FitMyLLM
ARM· Valhall

ARM Mali-G68 MC2

Running LLMs on the ARM Mali-G68 MC2 — the long read: which models fit at which quantisation, and the settings worth changing. · Or what a budget buys

VRAM
3 GB
Budget
Bandwidth
14
GB/s
Models Q4
87/498
17%
7B Q4 speed
~2
Slow
Model coverage at Q417% of all
Estimated speed· By model size at Q4

Average speeds at Q4 quantization. Actual performance varies by model architecture and context length.

3B
~4
tok/s
7B
—
3.9GB needed
14B
—
7.9GB needed
32B
—
18.0GB needed
70B
—
39.4GB needed
Measured rig reports

We rent the machine and time every model on it: decode, VRAM peak, concurrency, watts and cost per million tokens. Yours may already be one of them — and two are free to read in full.

See the reports →
Specifications
VRAM
3 GB
Bandwidth
14 GB/s
FP16 Compute
0.1 TFLOPS
TDP
2W
Memory
—
Architecture
Valhall
13
Fast models · >30 tok/s
Real-time chat speed
37
Usable · >10 tok/s
Comfortable for all tasks
87
Total compatible
Fit in VRAM at Q4
Don’t want to buy?

Test ARM Mali-G68 MC2 (or anything bigger) without committing. Pay by the second, cancel anytime.

Spin up in ~60s. Pay by the second. Cancel anytime.

Some links are affiliate links — we may earn a small commission at no extra cost to you. This helps keep FitMyLLM free and independent.

Compatible models· 87
S
Falcon-H1R Tiny 90M0.09B
Falcon·256K ctx· chat· reasoning
138
tok/s · 18% VRAM
›
S
SmolLM2 135M0.135B
Smollm·2K ctx· chat
92
tok/s · 19% VRAM
›
S
nomic-embed-text-v1.5 100M0.14B
Embedding·8K ctx· chat
89
tok/s · 19% VRAM
›
S
GPT-2 124M0.14B
Gpt2·1K ctx· chat
89
tok/s · 19% VRAM
›
A
SmolVLM 256M0.256B
Smollm·8K ctx· chat· vision
49
tok/s · 21% VRAM
›
A
Gemma 3 270M0.27B
Gemma·32K ctx· chat
46
tok/s · 22% VRAM
›
A
Snowflake Arctic Embed M v2.00.305B
Embedding·8K ctx· embedding· multilingual
41
tok/s · 22% VRAM
›
B
bge-large-en-v1.5 335M0.335B
Embedding·1K ctx· chat
37
tok/s · 23% VRAM
›
B
mxbai-embed-large-v10.335B
Embedding·1K ctx· embedding
37
tok/s · 23% VRAM
›
B
Snowflake Arctic Embed L0.335B
Embedding·1K ctx· embedding
37
tok/s · 23% VRAM
›
B
LFM2 350M0.35B
Lfm·125K ctx· chat· tool_use
36
tok/s · 23% VRAM
›
B
SmolLM2 360M0.36B
Smollm·8K ctx· chat
35
tok/s · 24% VRAM
›
B
GPT-2 Medium 345M0.38B
Gpt2·1K ctx· chat
33
tok/s · 24% VRAM
›
B
Qwen 2.5 0.5B0.5B
Qwen·32K ctx· chat
25
tok/s · 26% VRAM
›
B
SmolVLM 500M0.5B
Smollm·8K ctx· chat· vision
25
tok/s · 26% VRAM
›
C
Falcon-H1 0.5B0.52B
Falcon·128K ctx· chat
24
tok/s · 27% VRAM
›
C
BGE-M30.568B
Embedding·8K ctx· embedding
22
tok/s · 28% VRAM
›
C
Snowflake Arctic Embed L v2.00.568B
Embedding·8K ctx· embedding· multilingual
22
tok/s · 28% VRAM
›
C
Qwen3-Embedding 0.6B0.6B
Embedding·32K ctx· embedding· multilingual
21
tok/s · 29% VRAM
›
C
Falcon-H1R Tiny 0.6B0.6B
Falcon·32K ctx· chat· reasoning
21
tok/s · 29% VRAM
›
C
Falcon Perception 0.6B0.6B
Falcon·4K ctx· vision
21
tok/s · 29% VRAM
›
C
Qwen 1.5 0.5B0.62B
Qwen·32K ctx· chat
20
tok/s · 29% VRAM
›
C
Qwen3 0.6B0.75B
Qwen·32K ctx· chat· reasoning
17
tok/s · 32% VRAM
›
C
LFM2 700M0.74B
Lfm·125K ctx· chat· tool_use
17
tok/s · 31% VRAM
›
C
GPT-2 Large 774M0.81B
Gpt2·1K ctx· chat
15
tok/s · 33% VRAM
›
D
Qwen 3.5 0.8B0.87B
Qwen·256K ctx· chat· coding· multilingual
14
tok/s · 34% VRAM
›
D
InternVL3 1B0.94B
Other·32K ctx· chat· vision
13
tok/s · 35% VRAM
›
D
MiniCPM5 1B1.08B
Minicpm·128K ctx· chat· reasoning· multilingual
12
tok/s · 38% VRAM
›
D
K2 Horizon 0.9B1.08B
Other·128K ctx· chat· coding· reasoning
12
tok/s · 38% VRAM
›
D
TinyLlama 1.1B1.1B
Llama·2K ctx· chat
11
tok/s · 39% VRAM
›
D
EXAONE-4.0-1.2B1.3B
Exaone·64K ctx· chat
10
tok/s · 43% VRAM
›
D
OPT 1.3B1.3B
Opt·2K ctx· chat
10
tok/s · 43% VRAM
›
D
MiniCPM-V 4.61.3B
Other·256K ctx· chat· vision
10
tok/s · 43% VRAM
›
D
LFM2.5-1.2B-Thinking1.2B
Lfm·122K ctx· chat· reasoning· tool_use
10
tok/s · 41% VRAM
›
D
Llama-3.2-1B1.2B
Llama·4K ctx· chat
10
tok/s · 41% VRAM
›
D
LFM2 1.2B1.2B
Lfm·125K ctx· chat· tool_use· multilingual
10
tok/s · 41% VRAM
›
D
Zamba2 1.2B1.2B
Other·4K ctx· chat
10
tok/s · 41% VRAM
›
D
Phi-1 1.3B1.42B
Phi·2K ctx· coding
9
tok/s · 45% VRAM
›
D
Phi-1.5 1.3B1.42B
Phi·2K ctx· chat· coding
9
tok/s · 45% VRAM
›
D
DeepSeek Coder 1.3B1.35B
Deepseek·16K ctx· coding
9
tok/s · 44% VRAM
›
Next step

Get personalized recommendations.

See ranked models with benchmark scores, run commands, and precise speed estimates for your ARM Mali-G68 MC2.

What this card is worth

ARM Mali-G68 MC2 holds 87 of the models in our catalogue and is, in practice, a Q4_K_M card — the largest it takes is Falcon3-3B at Q4_K_M.

Tokens/sec per $100
—
8B at Q4_K_M, so cards compare like for like
VRAM per $100
—
what memory costs on this card
The biggest it takes
Falcon3-3B3.23B · Q4_K_M2.6 GB5 tok/sest
granite-4.0-h-micro 3.2B3.2B · Q5_K_M2.6 GB5 tok/sest
Llama-3.2-3B3.2B · Q4_K_M2.7 GB5 tok/sest
Falcon-H1 3B3.15B · Q5_K_M2.7 GB5 tok/sest
Qwen 2.5 3B3.1B · Q5_K_M2.6 GB5 tok/sest
SmolLM3-3B3.1B · Q4_K_M2.5 GB6 tok/sest
Ministral 3B3B · Q4_K_M2.6 GB6 tok/sest
StarCoder2 3B3B · Q5_K_M2.5 GB5 tok/sest
Against cards of similar memory
M2 (8GB)5 GB$5992.8 tok/s per $100
M3 (8GB)5 GB$5992.8 tok/s per $100
M1 (8GB)5 GB$4992.2 tok/s per $100
RTX 2060 6GB6 GB$15026.7 tok/s per $100
ARM Mali-G68 MC23 GB——

Neighbours in memory rather than in price: memory decides whether a card can do the job at all, so two cards of the same size at different prices is the comparison you are making. Ordered by memory, then bandwidth — not by the value column, which is worked out from bandwidth and price alone and therefore rewards a cheap card whatever its software stack does to that bandwidth in practice. Read it as one input, not as a ranking.

Device under test

ARM Mali-G68 MC2 — 3 GB VRAM.

ARM Mali-G68 MC2 spec
Brand
Apple
VRAM
3 GB
Bandwidth
14 GB/s
FP16 Compute
0.1 TFLOPS
FP32 Compute
0.05 TFLOPS
TDP
2 W
Architecture
Valhall
AI capability
87/ 498 models @ Q4

With 3 GB VRAM and 14 GB/s bandwidth, this GPU handles models up to 3.1B parameters.

Speed ≈ bandwidth / model_size × efficiency. A 7B model at Q4 runs at ~2 tok/s.

Top models for ARM Mali-G68 MC2
87 fit · showing 20
ModelSizeVRAM Q4tok/sAvg
Qwen 2.5 3B3.1B2.4 GB437.2
SmolLM3-3B3.1B2.4 GB435.1
Ministral 3B3B2.3 GB423.7
StarCoder2 3B3B2.3 GB49.5
Granite 4.1 3B3B2.3 GB417.5
xLAM-2 3B Function-Calling3B2.3 GB4—
Jamba 2 3B3B2.3 GB4—
Dolly v2 3B2.8B2.2 GB45.6
StableLM Zephyr 3B2.79B2.2 GB414.9
Zephyr 3B2.79B2.2 GB414.4
OPT 2.7B2.7B2.1 GB528.0
Phi-2 2.7B2.7B2.1 GB524.1
Zamba2 2.7B2.7B2.1 GB548.0
LFM2.5 2.6B2.7B2.1 GB528.3
Granite 3.0 2B2.63B2.1 GB535.8
gemma-2-2b2.6B2.1 GB522.9
LFM2 2.6B2.6B2.1 GB516.3
Granite 3.1 2B2.53B2.0 GB537.8
Granite 3.3 2B2.53B2.0 GB520.5
MiniCPM5 2B2.52B2.0 GB549.2