whatmodel.sh / Qwen3 Coder 30B-A3B Instruct

Qwen3 Coder 30B-A3B Instruct

+ contribute

Qwen3 Coder 30B-A3B Instruct is an open-weights language model from Alibaba, released July 2025. Its published size is 30.5B parameters (3.3B active per token) and a 160k-token context window. Licence: apache-2.0.

57.6±1.2#224 of 719 LLMsfrom 13 benchmark results across 3 sources

$ how it compares

The LLMs ranked either side of Qwen3 Coder 30B-A3B Instruct, out of 828 with a score, then more from Alibaba.

#	model	score	where	RAM	released
266	Qwen3-VL 8B Thinking Alibaba	58±1	api $0.66/M +local	~6.8GB	2025-10
267	Qwen3 32B Alibaba	58±0	api $0.13/M +local	~20GB	2025-04
268	Qwen3 Coder 30B-A3B Instruct Alibaba	58±1	api $0.12/M +local	~20GB	2025-07
269	Nova 2 Lite Amazon	58±2	api $0.94/M		2025-10
270	Llama 3.3 Nemotron Super 49B v1.5 NVIDIA	58±1	local	~18GB	2025-07
	more from Alibaba
18	Qwen3.8 Max (2.4T-A95B) Alibaba	78±1	api $3.00/M +local	~380GB	2026-08
34	Qwen3.8 Max (0902) Alibaba	76±1	api $3.00/M		2026-09
37	Qwen3.8-Flash-Next Alibaba	75±1	local	~59GB	2026-08

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #171 of 300 · 60% of its score · 8 tests

Beats o1 (2024-12-17), o3-mini, DeepSeek R1 0528 here, though they rank higher overall. Trails SmolLM3 3B, Jamba 1.5 Mini, Jamba 1.5 Large here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
SWE-bench Verified +1	31.8% ✎	▄▄·····▄▄▄▄█ beats 0/7	hard · avg 17%	▇▇▇▇▇▇▇▇ 21%	sharp · at its level · several runs
SWE-bench Multilingual +1	20.7% ✎	▄·▄·····█▄█▄ beats 0/7	hard · avg 18%	▇▇▇▇ 11%	at its level · several runs
Terminal-Bench 2.1	10.1% ✎	▄······█▄▄█▄ beats 0/7	very hard · avg 7.1%	▇▇▇▇ 9.4%	typical
SWE-bench Pro (split unspecified)	19.8% ✎	▄···▄···▄▄██ beats 1/7	very hard · avg 6.5%	▇▇▇ 7.2%	sharp
τ-bench Retail lab	68.7% ✎	▁····▃·▂▇█▂▆ beats 19/28	medium · avg 45%	▇▇ 6.1%	at its level · small run
τ-bench Airline lab	48.0% ✎	▁▁▁·▅▄▁█▇▁▁▇ beats 19/28	hard · avg 29%	▇▇ 4.2%	at its level · small run
BFCL v3 lab	62.2% ✎	▁▁▁▁▁▃▅█▆▅▄▁ beats 105/170	medium · avg 57%	▇ 1.6%	blunt · at its level · small run
SWE-bench Live lab	20.7% ✎	█··········█ beats 0/1		—	not counted yet: no other model in the same run to compare it with

coding▁▁▁▃▆██▇▆▄▃▁ #208 of 458 · 71% of its score · 7 tests

Beats Qwen3.5 4B, Qwen3-VL 32B Instruct, Solar Open 100B (102B-A12B) here, though they rank higher overall. Trails QwQ-32B-Preview, Ministral 3 3B Reasoning 2512, DeepSeek-Coder-V2 Instruct (0724) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
SWE-bench Verified +1	31.8% ✎	▄▄·····▄▄▄▄█ beats 0/7	hard · avg 17%	▇▇▇▇▇▇▇▇ 21%	sharp · at its level · several runs
Aider polyglot (225 exercises, two attempts) lab	33.3% ✎	▂▁▁▃▆▃▄▆█▃▄▂ beats 28/71	very hard · avg 10%	▇▇▇▇▇▇▇ 18%	sharp · at its level · small run
SWE-bench Multilingual +1	20.7% ✎	▄·▄·····█▄█▄ beats 0/7	hard · avg 18%	▇▇▇▇ 11%	at its level · several runs
Luau Leetcode fully solved (Torpedo Software)	38.5% ✎	▁▁▁▁·▃▂▃▅▃█▇ beats 6/49	hard · avg 18%	▇▇▇▇ 10%	at its level
SWE-bench Pro (split unspecified)	19.8% ✎	▄···▄···▄▄██ beats 1/7	very hard · avg 6.5%	▇▇▇ 7.2%	sharp
SciCode (level unspecified)	18.3% ✎	▄···▄▄█▄·▄·▄ beats 0/7	very hard · avg 13%	▇▇ 4.6%	typical
SWE-bench Live lab	20.7% ✎	█··········█ beats 0/1		—	not counted yet: no other model in the same run to compare it with

other tests1 not in a task · 6.6% of its score

test	result	vs the others in its run	difficulty	counts	why
RobloxQA v2.0 multiple choice (Torpedo Software)	85.0% ✎	▁▁▃▁█▄▃▃▆▂▂▂ beats 6/49	easy · avg 78%	▇▇▇ 6.6%	typical

$ where to run it

Through an API

api · Novita AI · $0.12/M · $0.07 in / $0.27 out per 1M tokens · 160k context · novita.ai ↗ · price sourceUSD per 1M text tokens, standard uncached input and output. Model endpoint qwen/qwen3-coder-30b-a3b-instruct. Product: Serverless chat completions. API model: qwen/qwen3-coder-30b-a3b-instruct. Settings: service tier=standard, batch=no, cache=uncached input. Additional/excluded charges: Tools, cache storage, priority service and other modalities excluded.. Currency: USD.

llama.cpp · recommended

version

Install llama.cpp · Homebrew and winget installs name the commands llama-cli and llama-server instead.
macOS / Linux
curl -LsSf https://llama.app/install.sh | sh
Windows
irm https://llama.app/install.ps1 | iex
other ways to install ↗
Run Qwen3 Coder 30B-A3B Instruct · needs ~18.6 GB · from huggingface.co/lmstudio-community
chat in the terminal
llama cli -hf lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-GGUF:Q4_K_M
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-GGUF:Q4_K_M

LM Studio

version

Install LM Studio · Open the app once before using the lms command.
download it (macOS / Windows / Linux) ↗
Run Qwen3 Coder 30B-A3B Instruct · needs ~18.6 GB · from huggingface.co/lmstudio-community
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-GGUF@Q4_K_M -y
Then run lms ls to see its name, and lms chat followed by that name

Ollama

version

Install Ollama
download it (macOS / Windows) ↗
Linux
curl -fsSL https://ollama.com/install.sh | sh
Run Qwen3 Coder 30B-A3B Instruct · needs ~18.8 GB · from ollama.com
chat in the terminal
ollama run qwen3-coder:30b

MLX (Mac)

version

Install MLX (Mac) · Apple silicon Macs only, text only.
with uv
uv tool install mlx-lm
or
pip install mlx-lm
Run Qwen3 Coder 30B-A3B Instruct · needs ~17.2 GB, Mac only · from huggingface.co/lmstudio-community
chat in the terminal
mlx_lm.chat --model lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-4bit
Or start an OpenAI-compatible server at http://localhost:8080/v1
mlx_lm.server --model lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-4bit

Official weights and documentation: huggingface.co/Qwen ↗

$ how the score was worked out · Qwen/Qwen3-Coder-30B-A3B-Inst… 42%, Kwaipilot/KAT-Coder-V2.5-Dev… 41%, Roblox LLM Leaderboard (Torpe… 16%

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

source	results	share of the score	licence
Qwen/Qwen3-Coder-30B-A3B-Instruct model card · lab-reported	6	▇▇▇▇▇▇▇▇▇▇▇▇▇ 42%	Apache-2.0
Kwaipilot/KAT-Coder-V2.5-Dev model card	5	▇▇▇▇▇▇▇▇▇▇▇▇ 41%	Apache-2.0
Roblox LLM Leaderboard (Torpedo Software): RobloxQA v2.0 and Luau Leetcode	2	▇▇▇▇▇ 16%	MIT
prior · pull toward the middle		▇ 1%
weighted average		= 57.6

not counted (1)

Qwen/Qwen3-Coder-30B-A3B-Instruct model card

SWE-bench Live [correct]

20.7%

fewer than three models have this test yet, or this source ran it for this model alone

$ published speed measurements · 36

Reference results for the hardware, engine and settings below. They aren't verified speeds of the downloads above or predictions for your machine.

NVIDIA DGX Spark (GB10)	Q8_0	61.06 tok/s	llama.cpp 11fb327bf (7941), CUDA · llama-bench; tg32; 32 generated tokens; context depth 0; n_ubatch=2048; flash attention=1; one sequence Published token-generation mean; not prompt processing or aggregate batch throughput. Hardware core count and installed RAM are not inferred.	source ↗
NVIDIA DGX Spark (GB10)	Q8_0	54.77 tok/s	llama.cpp 11fb327bf (7941), CUDA · llama-bench; tg32 @ d4096; 32 generated tokens; context depth 4096; n_ubatch=2048; flash attention=1; one sequence Published token-generation mean; not prompt processing or aggregate batch throughput. Hardware core count and installed RAM are not inferred.	source ↗
NVIDIA DGX Spark (GB10)	Q8_0	30.21 tok/s	llama.cpp 11fb327bf (7941), CUDA · llama-bench; tg32 @ d32768; 32 generated tokens; context depth 32768; n_ubatch=2048; flash attention=1; one sequence Published token-generation mean; not prompt processing or aggregate batch throughput. Hardware core count and installed RAM are not inferred.	source ↗
Apple M2 Ultra (GPU core count and RAM not specified)	Q8_0	78.97 tok/s	llama.cpp b828e18c7 (7948), Metal/BLAS · llama-bench; tg32; 32 generated tokens; context depth 0; n_ubatch=2048; flash attention=1; one sequence Published token-generation mean; not prompt processing or aggregate batch throughput. Hardware core count and installed RAM are not inferred.	source ↗
Apple M2 Ultra (GPU core count and RAM not specified)	Q8_0	71.18 tok/s	llama.cpp b828e18c7 (7948), Metal/BLAS · llama-bench; tg32 @ d4096; 32 generated tokens; context depth 4096; n_ubatch=2048; flash attention=1; one sequence Published token-generation mean; not prompt processing or aggregate batch throughput. Hardware core count and installed RAM are not inferred.	source ↗
Apple M2 Ultra (GPU core count and RAM not specified)	Q8_0	45.26 tok/s	llama.cpp b828e18c7 (7948), Metal/BLAS · llama-bench; tg32 @ d32768; 32 generated tokens; context depth 32768; n_ubatch=2048; flash attention=1; one sequence Published token-generation mean; not prompt processing or aggregate batch throughput. Hardware core count and installed RAM are not inferred.	source ↗
NVIDIA DGX Spark (GB10)	Q8_0	48.02 tok/s	llama.cpp 11fb327bf (7941), CUDA · llama-bench; tg32 @ d8192; 32 generated tokens; context depth 8192; n_ubatch=2048; flash attention=1; one sequence Published token-generation mean; not prompt processing or aggregate batch throughput. Hardware core count and installed RAM are not inferred.	source ↗
NVIDIA DGX Spark (GB10)	Q8_0	40.23 tok/s	llama.cpp 11fb327bf (7941), CUDA · llama-bench; tg32 @ d16384; 32 generated tokens; context depth 16384; n_ubatch=2048; flash attention=1; one sequence Published token-generation mean; not prompt processing or aggregate batch throughput. Hardware core count and installed RAM are not inferred.	source ↗
Apple M2 Ultra (GPU core count and RAM not specified)	Q8_0	65.11 tok/s	llama.cpp b828e18c7 (7948), Metal/BLAS · llama-bench; tg32 @ d8192; 32 generated tokens; context depth 8192; n_ubatch=2048; flash attention=1; one sequence Published token-generation mean; not prompt processing or aggregate batch throughput. Hardware core count and installed RAM are not inferred.	source ↗
Apple M2 Ultra (GPU core count and RAM not specified)	Q8_0	56.87 tok/s	llama.cpp b828e18c7 (7948), Metal/BLAS · llama-bench; tg32 @ d16384; 32 generated tokens; context depth 16384; n_ubatch=2048; flash attention=1; one sequence Published token-generation mean; not prompt processing or aggregate batch throughput. Hardware core count and installed RAM are not inferred.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	UD-Q4_K_XL	94.64 tok/s	llama.cpp c841aee (build_number 1 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence; file label r20 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 4 (2026-08-30, measured-local): Current-stack sentinel; 20 repeats; kernel 7.0.0-30, Mesa 26.1.7, desktop performance profile, DPM auto and recorded CPU-only background workload. Not a strict-clean speed headline or a same-condition regression comparison with older kernels. Artifact SHA-256 and exact effective flags are preserved in the raw bundle.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	Q4_K_S	100.99 tok/s	llama.cpp 0eca4d490 (build_number 9851 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence; file label r50 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 11 (2026-06-30, measured-local): Official llama.cpp b9851 Ubuntu Vulkan release binary with explicit -dev Vulkan0; exact Q4_K_S speed-first file; pp512/tg128 r50 crossed 100 t/s. Host kept T3 running; root RustDesk service remained present. Speed-first quant, not balanced default.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	UD-Q4_K_XL	99.55 tok/s	llama.cpp 0eca4d490 (build_number 9851 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence; file label r5 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 20 (2026-06-30, measured-local): Official llama.cpp b9851 Ubuntu Vulkan release binary with explicit -dev Vulkan0; balanced coding quant short r5 control. Useful latest-build improvement signal, but keep separate from the longer speed-first Q4_K_S r50 headline.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	Q4_K_S	98.48 tok/s	llama.cpp 4fc4ec554 (build_number 9859 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence; file label r50 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 22 (2026-07-02, measured-local): Official llama.cpp b9859 Ubuntu Vulkan release binary with explicit -dev Vulkan0 and b9851-matching -fa on -mmp 0 command shape; current-runtime control reproduces the 98-99 t/s class but does not replace the stronger b9851 100.99 t/s speed-first headline.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	UD-Q4_K_XL	97.01 tok/s	llama.cpp 4fc4ec554 (build_number 9859 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence; file label r5 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 23 (2026-07-02, measured-local): Official llama.cpp b9859 Ubuntu Vulkan release binary with explicit -dev Vulkan0 and b9851-matching -fa on -mmp 0 command shape; balanced coding quant current-runtime control.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	Q4_K_S	98.12 tok/s	llama.cpp cb295bf59 (build_number 9888 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence; file label r50 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 25 (2026-07-06, measured-local): Official llama.cpp b9888 Ubuntu Vulkan release binary with explicit -dev Vulkan0 and b9851/b9859-matching -fa on -mmp 0 command shape; current-runtime control reproduces the 98 t/s class but does not replace the stronger b9851 100.99 t/s speed-first headline; p0/tg128 r50 measured 98.59 t/s.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	UD-Q4_K_XL	96.53 tok/s	llama.cpp cb295bf59 (build_number 9888 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence; file label r5 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 26 (2026-07-06, measured-local): Official llama.cpp b9888 Ubuntu Vulkan release binary with explicit -dev Vulkan0 and b9851/b9859-matching -fa on -mmp 0 command shape; balanced coding quant current-runtime control.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	Q4_K_S	94.2 tok/s	llama.cpp ac4cddeb0 (build_number 9592 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence; file label r20 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 29 (2026-06-11, measured-local): Latest upstream Vulkan/RADV control for the exact speed-first Qwen3-Coder file; below the b9179 98.51 t/s strict-clean headline and the b9544 98 t/s refresh.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	UD-Q4_K_XL	92.84 tok/s	llama.cpp 2016bf2 (build_number 1 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=-1; type_k=f16; type_v=f16; n_gpu_layers=-1; threads=16; one sequence; file label r20 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 39 (2026-06-05, measured-local): Latest/int-dot generation-only p0/n128 r20 check; below the older b9049 96-97 t/s balanced row and below the b9179 98.51 t/s Q4_K_S speed-first headline.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	IQ4_XS	90.44 tok/s	llama.cpp 1fd5f4803 (build_number 134 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence; file label r5 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 40 (2026-06-03, measured-local): Current-model negative/control row; IQ4_XS did not beat the older Qwen3-Coder Q4_K_S 98.51 t/s speed-first headline. Generation-only p0/n128 r10 was 90.72 t/s.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	Q4_K_S	98.51 tok/s	llama.cpp b81c2cdd7 (build_number 9179 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence; file label r50 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 47 (2026-05-16, measured-local): Strict-clean r50 confirmation; speed-first lower-quality quant candidate; tuned active and power-profiles-daemon inactive; nonessential GUI/video/media noise paused; not the balanced UD default.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	Q4_K_S	95.55 tok/s	llama.cpp de6f727aa (build_number 127 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence; file label r10 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 48 (2026-06-01, measured-local): Latest upstream de6f727 direct rerun with mmap disabled; useful negative/control row, but it did not beat the b9179 98.51 t/s speed-first headline.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	Q4_K_S	93.85 tok/s	llama.cpp d4c8e2c29 (build_number 109 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence; file label r10 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 49 (2026-05-31, measured-local): Latest upstream b9442 direct rerun; useful negative/control row, but it did not beat the b9179 98.51 t/s speed-first headline.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	Q4_K_S	97.23 tok/s	llama.cpp 6b4e4bd58 (build_number 27 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence; file label r20 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 50 (2026-05-27, measured-local): Latest upstream b9360 direct rerun; better than b9334 direct, but it did not beat the b9179 98.51 t/s speed-first headline.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	Q4_K_S	96.27 tok/s	llama.cpp 192d8ae (build_number 1 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence; file label r20 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 51 (2026-05-26, measured-local): Latest upstream b9334 direct rerun; useful regression check, but it did not beat the b9179 98.51 t/s speed-first headline.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	UD-Q4_K_XL	96.76 tok/s	llama.cpp 2496f9c14 (build_number 9049 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 52 (2026-05-07, measured-local): Max-performance campaign guide-flags r20; confirms the balanced UD path remains 96-97 t/s rather than stable 100 t/s.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	UD-Q4_K_XL	92.6 tok/s	llama.cpp 6b4e4bd58 (build_number 27 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence; file label r20 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 53 (2026-05-27, measured-local): Latest upstream b9360 balanced-UD direct rerun; no new headline because generation was below the b9049/b9010 96-97 t/s rows.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	UD-Q4_K_XL	94.15 tok/s	llama.cpp 192d8ae (build_number 1 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence; file label r20 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 54 (2026-05-26, measured-local): Latest upstream b9334 balanced-UD direct rerun; no new headline because generation was below the b9049/b9010 96-97 t/s rows.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	Q4_K_S	97.22 tok/s	llama.cpp b81c2cdd7 (build_number 9179 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 55 (2026-05-16, measured-local): Initial break-100 sweep on current master before strict host-state fix; speed-first lower-quality quant candidate; superseded by strict-clean 98.51 t/s r50 row.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	UD-Q4_K_XL	96.15 tok/s	llama.cpp 2496f9c14 (build_number 9049 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence; file label r20 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 61 (2026-05-07, measured-local): Clean latest-stack r20 rerun; one lightweight workstation dependency remained running; no material change versus b9010 headline.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	UD-Q4_K_XL	73.69 tok/s	llama.cpp b1c70e2e5 (build_number 8460 as recorded), ROCm/HIP · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence; file label r10 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics. Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 97 (2026-05-03, measured-local): ROCm spot check with HSA override; rocBLASLt Tensile gfx1151 file missing warning	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	Q4_K_S	98.02 tok/s	llama.cpp 98d5e8ba8 (build_number 9544 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence; file label r50 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 99 (2026-06-07, measured-local): Exact SHA-matched Q4_K_S speed-first file redownloaded from Unsloth and rerun on b9544 with explicit -dev Vulkan0; r50 pp512/tg128 did not beat the older b9179 strict-clean 98.51 t/s row; generation-only p0/n128 r20 measured 98.49 t/s.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	UD-Q4_K_XL	97.08 tok/s	llama.cpp 98d5e8ba8 (build_number 9544 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence; file label r5 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 100 (2026-06-07, measured-local): Latest llama.cpp b9544 Vulkan/RADV balanced coding control with explicit -dev Vulkan0; r5; stays in the 96-97 t/s class and remains separate from the Q4_K_S speed-first control.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	UD-Q4_K_XL	87.11 tok/s	llama.cpp b8460, Vulkan (RADV) · llama-bench pp512/tg128 as tabulated (tg_tps column = tg128 generation); exact flags not retained in a raw CSV for this historical row Historical row sourced by the guide to its own BENCHMARKS.md/README.md; no raw CSV retained. Row date 2026-03-21, status historical-local, kernel 6.19.4, driver stack 26.0.2-26.0.6. Author note: Previous coding MoE headline result	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	Q8_0	51.4 tok/s	Ollama 0.20.4, Ollama Vulkan RADV · Ollama generation rate (eval tokens/s) as tabulated; prompt 12 tokens; one request at a time unless notes say otherwise; generated-token count not tabulated Ollama API eval rate from the guide's structured table; warm-run means where the note says so. Row date 2026-03-20, status historical-local, kernel 6.19.4, driver stack 26.0.2. Author note: Ollama path	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	Q8_0	51.3 tok/s	Ollama 0.20.4, Ollama Vulkan RADV · Ollama generation rate (eval tokens/s) as tabulated; prompt 21 tokens; one request at a time unless notes say otherwise; generated-token count not tabulated Ollama API eval rate from the guide's structured table; warm-run means where the note says so. Row date 2026-03-20, status historical-local, kernel 6.19.4, driver stack 26.0.2. Author note: Ollama path	source ↗

$ improve this page

Update model information Add a place to run it Add a test result Propose a benchmark Correct a benchmark score Edit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated 2026-10-11 02:20 UTC · sources and licences · json