FitMyLLM
▸ IMAGINATION· POWERVR ROGUE

Imagination PowerVR GM9446

Running LLMs on the Imagination PowerVR GM9446 — the long read: which models fit at which quantisation, and the settings worth changing. · Or what a budget buys

VRAM
4 GB
BUDGET
BANDWIDTH
17
GB/S
MODELS Q4
97/449
22%
7B Q4 SPEED
~2
SLOW
▸ MODEL COVERAGE @ Q422% OF ALL
▸ ESTIMATED SPEED· BY MODEL SIZE @ Q4

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

3B
~5
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
4 GB
BANDWIDTH
17 GB/s
FP16 COMPUTE
0.3 TFLOPS
TDP
3W
MEMORY
Shared
ARCHITECTURE
PowerVR Rogue
15
FAST MODELS · >30 TOK/S
Real-time chat speed
45
USABLE · >10 TOK/S
Comfortable for all tasks
97
TOTAL COMPATIBLE
Fit in VRAM at Q4
▸ DON’T WANT TO BUY?

Test Imagination PowerVR GM9446 (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· 97
S
Falcon-H1R Tiny 90M0.09B
FALCON·256K CTX· CHAT· REASONING
168
TOK/S · 14% VRAM
›
S
SmolLM2 135M0.135B
SMOLLM·2K CTX· CHAT
112
TOK/S · 14% VRAM
›
S
nomic-embed-text-v1.5 100M0.14B
EMBEDDING·8K CTX· CHAT
108
TOK/S · 14% VRAM
›
S
GPT-2 124M0.14B
GPT2·1K CTX· CHAT
108
TOK/S · 14% VRAM
›
A
SmolVLM 256M0.256B
SMOLLM·8K CTX· CHAT· VISION
59
TOK/S · 16% VRAM
›
A
Gemma 3 270M0.27B
GEMMA·32K CTX· CHAT
56
TOK/S · 16% VRAM
›
A
Snowflake Arctic Embed M v2.00.305B
EMBEDDING·8K CTX· EMBEDDING· MULTILINGUAL
50
TOK/S · 17% VRAM
›
A
bge-large-en-v1.5 335M0.335B
EMBEDDING·1K CTX· CHAT
45
TOK/S · 17% VRAM
›
A
mxbai-embed-large-v10.335B
EMBEDDING·1K CTX· EMBEDDING
45
TOK/S · 17% VRAM
›
A
Snowflake Arctic Embed L0.335B
EMBEDDING·1K CTX· EMBEDDING
45
TOK/S · 17% VRAM
›
A
LFM2 350M0.35B
LFM·125K CTX· CHAT· TOOL_USE
43
TOK/S · 18% VRAM
›
A
SmolLM2 360M0.36B
SMOLLM·8K CTX· CHAT
42
TOK/S · 18% VRAM
›
A
GPT-2 Medium 345M0.38B
GPT2·1K CTX· CHAT
40
TOK/S · 18% VRAM
›
B
Qwen 2.5 0.5B0.5B
QWEN·32K CTX· CHAT
30
TOK/S · 20% VRAM
›
B
SmolVLM 500M0.5B
SMOLLM·8K CTX· CHAT· VISION
30
TOK/S · 20% VRAM
›
B
Falcon-H1 0.5B0.52B
FALCON·128K CTX· CHAT
29
TOK/S · 20% VRAM
›
B
BGE-M30.568B
EMBEDDING·8K CTX· EMBEDDING
27
TOK/S · 21% VRAM
›
B
Snowflake Arctic Embed L v2.00.568B
EMBEDDING·8K CTX· EMBEDDING· MULTILINGUAL
27
TOK/S · 21% VRAM
›
B
Qwen3-Embedding 0.6B0.6B
EMBEDDING·32K CTX· EMBEDDING· MULTILINGUAL
25
TOK/S · 21% VRAM
›
B
Falcon-H1R Tiny 0.6B0.6B
FALCON·32K CTX· CHAT· REASONING
25
TOK/S · 21% VRAM
›
B
Falcon Perception 0.6B0.6B
FALCON·4K CTX· VISION
25
TOK/S · 21% VRAM
›
C
Qwen 1.5 0.5B0.62B
QWEN·32K CTX· CHAT
24
TOK/S · 22% VRAM
›
C
Qwen3 0.6B0.75B
QWEN·32K CTX· CHAT· REASONING
20
TOK/S · 24% VRAM
›
C
LFM2 700M0.74B
LFM·125K CTX· CHAT· TOOL_USE
20
TOK/S · 24% VRAM
›
C
GPT-2 Large 774M0.81B
GPT2·1K CTX· CHAT
19
TOK/S · 25% VRAM
›
C
Qwen 3.5 0.8B0.87B
QWEN·256K CTX· CHAT· CODING· MULTILINGUAL
17
TOK/S · 26% VRAM
›
C
InternVL3 1B0.94B
OTHER·32K CTX· CHAT· VISION
16
TOK/S · 27% VRAM
›
D
TinyLlama 1.1B1.1B
LLAMA·2K CTX· CHAT
14
TOK/S · 29% VRAM
›
D
MiniCPM5 1B1.08B
MINICPM·128K CTX· CHAT· REASONING· MULTILINGUAL
14
TOK/S · 29% VRAM
›
D
LFM2.5-1.2B-Thinking1.2B
LFM·122K CTX· CHAT· REASONING· TOOL_USE
13
TOK/S · 31% VRAM
›
D
Llama-3.2-1B1.2B
LLAMA·4K CTX· CHAT
13
TOK/S · 31% VRAM
›
D
LFM2 1.2B1.2B
LFM·125K CTX· CHAT· TOOL_USE· MULTILINGUAL
13
TOK/S · 31% VRAM
›
D
Zamba2 1.2B1.2B
OTHER·4K CTX· CHAT
13
TOK/S · 31% VRAM
›
D
EXAONE-4.0-1.2B1.3B
EXAONE·64K CTX· CHAT
12
TOK/S · 32% VRAM
›
D
OPT 1.3B1.3B
OPT·2K CTX· CHAT
12
TOK/S · 32% VRAM
›
D
MiniCPM-V 4.61.3B
OTHER·256K CTX· CHAT· VISION
12
TOK/S · 32% VRAM
›
D
Phi-1 1.3B1.42B
PHI·2K CTX· CODING
11
TOK/S · 34% VRAM
›
D
Phi-1.5 1.3B1.42B
PHI·2K CTX· CHAT· CODING
11
TOK/S · 34% VRAM
›
D
DeepSeek Coder 1.3B1.35B
DEEPSEEK·16K CTX· CODING
11
TOK/S · 33% VRAM
›
D
Falcon-H1 1.5B1.55B
FALCON·128K CTX· CHAT· CODING
10
TOK/S · 36% VRAM
›
▸ NEXT STEP

Get personalized recommendations.

See ranked models with benchmark scores, run commands, and precise speed estimates for your Imagination PowerVR GM9446.

WHAT THIS CARD IS WORTH

Imagination PowerVR GM9446 holds 97 of the models in our catalogue and is, in practice, a Q4_K_M card — the largest it takes is InternLM2 5B 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
InternLM2 5B4.5B · Q4_K_M3.6 GB5 tok/sEST
Ministral 3 3B Reasoning4.25B · Q4_K_M3.3 GB5 tok/sEST
Qwen3 4B4B · Q4_K_M3.3 GB5 tok/sEST
Qwen3-4B Instruct 25074B · Q4_K_M3.3 GB5 tok/sEST
Qwen3-Embedding 4B4B · Q4_K_M3.3 GB5 tok/sEST
Nemotron 3 Nano 4B3.97B · Q4_K_M3.4 GB5 tok/sEST
Ministral 3 3B3.85B · Q5_K_M3.5 GB5 tok/sEST
phi-3-mini-4k 3.8B3.8B · Q4_K_M3.4 GB5 tok/sEST
AGAINST CARDS OF SIMILAR MEMORY
RTX 3070 Ti8 GB$49912.6 tok/s per $100
RTX 3060 Ti GDDR6X8 GB$39915.8 tok/s per $100
RTX 3070 Ti 8 GB GA1028 GB$59910.5 tok/s per $100
Arc A7508 GB$19910.1 tok/s per $100
Imagination PowerVR GM94464 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

Imagination PowerVR GM9446 — 4 GB VRAM.

▸ IMAGINATION POWERVR GM9446 SPEC
BRAND
Apple
VRAM
4 GB Shared
BANDWIDTH
17 GB/s
FP16 COMPUTE
0.3 TFLOPS
TDP
3 W
ARCHITECTURE
PowerVR Rogue
▸ AI CAPABILITY
97/ 449 models @ Q4

With 4 GB VRAM and 17 GB/s bandwidth, this GPU handles models up to 4.25B parameters.

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

§ 01TOP MODELS FOR IMAGINATION POWERVR GM9446
97 FIT · SHOWING 20
MODELSIZEVRAM Q4TOK/SAVG
Ministral 3 3B Reasoning4.25B3.1 GB4—
Qwen 1.5 4B4B2.9 GB412.6
Qwen3 4B4B2.9 GB440.7
Qwen3-4B Instruct 25074B2.9 GB437.2
Qwen3-Embedding 4B4B2.9 GB4—
Nemotron 3 Nano 4B3.97B2.9 GB432.0
Ministral 3 3B3.85B2.8 GB421.4
Phi-3.5 Mini 3.8B3.82B2.8 GB446.6
phi-3-mini-4k 3.8B3.8B2.8 GB430.5
Phi-4-mini 3.8B3.8B2.8 GB449.0
Cogito 3B3.61B2.7 GB422.1
Falcon3-3B3.23B2.5 GB525.7
granite-4.0-h-micro 3.2B3.2B2.4 GB518.4
Llama-3.2-3B3.2B2.4 GB517.9
Falcon-H1 3B3.15B2.4 GB549.5
Qwen 2.5 3B3.1B2.4 GB537.2
SmolLM3-3B3.1B2.4 GB530.5
Ministral 3B3B2.3 GB529.6
StarCoder2 3B3B2.3 GB59.5
Granite 4.1 3B3B2.3 GB516.6