RUNLOCALAIv38
->Will it run?Best GPUCompareTroubleshootStartLearnPulseModelsHardwareToolsBench
Run check
RUNLOCALAI

Independently operated catalog for local-AI hardware and software. Hand-written verdicts. Source-cited claims. Reproducible commands when we have them.

OP·Fredoline Eruo
DIR
  • Models
  • Hardware
  • Tools
  • Benchmarks
TOOLS
  • Will it run?
  • Compare hardware
  • Cost vs cloud
  • Choose my GPU
  • Prompting kits
  • Quick answers
REF
  • All buyer guides
  • Learn local AI
  • Methodology
  • Glossary
  • Errors KB
  • Trust
EDITOR
  • About
  • Author
  • How we make money
  • Editorial policy
  • Contact
LEGAL
  • Privacy
  • Terms
  • Sitemap
MAIL · MONTHLY DIGEST
Get monthly local AI changes
Monthly recap. No spam.
DISCLOSURE

Some links on this site are affiliate links (Amazon Associates and other first-class retailers). When you buy through them, we earn a small commission at no extra cost to you. Affiliate links do not influence our verdicts — there are cards we rate highly that we don't have affiliate relationships with, and cards that sell well that we refuse to recommend. Read more →

© 2026 runlocalai.coIndependently operated
RUNLOCALAI · v38
Will it run? / NVIDIA RTX 4090 48GB (China-mod) / vision

What can NVIDIA RTX 4090 48GB (China-mod) run for vision?

Build: NVIDIA RTX 4090 48GB (China-mod) + — + 32 GB RAM (windows)

Memory: 48 GB VRAM + 32 GB system RAM
Runner: llama.cpp / Ollama (CUDA)
AnyChatCodingAgentsReasoningVisionLong contextCreative

Runs comfortably
28 models

Ranked by fit for vision use case + predicted speed. Click a row for VRAM breakdown.

#1Gemma 4 E2B (Effective 2B)
2B
gemma
Commercial OK
Quant: Q8_0Context: 8,192VRAM: 5.0 GBHeadroom: 43.0 GBTTFT: instant
ollama run gemma4:e2b
308
tok/s
Estimated
Weights
2.13 GB
KV cache
1.00 GB
Activations
0.11 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~62 ms (instant)
Model details →
#2Phi-3.5 Vision
4.2B
phi
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 6.6 GBHeadroom: 41.4 GBTTFT: fast
258
tok/s
Estimated
Weights
2.54 GB
KV cache
2.10 GB
Activations
0.14 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~130 ms (fast)
Model details →
#3Gemma 4 E4B (Effective 4B)
4B
gemma
Commercial OK
Quant: Q8_0Context: 8,192VRAM: 8.3 GBHeadroom: 39.7 GBTTFT: fast
ollama run gemma4:e4b
154
tok/s
Estimated
Weights
4.25 GB
KV cache
2.00 GB
Activations
0.22 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~124 ms (fast)
Model details →
#4Gemma 3 4B
4B
gemma
Commercial OK
Quant: Q8_0Context: 8,192VRAM: 8.3 GBHeadroom: 39.7 GBTTFT: fast
ollama run gemma3:4b
154
tok/s
Estimated
Weights
4.25 GB
KV cache
2.00 GB
Activations
0.22 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~124 ms (fast)
Model details →
#5Molmo 7B-D
8B
other
Commercial OK
Quant: Q4_K_MContext: 4,096VRAM: 8.9 GBHeadroom: 39.1 GBTTFT: fast
136
tok/s
Estimated
Weights
4.83 GB
KV cache
2.00 GB
Activations
0.25 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~248 ms (fast)
Model details →
#6LLaVA-OneVision 7B
7B
other
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 9.7 GBHeadroom: 38.3 GBTTFT: fast
155
tok/s
Estimated
Weights
4.23 GB
KV cache
3.50 GB
Activations
0.22 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~217 ms (fast)
Model details →
#7Qwen 2.5-VL 7B
7B
qwen
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 9.7 GBHeadroom: 38.3 GBTTFT: fast
155
tok/s
Estimated
Weights
4.23 GB
KV cache
3.50 GB
Activations
0.22 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~217 ms (fast)
Model details →
#8Janus-Pro 7B
7B
janus
Commercial OK
Quant: Q4_K_MContext: 4,096VRAM: 8.0 GBHeadroom: 40.0 GBTTFT: fast
155
tok/s
Estimated
Weights
4.23 GB
KV cache
1.75 GB
Activations
0.22 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~217 ms (fast)
Model details →
#9MiniCPM-V 3 8B
8B
minicpm
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 10.9 GBHeadroom: 37.1 GBTTFT: fast
136
tok/s
Estimated
Weights
4.83 GB
KV cache
4.00 GB
Activations
0.25 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~248 ms (fast)
Model details →
#10Qwen 2-VL 7B
7B
qwen
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 9.7 GBHeadroom: 38.3 GBTTFT: fast
155
tok/s
Estimated
Weights
4.23 GB
KV cache
3.50 GB
Activations
0.22 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~217 ms (fast)
Model details →
#11Qwen 2.5-VL 3B
3B
qwen
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 5.2 GBHeadroom: 42.8 GBTTFT: instant
362
tok/s
Estimated
Weights
1.81 GB
KV cache
1.50 GB
Activations
0.10 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~93 ms (instant)
Model details →
#12Moondream 2
1.9B
other
Commercial OK
Quant: Q4_K_MContext: 2,048VRAM: 3.2 GBHeadroom: 44.8 GBTTFT: instant
571
tok/s
Estimated
Weights
1.15 GB
KV cache
0.24 GB
Activations
0.06 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~59 ms (instant)
Model details →

Runs with tradeoffs
3 models

Tight VRAM, partial CPU offload, or context-limited.

Gemma 3 27B
27B
gemma
Commercial OK
Quant: Q8_0Context: 8,192VRAM: 45.4 GBHeadroom: 2.6 GBTTFT: noticeable
  • • Tight VRAM fit — only 2.6 GB headroom left for context growth
ollama run gemma3:27b
23
tok/s
Estimated
Weights
28.69 GB
KV cache
13.50 GB
Activations
1.44 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~837 ms (noticeable)
Model details →
Molmo 72B
72B
other
Commercial OK
Quant: Q4_K_MContext: 4,096VRAM: 65.4 GBHeadroom: 1.8 GBTTFT: slow
  • • Partial CPU offload: ~27% of layers run on CPU
15
tok/s
Estimated
Weights
43.47 GB
KV cache
18.00 GB
Activations
2.18 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~2231 ms (slow)
Model details →
InternVL 2.5 78B
78B
other
Commercial OK
Quant: Q4_K_MContext: 2,048VRAM: 61.0 GBHeadroom: 6.2 GBTTFT: slow
  • • Partial CPU offload: ~21% of layers run on CPU
14
tok/s
Estimated
Weights
47.09 GB
KV cache
9.75 GB
Activations
2.36 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~2417 ms (slow)
Model details →

What if you upgraded?

Hypothetical scenarios. We re-ran the compatibility engine for each.

+32 GB system RAM

~$80–150

Doubles your CPU-offload working set. Helps when models don't quite fit in VRAM.

Unlocks: 230 new comfortable, 24 new tradeoff

  • • all-MiniLM-L6-v2
  • • Piper
  • • Whisper Tiny
  • • Whisper Base
Shop this upgrade↗

Upgrade to NVIDIA A100 80GB SXM

see current pricing

80 GB VRAM (vs your 48 GB) plus a bandwidth jump from ~1008 GB/s to ~2039 GB/s.

Unlocks: 245 new comfortable

  • • all-MiniLM-L6-v2
  • • Piper
  • • Whisper Tiny
  • • Whisper Base
Shop this upgrade↗

Add a second NVIDIA RTX 4090 48GB (China-mod)

~$2400

Tensor parallelism splits the model across both cards, effectively doubling VRAM. Bandwidth doesn't double — runs ~1.5× the single-card speed in practice.

Unlocks: 255 new comfortable

  • • all-MiniLM-L6-v2
  • • Piper
  • • Whisper Tiny
  • • Whisper Base
Shop this upgrade↗

Some links above are affiliate links. We may earn a commission at no extra cost to you. How we make money.

Won't run
top 5 popular models

Need more memory than you have. Shown for orientation.

DeepSeek V4 Pro (1.6T MoE)
1600B
deepseek
Commercial OK

Even with CPU offload, needs more memory than your VRAM (48 GB) + 60% of system RAM (19 GB) combined.

—
Qwen 3.5 235B-A17B (MoE)
397B
qwen
Commercial OK

Even with CPU offload, needs more memory than your VRAM (48 GB) + 60% of system RAM (19 GB) combined.

—
Qwen 3 235B-A22B
235B
qwen
Commercial OK

Even with CPU offload, needs more memory than your VRAM (48 GB) + 60% of system RAM (19 GB) combined.

—
DeepSeek R1 (671B reasoning)
671B
deepseek
Commercial OK

Even with CPU offload, needs more memory than your VRAM (48 GB) + 60% of system RAM (19 GB) combined.

—
Llama 4 Scout
109B
llama
Commercial OK

Even with CPU offload, needs more memory than your VRAM (48 GB) + 60% of system RAM (19 GB) combined.

—

How to read these numbers

Measured here
Measured here - RunLocalAI ran this exact combo on owner hardware with public evidence.

Source-backed
Source-backed / community - a reproduced public source supports the speed, but it is not labeled as owner-measured.

Extrapolated
Extrapolated - predicted from a measured benchmark on similar-bandwidth hardware.

Estimated
Estimated - formula based on VRAM bandwidth and model architecture; not a benchmark row.

RunLocalAI Will-It-Run Framework →

Want a specific benchmark we don't have? Email Contact support and we'll prioritize it.