Home / Models / Phi-3 Medium 128K Instruct
Phi-3 Medium 128K Instruct
14B parameters · native context 131,072 · phi3
Estimated VRAM requirement
Weight memory uses parameter count × bits per weight, plus a quantization overhead factor. KV cache uses standard GQA formulas. Values are calculated estimates, not measured allocations.
| Quantization | Estimated weights | 2K total | 4K total | 8K total | 16K total | 32K total | 64K total | 128K total |
|---|---|---|---|---|---|---|---|---|
| FP16 | 26.6 GB | 29.9 GB | 30.3 GB | 31.0 GB | 32.6 GB | 35.7 GB | 42.0 GB | 54.5 GB |
| Q8 | 15.0 GB | 17.3 GB | 17.7 GB | 18.5 GB | 20.0 GB | 23.2 GB | 29.4 GB | 41.9 GB |
| Q6 | 11.7 GB | 13.7 GB | 14.1 GB | 14.9 GB | 16.5 GB | 19.6 GB | 25.8 GB | 38.3 GB |
| Q5 | 9.9 GB | 11.8 GB | 12.2 GB | 13.0 GB | 14.5 GB | 17.6 GB | 23.9 GB | 36.4 GB |
| Q4 | 8.2 GB | 10.0 GB | 10.4 GB | 11.2 GB | 12.7 GB | 15.9 GB | 22.1 GB | 34.6 GB |
| Q3 | 6.6 GB | 8.2 GB | 8.6 GB | 9.4 GB | 11.0 GB | 14.1 GB | 20.3 GB | 32.8 GB |
Find a GPU for this model
Run at the quantization and context selected below. Defaults are Q4 and 8K when you do not change the controls. Results are full-VRAM fits only — not a speed ranking.
- Run at
- Q4 · 8K context
- Required VRAM
- 11.2 GB
No fresh matched Amazon price is currently available for a full-VRAM fit. See compatible GPUs · Workstations · Rent a server · Buy vs rent.
Lowest-cost full-VRAM option
Sorted by lowest current Amazon price among GPUs that fit entirely in VRAM. This is not a performance ranking.
No compatible GPU currently has a fresh Amazon price.
More VRAM headroom
Sorted by leftover usable VRAM after the calculated requirement. Headroom is not tokens/sec.
NVIDIA B200 192 GB
- GPU VRAM
- 192 GB advertised
- Required VRAM
- 11.2 GB
- Usable GPU VRAM
- 172.8 GB
- VRAM headroom
- 161.6 GB — leftover usable memory after the calculated requirement, not a speed ranking
- Compatible
- Yes — fits entirely in VRAM
- Product SKUs
- 0 mapped · 0 Amazon-matched
- Amazon
- Check availability
NVIDIA H200 141 GB
- GPU VRAM
- 141 GB advertised
- Required VRAM
- 11.2 GB
- Usable GPU VRAM
- 126.9 GB
- VRAM headroom
- 115.7 GB — leftover usable memory after the calculated requirement, not a speed ranking
- Compatible
- Yes — fits entirely in VRAM
- Product SKUs
- 0 mapped · 0 Amazon-matched
- Amazon
- Check availability
NVIDIA GH200 96 GB
- GPU VRAM
- 96 GB advertised
- Required VRAM
- 11.2 GB
- Usable GPU VRAM
- 86.4 GB
- VRAM headroom
- 75.2 GB — leftover usable memory after the calculated requirement, not a speed ranking
- Compatible
- Yes — fits entirely in VRAM
- Product SKUs
- 0 mapped · 0 Amazon-matched
- Amazon
- Check availability
Other compatible GPUs
NVIDIA RTX 4500 24 GB
- GPU VRAM
- 24 GB advertised
- Required VRAM
- 11.2 GB
- Usable GPU VRAM
- 21.6 GB
- VRAM headroom
- 10.4 GB — leftover usable memory after the calculated requirement, not a speed ranking
- Compatible
- Yes — fits entirely in VRAM
- Product SKUs
- 2 mapped · 0 Amazon-matched
- Amazon
- Check availability
NVIDIA RTX 3090 24 GB
- GPU VRAM
- 24 GB advertised
- Required VRAM
- 11.2 GB
- Usable GPU VRAM
- 21.6 GB
- VRAM headroom
- 10.4 GB — leftover usable memory after the calculated requirement, not a speed ranking
- Compatible
- Yes — fits entirely in VRAM
- Product SKUs
- 18 mapped · 9 Amazon-matched
- Amazon
- Check price on Amazon
NVIDIA RTX 5000 32 GB
- GPU VRAM
- 32 GB advertised
- Required VRAM
- 11.2 GB
- Usable GPU VRAM
- 28.8 GB
- VRAM headroom
- 17.6 GB — leftover usable memory after the calculated requirement, not a speed ranking
- Compatible
- Yes — fits entirely in VRAM
- Product SKUs
- 2 mapped · 1 Amazon-matched
- Amazon
- Check price on Amazon
Compatible GPUs
Status is a calculated memory fit against usable VRAM (advertised × 0.9). Open a GPU to see product SKUs and Amazon listings.
Compare GPUs for this model
Published chip comparisons where both GPUs fully fit this model at Q4 / 8K. Opens with this model preselected; that query is not indexed.
How fast can this model run?
llama.cpp CUDA decode at Q4 under llama-bench tg128 conditions. Not a ranking. Open a pair for prefill and methodology.
| GPU | Decode | State | Confidence |
|---|---|---|---|
| NVIDIA RTX 5090 32 GB | ~130 tok/s | Estimated | MEDIUM |
| NVIDIA RTX 4090 24 GB | ~82 tok/s | Estimated | MEDIUM |
| NVIDIA RTX 5080 16 GB | ~80 tok/s | Estimated | MEDIUM |
| NVIDIA RTX 5070 Ti 16 GB | ~79 tok/s | Estimated | MEDIUM |
| NVIDIA RTX 3090 Ti 24 GB | ~75 tok/s | Estimated | MEDIUM |
| NVIDIA RTX 3090 24 GB | ~70 tok/s | Estimated | MEDIUM |
| NVIDIA RTX 4080 SUPER 16 GB | ~64 tok/s | Estimated | MEDIUM |
| NVIDIA RTX A6000 48 GB | ~63 tok/s | Estimated | MEDIUM |
| NVIDIA RTX 4080 16 GB | ~62 tok/s | Estimated | MEDIUM |
| NVIDIA RTX 4070 Ti SUPER 16 GB | ~58 tok/s | Estimated | MEDIUM |
Sources
Model specifications: official repository microsoft/Phi-3-medium-128k-instruct · revision main. Compatibility: RigForAI calculator v1.0.0 (estimated VRAM). GPU products: Icecat. Amazon is a separate affiliate match and is not used as a spec source.