Qwen3.5 2B

+ contribute

Qwen3.5 2B is an open-weights language model from Alibaba, released March 2026. Its published size is 2.27B parameters and a 262k-token context window. Licence: apache-2.0.

47.2±0.6#392 of 719 LLMsfrom 109 benchmark results across 13 sources

$ how it compares

The LLMs ranked either side of Qwen3.5 2B, out of 828 with a score, then more from Alibaba.

#	model	score	where	RAM	released
467	MiniCPM3-4B OpenBMB	48±2	local	~4GB	2024-09
468	Bielik 4.5B v3.0 Instruct SpeakLeash	48±2	local	~4.3GB	2025-04
469	Qwen3.5 2B Alibaba	47±1	local	~2.9GB	2026-03
470	Gemma 2 9B IT Google	47±1	local	~7GB	2024-06
471	Command R Plus (08-2024) Cohere	47±1	api $4.38/M +local	~63GB	2024-08
	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▁▁▃▆█▆▅▅▅▄▃▁ #275 of 300 · 9.3% of its score · 10 tests

Beats Magistral Small 1.0 (2506), Bielik 11B v2.3 Instruct, Mixtral 8x22B Instruct v0.1 here, though they rank higher overall. Trails Granite 4.0 H 350M, Qwen3 0.6B, Qwen2.5 1.5B Instruct here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
BFCL V4 (December 2025 archive) +1	43.6% ✎	▄·█·▄▄▄▄·▄·▄ beats 1/8	medium · avg 43%	▇ 1.8%	at its level · several runs
SWE-bench Verified +1	5.0% ✎	▄··▄▄█··▄·▄█ beats 3/8	very hard · avg 15%	▇ 1.7%	sharp · small run · several runs
SWE-bench Pro (split unspecified) +1	0.8% ✎	█▃···▂···▄·▂ beats 4/8	very hard · avg 8.0%	▇ 1.2%	sharp · too hard for it · small run · several runs
τ²-bench aggregate lab	48.8% ✎	▁▁▁·▂▃▄▃▃▅█▂ beats 5/37	medium · avg 44%	▇ 1.1%	at its level · small run
MCP-Atlas (public set)	14.8% ✎	▃·▁▂▁·▁▃█▆▆▂ beats 1/37	hard · avg 21%	▇ 1.0%	typical
SWE-bench Multilingual	5.0% ✎	▃··▁▃▂▅▇█▃▃▁ beats 0/60	hard · avg 18%	▇ 0.8%	typical
BrowseComp-ZH	18.2% ✎	▄▄·▄▄▄·█···█ beats 5/8	very hard · avg 13%	▇ 0.6%	typical
τ³-bench Banking	2.1% ✎	▂▃·▅█▂▂▃▂▅▂▃ beats 1/20	very hard · avg 3.7%	▇ 0.4%	blunt · too hard for it

2 more tests

test	result	vs the others in its run	difficulty	counts	why
τ³-bench	4.1% ✎	▃·▃·█·▃··▃·▃ beats 0/7	very hard · avg 14%	▇ 0.4%	blunt
BrowseComp	3.1% ✎	▅···█·▃···▃▃ beats 0/7	very hard · avg 10%	▇ 0.3%	noisy

coding▁▁▁▃▆██▇▆▄▃▁ #439 of 458 · 8.3% of its score · 7 tests

Trails Falcon3 1B Instruct, Granite 3.0 1B-A400M Instruct, Gemma 2B (base) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
LiveCodeBench v6 (window unspecified)	20.2% ✎	▃····█·▃▅▃·▃ beats 0/8	medium · avg 39%	▇ 2.0%	at its level
SWE-bench Verified +1	5.0% ✎	▄··▄▄█··▄·▄█ beats 3/8	very hard · avg 15%	▇ 1.7%	sharp · small run · several runs
LiveCodeBench Pro 2025Q2 (Easy)	10.3% ✎	▄···█▄··▄█▄▄ beats 0/8	hard · avg 33%	▇ 1.5%	at its level
SWE-bench Pro (split unspecified) +1	0.8% ✎	█▃···▂···▄·▂ beats 4/8	very hard · avg 8.0%	▇ 1.2%	sharp · too hard for it · small run · several runs
SWE-bench Multilingual	5.0% ✎	▃··▁▃▂▅▇█▃▃▁ beats 0/60	hard · avg 18%	▇ 0.8%	typical
SciCode (level unspecified)	6.0% ✎	█·····██████ beats 0/6	hard · avg 17%	▇ 0.6%	typical
LiveCodeBench Pro 2025Q2 (Medium)	0.0% ✎	█···▂··▄▂··▂ beats 2/8	very hard · avg 2.0%	▇ 0.5%	sharp · too hard for it · noisy

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #247 of 393 · 1.0% of its score · 1 test

Beats Mistral Small 4 (119B-A6.5B), Ministral 3 14B Reasoning 2512, Qwen2.5 Coder 32B Instruct here, though they rank higher overall. Trails Flan-T5 XXL, Granite 4.0 H 1B, Granite 4.0 1B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
IFEval (aggregation unspecified) +2	77.5% ✎	▄▄····▄··███ beats 2/8	easy · avg 79%	▇ 1.0%	several runs

math▁▂▃▅▆▅▄█▄▁▁▁ #286 of 398 · 26% of its score · 13 tests

Beats Llama 3.1 Nemotron 70B Instruct, GPT-4 (0613), Athene-70B here, though they rank higher overall. Trails DeepSeek R1 Distill Qwen 1.5B, Phi-2 (2.7B), StarCoder2 15B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
HMMT February 2026 +1	20.5% ✎	▅··▃▃·▃··▃·█ beats 1/8	hard · avg 32%	▇▇▇ 7.1%	sharp · at its level · several runs
AIME 2026 +1	29.0% ✎	▄·▄█·▄▄···█▄ beats 0/8	medium · avg 53%	▇▇ 6.0%	sharp · at its level · several runs
AIME 2025	29.6% ✎	█▄▄▄▄····█·▄ beats 0/8	medium · avg 42%	▇ 2.8%	sharp · at its level
MATH-500	85.8% ✎	██▄▄▄··▄···▄ beats 1/8	very easy · avg 89%	▇ 2.5%	typical
MathArena ArXivMath April 2026 (40 problems)	1.9% ✎	▃·····▅██▄·▁ beats 0/20	very hard · avg 7.2%	▇ 1.4%	sharp · too hard for it
HMMT February 2025 lab	22.9% ✎	▁▃▁▃▁▅▅▃██▁▄ beats 1/31	medium · avg 36%	▇ 1.4%	at its level · small run
MathArena ArXivMath June 2026 (48 problems)	4.2% ✎	▃···▄·██▆▆▄▂ beats 0/19	very hard · avg 4.6%	▇ 1.0%	sharp · too hard for it
MathVista (testmini) lab	76.7% ✎	▁··▁▄▄▇█▅▆█▄ beats 30/47	easy · avg 73%	▇ 0.9%	at its level · small run

5 more tests

test	result	vs the others in its run	difficulty	counts	why
HMMT November 2025 lab	19.6% ✎	▁▃····▁▅▃█▇▁ beats 1/23	medium · avg 41%	▇ 0.9%	at its level · small run
PolyMATH lab	26.1% ✎	▁···▃▁·▇█▄▄▄ beats 2/23	hard · avg 29%	▇ 0.9%	at its level · small run
MathArena ArXivMath May 2026 (40 problems)	10.6% ✎	▃··▁▄█▄▁·▃·▁ beats 0/17	very hard · avg 10%	▇ 0.7%	typical
DynaMath lab	73.6% ✎	▂·····▃··█▄▆ beats 1/10	easy · avg 70%	▇ 0.7%	at its level · small run
MathArena BrokenArXiv June 2026 (54 problems)	2.8% ✎	▄··█▂▂▂▆▄··▂ beats 0/14	very hard · avg 1.1%	▇ 0.1%	sharp · too hard for it · noisy

vision & documents▁▁▃▂▁▃▄▂▆▇█▄ #114 of 147 · 14% of its score · 20 tests

Beats DeepSeek V4 Flash Vision (Exp), GPT-4 Turbo (2024-04-09), Qwen2-VL 72B Instruct here, though they rank higher overall.

test	result	vs the others in its run	difficulty	counts	why
MMStar +1	61.4% ✎	▄▄▄█·······▄ beats 5/5	medium · avg 62%	▇ 1.9%	at its level · several runs
CharXiv descriptive questions	76.1% ✎	█·▄·······▄█ beats 4/5	easy · avg 78%	▇ 1.4%	at its level
ChartArena chart parsing (English, mAP_high average)	16.2% ✎	▁·▃·▃▃▁▇▃█·▄ beats 1/22	hard · avg 27%	▇ 1.2%	blunt
CharXiv (reasoning questions) lab	58.8% ✎	▂▅▃▅▅▆▆▅▆█▆▂ beats 43/100	medium · avg 55%	▇ 1.1%	at its level · small run
MMBench-EN dev v1.1 +1	76.0% ✎	█·····▃··▃·▃ beats 5/5	easy · avg 79%	▇ 1.0%	blunt · several runs
MMMU (val) lab	64.2% ✎	▁····▁▁▄█▇▆▆ beats 27/52	medium · avg 63%	▇ 1.0%	at its level · small run
MathVista (testmini) lab	76.7% ✎	▁··▁▄▄▇█▅▆█▄ beats 30/47	easy · avg 73%	▇ 0.9%	at its level · small run
MMMU Pro lab	50.3% ✎	▂▃▁▃▆▇▃▇▇██▅ beats 12/52	medium · avg 49%	▇ 0.9%	at its level · small run

12 more tests

test	result	vs the others in its run	difficulty	counts	why
HallusionBench +1	65.5% ✎	▄▄·▄·····█·▄ beats 4/5	medium · avg 65%	▇ 0.8%	blunt · at its level · several runs
CountBench +1	80.5% ✎	▄··▄··█··▄·▄ beats 4/5	very easy · avg 90%	▇ 0.7%	small run · several runs
OmniDocBench 1.5 overall (higher is better) lab	79.8% ✎	▂··▂▂▃··▂▂█▅ beats 3/13	easy · avg 75%	▇ 0.7%	at its level · small run
SimpleVQA lab	38.5% ✎	▃·▃·▃·▃·█▅▃▃ beats 1/10	medium · avg 40%	▇ 0.6%	at its level · small run
PMC-VQA lab	48.8% ✎	▃···▃··▃▃·▃█ beats 1/7	medium · avg 48%	▇ 0.6%	at its level · small run
RealWorldQA +1	69.3% ✎	▄···█··▄▄··▄ beats 5/5	easy · avg 69%	▇ 0.5%	blunt · at its level · several runs
OCRBench +1	86.1% ✎	█··███·█···█ beats 5/5	easy · avg 81%	▇ 0.4%	blunt · several runs
AI2D test +1	75.2% ✎	█···▄▄···▄·▄ beats 3/5	easy · avg 82%	▇ 0.2%	blunt · several runs
CC-OCR lab	72.9% ✎	▁·▂··▃▂▃▃▂▆█ beats 4/26	easy · avg 73%	▇ 0.2%	blunt · at its level · small run
SpaceCast-Bench predictive spatial reasoning (overall, 16 task types)	31.3% ✎	▃·▅▃▅▃··▅··█ beats 0/11	very hard · avg 32%	▇ 0.1%	blunt · too hard for it · noisy
ChartQA	77.5% ✎	▄······▄··██ beats 2/5	easy · avg 80%	▇ 0.1%	blunt · noisy
PureDocBench document parsing (Avg3 over Clean/Digital/Real)	62.5% ✎	▂·▁·▃▂▅█·▁▂▂ beats 4/22	medium · avg 62%	▇ 0.1%	blunt · at its level

web search█·█▄██·▄▄·▄▄ #8 of 13 * · 0.9% of its score · 2 tests

Beats Nemotron 3 Nano 4B, Gemma 4 E4B, LFM2.5 2.6B here, though they rank higher overall.

test	result	vs the others in its run	difficulty	counts	why
BrowseComp-ZH	18.2% ✎	▄▄·▄▄▄·█···█ beats 5/8	very hard · avg 13%	▇ 0.6%	typical
BrowseComp	3.1% ✎	▅···█·▃···▃▃ beats 0/7	very hard · avg 10%	▇ 0.3%	noisy

other tests39 not in a task · 45% of its score

test	result	vs the others in its run	difficulty	counts	why
MMLU-Pro +3	64.3% ✎	▂··▂█·▂▂···▂ beats 2/8	medium · avg 59%	▇▇ 6.2%	at its level · several runs
GPQA (subset unspecified) +1	44.6% ✎	█···▃···▅·▃▃ beats 2/7	medium · avg 51%	▇ 3.6%	sharp · at its level · small run · several runs
MMLU-Redux +1	80.0% ✎	▂···█··▂▂··▂ beats 4/8	easy · avg 77%	▇ 3.0%	at its level · several runs
EuroEval Life in the UK (English knowledge MC), validation split, 5-shot, accuracy	70.6% ✎	▂·▂▃·▂·▃▃·█▃ beats 2/14	easy · avg 80%	▇ 2.8%	at its level
SuperGPQA +1	38.6% ✎	▃·▃·▃▅█····▃ beats 4/8	medium · avg 39%	▇ 1.9%	at its level · small run · several runs
IFBench (aggregation unspecified) +1	46.0% ✎	▅····▃▃·█·▃▃ beats 2/8	medium · avg 40%	▇ 1.7%	at its level · small run · several runs
INCLUDE lab	55.4% ✎	▂····▂▂▄▄▄█▄ beats 10/36	medium · avg 59%	▇ 1.5%	at its level · small run
MMLU	54.3% ✎	▄·····▄▄··▄█ beats 2/5	medium · avg 70%	▇ 1.4%	at its level

31 more tests

test	result	vs the others in its run	difficulty	counts	why
C-Eval lab	73.2% ✎	▁···▁▁▂▁▂▃▃█ beats 11/33	easy · avg 75%	▇ 1.4%	at its level · small run
LongBench v2 +1	24.9% ✎	▃······▅▅▃·█ beats 0/8	very hard · avg 35%	▇ 1.3%	noisy · small run · several runs
EuroEval HellaSwag (English), validation split, 5-shot generative MC, accuracy	56.2% ✎	▂··▂▃·▂▆▆█·▄ beats 2/15	medium · avg 63%	▇ 1.2%	at its level
MMMLU lab	63.1% ✎	▂▁▁▂▃▂▄▄▅██▃ beats 16/46	medium · avg 65%	▇ 1.2%	at its level · small run
MedXPertQA MM lab	26.9% ✎	▂···▂·▃··▆▂█ beats 2/13	very hard · avg 28%	▇ 1.2%	sharp · too hard for it · small run
Multi-IF (aggregation unspecified) +1	57.1% ✎	▄▄···▄·▄·▄██ beats 2/8	medium · avg 58%	▇ 1.1%	blunt · at its level · several runs
Humanity's Last Exam	2.1% ✎	▄▄·▄··▄·▄·▄█ beats 0/7	very hard · avg 3.8%	▇ 1.1%	too hard for it
Global PIQA lab	69.3% ✎	▃▃··▃··▃▃▅█▅ beats 2/11	medium · avg 70%	▇ 1.1%	at its level · small run
MultiChallenge lab	33.7% ✎	▁·▁·▃▃▆▅▃▁█▅ beats 2/29	hard · avg 34%	▇ 1.0%	at its level · small run
KINA (Knowledge Index of Noah's Ark) accuracy, 899 ten-option items	20.5% ✎	▁▁▂▃▃▄█▆▄▅▂▃ beats 4/39	very hard · avg 22%	▇ 1.0%	blunt
MAXIFE lab	60.6% ✎	▂·▂·▃···▆·▄█ beats 2/11	medium · avg 63%	▇ 0.9%	at its level · small run
MMLU-ProX lab	52.3% ✎	▁·▁▃▁▂▁▃▂█▅▄ beats 2/35	medium · avg 57%	▇ 0.9%	at its level · small run
OJBench	2.6% ✎	▃····▃▅··█▃▃ beats 0/8	very hard · avg 12%	▇ 0.8%	typical
VideoMMMU lab	62.1% ✎	▂···▃···▂▂█▆ beats 1/10	medium · avg 63%	▇ 0.7%	at its level · small run
LVBench lab	57.1% ✎	▇·█▅▅▄▄··▃▁▁ beats 9/28	medium · avg 55%	▇ 0.7%	at its level · small run
RefSpatial-Bench lab	32.9% ✎	▃·▃····▃▃▅▅█ beats 1/10	medium · avg 39%	▇ 0.6%	at its level · small run
MMVU lab	48.6% ✎	▃···▃···▃▃▅█ beats 1/8	medium · avg 51%	▇ 0.6%	at its level · small run
ERQA lab	43.8% ✎	▃··▃··█·▃█▅▅ beats 1/12	medium · avg 43%	▇ 0.6%	at its level · small run
ZEROBench subquestions lab	17.1% ✎	▄·▄···▄·▄██▄ beats 1/8	hard · avg 19%	▇ 0.6%	small run
Video-MME (with subtitles) lab	75.6% ✎	▁▁▁▁▁·▃▃▁▁▁█ beats 5/16	easy · avg 75%	▇ 0.6%	at its level · small run
VLMs Are Blind lab	75.8% ✎	▂·▃·····▃··█ beats 1/8	easy · avg 83%	▇ 0.6%	small run
MMLongBench-Doc lab	45.4% ✎	▃····▃··▃▃▅█ beats 1/8	medium · avg 44%	▇ 0.6%	at its level · small run
SLAKE lab	74.4% ✎	▄·····▄▄·██▄ beats 1/7	easy · avg 72%	▇ 0.6%	at its level · small run
NOVA-63 lab	46.4% ✎	▃▃·▃···▃█▃▃█ beats 2/11	medium · avg 46%	▇ 0.6%	blunt · at its level · small run
MLVU lab	76.2% ✎	▂···▃···▂▄▄█ beats 1/10	easy · avg 75%	▇ 0.6%	at its level · small run
EmbSpatial-Bench lab	77.9% ✎	▁·····▃·▁·▃█ beats 1/10	easy · avg 75%	▇ 0.5%	at its level · small run
Video-MME (without subtitles) lab	69.0% ✎	▁▃▁▃▁·█▄▄▅·▇ beats 8/27	medium · avg 68%	▇ 0.5%	blunt · at its level · small run
MVBench lab	64.9% ✎	▁··▃▄▁▃▄▄▄█▄ beats 5/26	medium · avg 66%	▇ 0.2%	blunt · at its level · small run
LIBRA Mini (Russian long context, May 2026 version)	37.4% ✎	▄···▄█·██··▄ beats 2/8	medium · avg 38%	▇ 0.1%	blunt · at its level
WMT24++ lab	45.8% ✎	▁·▁▃·▁·▁▂▆█▃ beats 2/27	easy · avg 66%	▇ 0.1%	blunt · at its level · small run
Chess Puzzles (Epoch)	0.0% ✎	█▁▁▁▁▂▁▄▂▂▁▁ beats 36/135	very hard · avg 5.9%	▇ 0.1%	sharp · too hard for it · noisy

$ where to run it

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 Qwen3.5 2B · needs ~1.3 GB · from huggingface.co/lmstudio-community
chat in the terminal
llama cli -hf lmstudio-community/Qwen3.5-2B-GGUF:Q4_K_M
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf lmstudio-community/Qwen3.5-2B-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 Qwen3.5 2B · needs ~1.3 GB · from huggingface.co/lmstudio-community
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/lmstudio-community/Qwen3.5-2B-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 Qwen3.5 2B · needs ~1.86 GB · from ollama.com
chat in the terminal
ollama run qwen3.5:2b-q4_K_M

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 Qwen3.5 2B · needs ~1.7 GB, Mac only · from huggingface.co/lmstudio-community
chat in the terminal
mlx_lm.chat --model lmstudio-community/Qwen3.5-2B-MLX-4bit
Or start an OpenAI-compatible server at http://localhost:8080/v1
mlx_lm.server --model lmstudio-community/Qwen3.5-2B-MLX-4bit

Official weights and documentation: huggingface.co/Qwen ↗

$ how the score was worked out · Qwen/Qwen3.5-2B model card 39%, openbmb/MiniCPM5-2B model card 26%, XHToken/Spark-X2.5-4B model c… 15% and 10 more

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/Qwen3.5-2B model card · lab-reported	52	▇▇▇▇▇▇▇▇▇▇▇▇ 39%	apache-2.0
openbmb/MiniCPM5-2B model card	20	▇▇▇▇▇▇▇▇ 26%	Apache-2.0
XHToken/Spark-X2.5-4B model card	11	▇▇▇▇▇ 15%	Apache-2.0
CohereLabs/North-Micro-Vision-Instruct model card (benchmark table)	13	▇▇▇ 9%	Apache-2.0
EuroEval leaderboards (results collection), English generative leaderboard	2	▇ 4%	MIT
MathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets	4	▇ 3%	CC BY-SA 4.0
Qwen/Qwen3.5-0.8B model card · lab-reported	1	▇ 2%	apache-2.0
ChartArena: Benchmarking Chart Parsing across Languages, Scenarios, and Formats (arXiv:2606.01348v3)	1	▇ 1%	CC BY 4.0
Knowledge Index of Noah's Ark (arXiv:2606.05104v2)	1	▇ 1%	CC BY 4.0
ai-forever LIBRA leaderboard (Russian long-context benchmark)	1	▇ 0%	MIT
SpaceCast-Bench: Evaluating Predictive Spatial Reasoning in Vision-Language Models (arXiv:2610.12402v1)	1	▇ 0%	CC BY 4.0
Epoch AI — internal evaluations	1	▇ 0%	CC BY 4.0
How Far Is Document Parsing from Solved? PureDocBench: A Source-Traceable Benchmark across Clean, Degraded, and Real-World Settings (arXiv:2605.07492v2)	1	▇ 0%	CC BY 4.0
prior · pull toward the middle		▇ 0%
weighted average		= 47.2

$ published speed measurements · 4

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

Apple M5 Pro (48GB; CPU/GPU core counts not stated)	Q4 GGUF (4-bit; exact llama.cpp type not stated)	147.1 tok/s	llama.cpp b9960, Metal backend, default settings · tg128: token-generation throughput over 128 generated tokens; average of five repetitions; macOS (Darwin 25.4) Baseline runtime column from the BaseRT M5 paper; BaseRT's own column is not recorded.	source ↗
Apple M5 Pro (48GB; CPU/GPU core counts not stated)	MLX 4-bit (group size not stated)	208.7 tok/s	mlx-lm 0.31.3 (MLX 0.32.0), default generation settings · tg128: token-generation throughput over 128 generated tokens; average of five repetitions; macOS (Darwin 25.4) Baseline runtime column from the BaseRT M5 paper; BaseRT's own column is not recorded.	source ↗
Apple M5 Pro (48GB; CPU/GPU core counts not stated)	Q8 GGUF (8-bit; exact llama.cpp type not stated)	110.6 tok/s	llama.cpp b9960, Metal backend, default settings · tg128: token-generation throughput over 128 generated tokens; average of five repetitions; macOS (Darwin 25.4) Baseline runtime column from the BaseRT M5 paper; BaseRT's own column is not recorded.	source ↗
Apple M5 Pro (48GB; CPU/GPU core counts not stated)	MLX 8-bit (group size not stated)	125.9 tok/s	mlx-lm 0.31.3 (MLX 0.32.0), default generation settings · tg128: token-generation throughput over 128 generated tokens; average of five repetitions; macOS (Darwin 25.4) Baseline runtime column from the BaseRT M5 paper; BaseRT's own column is not recorded.	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