_K_M	7.6 / 8.0 GB + 13.0 GB RAM	MoE experts in RAM	up to 32k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$339 street (as of 2026-08-09)
GeForce RTX 5060 Ti 8GB	Runs well	est. 37.1 tok/s26.0–48.3theoretical estimate ±30%	Q4_K_M	7.6 / 8.0 GB + 13.0 GB RAM	MoE experts in RAM	up to 32k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$429 street (as of 2026-08-09)
GeForce RTX 4060 Ti 8GB	Runs well	est. 33.0 tok/s23.1–42.8theoretical estimate ±30%	Q4_K_M	7.6 / 8.0 GB + 13.0 GB RAM	MoE experts in RAM	up to 32k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$399 launch MSRP
GeForce RTX 4060	Runs wellQwen3 30B-A3B (2507) verdict on the GeForce RTX 4060	est. 32.4 tok/s22.6–42.1theoretical estimate ±30%	Q4_K_M	7.6 / 8.0 GB + 13.0 GB RAM	MoE experts in RAM	up to 32k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$299 launch MSRP
GeForce RTX 4060 Laptop	Runs well	est. 31.7 tok/s22.2–41.2theoretical estimate ±30%	Q4_K_M	7.6 / 8.0 GB + 13.0 GB RAM	MoE experts in RAM	up to 32k	Not fully on the GPU — Runs great needs the whole model in GPU memory	—
GeForce RTX 4070 Laptop	Runs well	est. 31.7 tok/s22.2–41.2theoretical estimate ±30%	Q4_K_M	7.6 / 8.0 GB + 13.0 GB RAM	MoE experts in RAM	up to 32k	Not fully on the GPU — Runs great needs the whole model in GPU memory	—
GeForce RTX 3050 8GB	Runs well	est. 30.2 tok/s21.2–39.3theoretical estimate ±30%	Q4_K_M	7.6 / 8.0 GB + 13.0 GB RAM	MoE experts in RAM	up to 16k	Not fully on the GPU — Runs great needs the whole model in GPU memory	$249 launch MSRP
DDR5-6000 dual-channel CPU · 32 GB	Runs slowly	est. 14.7 tok/s11.7–17.6calibrated estimate ±20%	Q4_K_M	19.4 GB RAM	CPU only	up to 64k	CPU inference — capped at Runs slowly	—
DDR5-5600 dual-channel CPU · 32 GB	Runs slowly	est. 13.7 tok/s11.0–16.4calibrated estimate ±20%	Q4_K_M	19.4 GB RAM	CPU only	up to 64k	CPU inference — capped at Runs slowly	—
DDR4-3200 dual-channel CPU · 32 GB	Runs slowly	est. 7.8 tok/s6.3–9.4calibrated estimate ±20%	Q4_K_M	19.4 GB RAM	CPU only	up to 64k	—	—

Cheapest GPUs that run it

Runs greatGeForce RTX 3090$1,050 street (as of 2026-08-09), est. 140.6 tok/s (112.5–168.7, calibrated estimate ±20%).
Runs wellGeForce RTX 3060 12GB$250 street (as of 2026-08-09), est. 42.0 tok/s (29.4–54.6, theoretical estimate ±30%).

Measured results for Qwen3 30B-A3B (2507)

Public benchmarks we calibrate against. Their conditions (context, backend, flags) can differ from the estimates above.

Measured results for Qwen3 30B-A3B (2507)
Hardware	Quant	Backend	Context	Prompt (tok/s)	Generation (tok/s)	Flags	Source	Measured
GeForce RTX 5090	Q4_K_XL	llama.cpp	16k	—	141.6	—	hardware-corner.net	2026-08-09
GeForce RTX 4090	Q4_K_XL	llama.cpp	16k	—	139.7	—	hardware-corner.net	2026-08-09
GeForce RTX 3090	Q4_K_XL	llama.cpp	16k	—	113.8	—	hardware-corner.net	2026-08-09

Frequently asked questions

How much VRAM does Qwen3 30B-A3B (2507) need?

At Q4_K_M the file is 18.56 GB; with 8k context a graphics card needs about 19.8 GB (weights 18.37 GB, KV cache 0.8 GB, compute buffer 0.6 GB). The smallest tracked file (Q2_K (heavy quality loss)) needs about 12.5 GB. Part of the file stays off the graphics card: llama.cpp keeps the input embedding table in system RAM. A Mac counts the whole file, so it needs a little more.

What is the cheapest GPU that runs Qwen3 30B-A3B (2507) well?

Cheapest that runs it great: GeForce RTX 3090 at $1,050 (street price as of 2026-08-09), est. 140.6 tok/s (112.5–168.7, calibrated estimate ±20%). Cheaper and still Runs well: GeForce RTX 3060 12GB at $250 (street price as of 2026-08-09), est. 42.0 tok/s (29.4–54.6, theoretical estimate ±30%).

Can Qwen3 30B-A3B (2507) run on an 8 GB, 16 GB or 24 GB GPU?

At Q4_K_M with 8k context: GeForce RTX 4060 — Runs well, est. 32.4 tok/s (22.6–42.1, theoretical estimate ±30%); GeForce RTX 5060 Ti 16GB — Runs well, est. 57.7 tok/s (40.4–75.1, theoretical estimate ±30%); GeForce RTX 3090 — Runs great, est. 140.6 tok/s (112.5–168.7, calibrated estimate ±20%).

Speeds are estimates from memory bandwidth, calibrated against public benchmarks, and each one comes with an error band and a confidence label. Real results vary with drivers, backend, context length and thermals.