FitMyLLM
▸ ARM· IMMORTALIS GEN3

Google Tensor G3 GPU

Running LLMs on the Google Tensor G3 GPU — the long read: which models fit at which quantisation, and the settings worth changing. · Or what a budget buys

VRAM
12 GB
ENTRY-LEVEL
BANDWIDTH
34
GB/S
MODELS Q4
248/449
55%
7B Q4 SPEED
~4
SLOW
▸ MODEL COVERAGE @ Q455% OF ALL
▸ ESTIMATED SPEED· BY MODEL SIZE @ Q4

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

3B
~10
TOK/S
7B
~4
TOK/S
14B
~2
TOK/S
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
12 GB
BANDWIDTH
34 GB/s
FP16 COMPUTE
1.5 TFLOPS
TDP
7W
MEMORY
Shared
ARCHITECTURE
Immortalis Gen3
27
FAST MODELS · >30 TOK/S
Real-time chat speed
85
USABLE · >10 TOK/S
Comfortable for all tasks
248
TOTAL COMPATIBLE
Fit in VRAM at Q4
▸ DON’T WANT TO BUY?

Test Google Tensor G3 GPU (or anything bigger) without committing. Pay by the second, cancel anytime.

Rent on Vast.ai →Or RunPod →

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· 248
S
Falcon-H1R Tiny 90M0.09B
FALCON·256K CTX· CHAT· REASONING
336
TOK/S · 5% VRAM
›
S
SmolLM2 135M0.135B
SMOLLM·2K CTX· CHAT
224
TOK/S · 5% VRAM
›
S
nomic-embed-text-v1.5 100M0.14B
EMBEDDING·8K CTX· CHAT
216
TOK/S · 5% VRAM
›
S
GPT-2 124M0.14B
GPT2·1K CTX· CHAT
216
TOK/S · 5% VRAM
›
S
SmolVLM 256M0.256B
SMOLLM·8K CTX· CHAT· VISION
118
TOK/S · 5% VRAM
›
S
Gemma 3 270M0.27B
GEMMA·32K CTX· CHAT
112
TOK/S · 5% VRAM
›
S
Snowflake Arctic Embed M v2.00.305B
EMBEDDING·8K CTX· EMBEDDING· MULTILINGUAL
99
TOK/S · 6% VRAM
›
S
bge-large-en-v1.5 335M0.335B
EMBEDDING·1K CTX· CHAT
90
TOK/S · 6% VRAM
›
S
mxbai-embed-large-v10.335B
EMBEDDING·1K CTX· EMBEDDING
90
TOK/S · 6% VRAM
›
S
Snowflake Arctic Embed L0.335B
EMBEDDING·1K CTX· EMBEDDING
90
TOK/S · 6% VRAM
›
S
LFM2 350M0.35B
LFM·125K CTX· CHAT· TOOL_USE
86
TOK/S · 6% VRAM
›
S
SmolLM2 360M0.36B
SMOLLM·8K CTX· CHAT
84
TOK/S · 6% VRAM
›
S
GPT-2 Medium 345M0.38B
GPT2·1K CTX· CHAT
80
TOK/S · 6% VRAM
›
S
Qwen 2.5 0.5B0.5B
QWEN·32K CTX· CHAT
60
TOK/S · 7% VRAM
›
S
SmolVLM 500M0.5B
SMOLLM·8K CTX· CHAT· VISION
60
TOK/S · 7% VRAM
›
A
Falcon-H1 0.5B0.52B
FALCON·128K CTX· CHAT
58
TOK/S · 7% VRAM
›
A
BGE-M30.568B
EMBEDDING·8K CTX· EMBEDDING
53
TOK/S · 7% VRAM
›
A
Snowflake Arctic Embed L v2.00.568B
EMBEDDING·8K CTX· EMBEDDING· MULTILINGUAL
53
TOK/S · 7% VRAM
›
A
Qwen3-Embedding 0.6B0.6B
EMBEDDING·32K CTX· EMBEDDING· MULTILINGUAL
50
TOK/S · 7% VRAM
›
A
Falcon-H1R Tiny 0.6B0.6B
FALCON·32K CTX· CHAT· REASONING
50
TOK/S · 7% VRAM
›
A
Falcon Perception 0.6B0.6B
FALCON·4K CTX· VISION
50
TOK/S · 7% VRAM
›
A
Qwen 1.5 0.5B0.62B
QWEN·32K CTX· CHAT
49
TOK/S · 7% VRAM
›
A
LFM2 700M0.74B
LFM·125K CTX· CHAT· TOOL_USE
41
TOK/S · 8% VRAM
›
A
Qwen3 0.6B0.75B
QWEN·32K CTX· CHAT· REASONING
40
TOK/S · 8% VRAM
›
B
GPT-2 Large 774M0.81B
GPT2·1K CTX· CHAT
37
TOK/S · 8% VRAM
›
B
Qwen 3.5 0.8B0.87B
QWEN·256K CTX· CHAT· CODING· MULTILINGUAL
35
TOK/S · 9% VRAM
›
B
InternVL3 1B0.94B
OTHER·32K CTX· CHAT· VISION
32
TOK/S · 9% VRAM
›
B
MiniCPM5 1B1.08B
MINICPM·128K CTX· CHAT· REASONING· MULTILINGUAL
28
TOK/S · 10% VRAM
›
B
TinyLlama 1.1B1.1B
LLAMA·2K CTX· CHAT
27
TOK/S · 10% VRAM
›
B
LFM2.5-1.2B-Thinking1.2B
LFM·122K CTX· CHAT· REASONING· TOOL_USE
25
TOK/S · 10% VRAM
›
B
Llama-3.2-1B1.2B
LLAMA·4K CTX· CHAT
25
TOK/S · 10% VRAM
›
B
LFM2 1.2B1.2B
LFM·125K CTX· CHAT· TOOL_USE· MULTILINGUAL
25
TOK/S · 10% VRAM
›
B
Zamba2 1.2B1.2B
OTHER·4K CTX· CHAT
25
TOK/S · 10% VRAM
›
C
EXAONE-4.0-1.2B1.3B
EXAONE·64K CTX· CHAT
23
TOK/S · 11% VRAM
›
C
OPT 1.3B1.3B
OPT·2K CTX· CHAT
23
TOK/S · 11% VRAM
›
C
MiniCPM-V 4.61.3B
OTHER·256K CTX· CHAT· VISION
23
TOK/S · 11% VRAM
›
C
DeepSeek Coder 1.3B1.35B
DEEPSEEK·16K CTX· CODING
22
TOK/S · 11% VRAM
›
C
Phi-1 1.3B1.42B
PHI·2K CTX· CODING
21
TOK/S · 11% VRAM
›
C
Phi-1.5 1.3B1.42B
PHI·2K CTX· CHAT· CODING
21
TOK/S · 11% VRAM
›
C
LFM2 8B A1B8.3BMoE
LFM·125K CTX· CHAT· TOOL_USE· MULTILINGUAL
20
TOK/S · 46% VRAM
›
▸ NEXT STEP

Get personalized recommendations.

See ranked models with benchmark scores, run commands, and precise speed estimates for your Google Tensor G3 GPU.

▸Find models for Google Tensor G3 GPUCompare with other GPUs
WHAT THIS CARD IS WORTH

Google Tensor G3 GPU holds 248 of the models in our catalogue and is, in practice, a IQ4_XS card — the largest it takes is Ling-lite 16.8B at IQ4_XS.

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
Ling-lite 16.8B16.8B · IQ4_XS10.7 GB19 tok/sEST
DeepSeek V2 Lite 16B16B · Q4_K_S10.6 GB18 tok/sEST
StarCoder2 15B15.96B · Q4_K_S10.7 GB3 tok/sEST
DeepSeek-Coder-V2-Lite 15.7B15.7B · Q4_K_S10.4 GB18 tok/sEST
DeepSeek R1 Distill Qwen 14B14.8B · Q4_K_S10.4 GB3 tok/sEST
DeepCoder 14B14.8B · Q4_K_S10.4 GB3 tok/sEST
Qwen2.5-Coder-14B14.8B · Q4_K_S10.4 GB3 tok/sEST
Qwen2.5-14B14.8B · Q4_K_S10.4 GB3 tok/sEST
AGAINST CARDS OF SIMILAR MEMORY
M4 Pro (24GB)16 GB$1,3992.5 tok/s per $100
M4 (24GB)16 GB$6992.9 tok/s per $100
M2 (24GB)16 GB$9991.7 tok/s per $100
M3 (24GB)16 GB$9991.7 tok/s per $100
Google Tensor G3 GPU12 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

Google Tensor G3 GPU — 12 GB VRAM.

▸ GOOGLE TENSOR G3 GPU SPEC
BRAND
Apple
VRAM
12 GB Shared
BANDWIDTH
34 GB/s
FP16 COMPUTE
1.5 TFLOPS
TDP
7 W
ARCHITECTURE
Immortalis Gen3
▸ AI CAPABILITY
248/ 449 models @ Q4

With 12 GB VRAM and 34 GB/s bandwidth, this GPU handles models up to 14.8B parameters.

Speed ≈ bandwidth / model_size × efficiency. A 7B model at Q4 runs at ~4 tok/s.

RENT IT — LIVE PRICES
Checking what the clouds are charging…
§ 01TOP MODELS FOR GOOGLE TENSOR G3 GPU
248 FIT · SHOWING 20
MODELSIZEVRAM Q4TOK/SAVG
DeepSeek R1 Distill Qwen 14B14.8B9.5 GB243.9
DeepCoder 14B14.8B9.5 GB238.7
Qwen2.5-Coder-14B14.8B9.5 GB241.3
Qwen2.5-14B14.8B9.5 GB241.3
Qwen3 14B14.8B9.5 GB245.7
phi-4 14B14.66B9.4 GB233.7
Phi-4-reasoning 14B14.66B9.4 GB233.7
Phi-4-reasoning-plus 14B14.66B9.4 GB275.5
Phi-3-medium-14b14B9.0 GB233.7
Qwen 1.5 14B14B9.0 GB241.3
Ministral 3 14B Reasoning13.95B9.0 GB2—
Baichuan2 13B13B8.4 GB223.6
Llama 2 13B13B8.4 GB217.2
CodeLlama 13B13B8.4 GB219.7
Vicuna 13B13B8.4 GB211.8
LLaMA 1 13B13B8.4 GB232.9
OPT 13B13B8.4 GB235.8
Orca 2 13B13B8.4 GB225.4
WizardCoder Python 13B13B8.4 GB260.1
WizardLM 13B13B8.4 GB219.5