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) / reasoning

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

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
160 models

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

#1Llama 3.1 Nemotron Nano 8B
8B
llama
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 →
#2NVIDIA Nemotron Nano 9B v2 Japanese
9B
other
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 12.0 GBHeadroom: 36.0 GBTTFT: fast
121
tok/s
Estimated
Weights
5.43 GB
KV cache
4.50 GB
Activations
0.28 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~279 ms (fast)
Model details →
#3RefinedNeuro RN TR R1
8B
llama
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 10.9 GBHeadroom: 37.1 GBTTFT: fast
ollama run RefinedNeuro/RN_TR_R1:latest
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 →
#4DeepSeek R1 Distill Llama 8B
8B
deepseek
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 →
#5DeepSeek R1 Distill Qwen 7B
7B
deepseek
Commercial OK
Quant: Q8_0Context: 8,192VRAM: 13.1 GBHeadroom: 34.9 GBTTFT: fast
ollama run deepseek-r1:7b
88
tok/s
Estimated
Weights
7.44 GB
KV cache
3.50 GB
Activations
0.38 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~217 ms (fast)
Model details →
#6Phi-4 Reasoning 14B
14B
phi
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 17.7 GBHeadroom: 30.3 GBTTFT: fast
ollama run phi4-reasoning:14b
78
tok/s
Estimated
Weights
8.45 GB
KV cache
7.00 GB
Activations
0.43 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~434 ms (fast)
Model details →
#7DeepSeek R1 Distill Qwen 14B
14B
deepseek
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 17.7 GBHeadroom: 30.3 GBTTFT: fast
ollama run deepseek-r1:14b
78
tok/s
Estimated
Weights
8.45 GB
KV cache
7.00 GB
Activations
0.43 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~434 ms (fast)
Model details →
#8DeepSeek V3 Lite (16B MoE)
16B
deepseek
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 20.0 GBHeadroom: 28.0 GBTTFT: instant
452
tok/s
Estimated
Weights
9.66 GB
KV cache
8.00 GB
Activations
0.49 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~74 ms (instant)
Model details →
#9DeepSeek R1 Distill Mistral 24B
24B
deepseek
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 29.0 GBHeadroom: 19.0 GBTTFT: noticeable
45
tok/s
Estimated
Weights
14.49 GB
KV cache
12.00 GB
Activations
0.73 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~744 ms (noticeable)
Model details →
#10Omni 31B Turkish Reasoning
31B
other
Quant: Q4_K_MContext: 8,192VRAM: 37.0 GBHeadroom: 11.0 GBTTFT: noticeable
35
tok/s
Estimated
Weights
18.72 GB
KV cache
15.50 GB
Activations
0.94 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~961 ms (noticeable)
Model details →
#11DeepSeek R1 Distill Qwen 32B
32B
deepseek
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 38.1 GBHeadroom: 9.9 GBTTFT: noticeable
ollama run deepseek-r1:32b
34
tok/s
Estimated
Weights
19.32 GB
KV cache
16.00 GB
Activations
0.97 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~992 ms (noticeable)
Model details →
#12QwQ 32B Preview
32B
qwen
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 38.1 GBHeadroom: 9.9 GBTTFT: noticeable
ollama run qwq:32b
34
tok/s
Estimated
Weights
19.32 GB
KV cache
16.00 GB
Activations
0.97 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~992 ms (noticeable)
Model details →

Runs with tradeoffs
15 models

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

DeepSeek R1 Distill Llama 70B
70B
deepseek
Commercial OK
Quant: Q4_K_MContext: 2,048VRAM: 54.9 GBHeadroom: 12.3 GBTTFT: slow
  • • Partial CPU offload: ~13% of layers run on CPU
ollama run deepseek-r1:70b
16
tok/s
Estimated
Weights
42.26 GB
KV cache
8.75 GB
Activations
2.12 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~2169 ms (slow)
Model details →
Hermes 4 70B FP8
70B
llama
Commercial OK
Quant: Q4_K_MContext: 2,048VRAM: 54.9 GBHeadroom: 12.3 GBTTFT: slow
  • • Partial CPU offload: ~13% of layers run on CPU
16
tok/s
Estimated
Weights
42.26 GB
KV cache
8.75 GB
Activations
2.12 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~2169 ms (slow)
Model details →
Qwen 2.5 Math 72B
72B
qwen
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 →
Llama 3.1 Nemotron 70B Instruct
70B
llama
Commercial OK
Quant: Q4_K_MContext: 2,048VRAM: 54.9 GBHeadroom: 12.3 GBTTFT: slow
  • • Partial CPU offload: ~13% of layers run on CPU
ollama run nemotron:70b
16
tok/s
Estimated
Weights
42.26 GB
KV cache
8.75 GB
Activations
2.12 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~2169 ms (slow)
Model details →
Nemotron 3 Super 49B
49B
other
Commercial OK
Quant: AWQ-INT4Context: 2,048VRAM: 59.4 GBHeadroom: 7.8 GBTTFT: noticeable
  • • Partial CPU offload: ~19% of layers run on CPU
13
tok/s
Estimated
Weights
49.00 GB
KV cache
6.13 GB
Activations
2.45 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~1519 ms (noticeable)
Model details →
ALIA 40b instruct 2601
40B
llama
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 47.2 GBHeadroom: 0.8 GBTTFT: noticeable
  • • Tight VRAM fit — only 0.8 GB headroom left for context growth
27
tok/s
Estimated
Weights
24.15 GB
KV cache
20.00 GB
Activations
1.22 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~1240 ms (noticeable)
Model details →
Llama 3.3 70B Instruct
70B
llama
Commercial OK
Quant: Q4_K_MContext: 2,048VRAM: 46.8 GBHeadroom: 1.2 GBTTFT: slow
  • • Tight VRAM fit — only 1.2 GB headroom left for context growth
ollama run llama3.3:70b
16
tok/s
Estimated
Weights
42.26 GB
KV cache
0.67 GB
Activations
2.12 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~2169 ms (slow)
Model details →
Llama 3.1 70B Instruct
70B
llama
Commercial OK
Quant: Q4_K_MContext: 2,048VRAM: 54.9 GBHeadroom: 12.3 GBTTFT: slow
  • • Partial CPU offload: ~13% of layers run on CPU
ollama run llama3.1:70b
16
tok/s
Estimated
Weights
42.26 GB
KV cache
8.75 GB
Activations
2.12 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~2169 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: 98 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: 113 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: 123 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.