Phi-4 14B
Phi-4 14B needs 11 GB of VRAM at Q4_K_M with 8k context, computed from its config.json as checked on 16 October 2026. The smallest card that holds it has 12 GB, a tight fit, and the fastest desktop card that runs it, the RTX 5090, generates at most 147 tokens per second.
- Parameters
- 14.7B
- Context
- 16,384 tokens
- Licence
- MIT
- Released
- December 2024
- Config checked
- 16 October 2026
Open Phi-4 14B in the LLM checker
Which card size runs it
11 GB against each memory size. A model may use 95% of a card, and only what is left over, or missing, is coloured; tight means above 85%, with little room for a longer conversation.
Other quantisations, context lengths or a Q8 KV cache: the VRAM calculator.
Computed 16 October 2026Which GPU to pick
The smallest card here that loads it at Q4_K_M with 8k of context is the RTX 3060 (12 GB): 11 GB needed, an estimated 30 tokens per second, and a tight fit that leaves little for the conversation. For room to work, meaning 16,384 tokens of context while staying under 85% of memory, step up to the RTX 4060 Ti (16 GB), which needs 12.7 GB for that and generates at about 24 tokens per second.
At Q8_0, which is near-lossless, the card is the RX 7900 XT (20 GB), needing 18.4 GB. On Apple silicon, a Mac mini M4 Pro holds it with 16,384 tokens of context at an estimated 22 tokens per second; prompt processing is slower than on Nvidia, which long documents make obvious. Among unified-memory desktops, the Ryzen AI Max+ 395 holds it at 16,384 tokens in 12.7 GB, at an estimated 21 tokens per second.
Speed on each card
| GPU | Fits | Est. tokens/s | Feels like | Memory | Bandwidth |
|---|---|---|---|---|---|
| H100 SXM | Yes | ≤ 275 | faster than you can read | 80 GB | 3,350 GB/s |
| A100 | Yes | ≤ 167 | faster than you can read | 80 GB | 2,039 GB/s |
| RTX 5090 | Yes | ≤ 147 | faster than you can read | 32 GB | 1,792 GB/s |
| RTX PRO 6000 Blackwell | Yes | ≤ 147 | faster than you can read | 96 GB | 1,792 GB/s |
| RTX PRO 5000 Blackwell | Yes | ≤ 110 | faster than you can read | 48 GB | 1,344 GB/s |
| RTX 4090 | Yes | 83 | faster than you can read | 24 GB | 1,008 GB/s |
| RTX 5080 | Yes | 79 | faster than you can read | 16 GB | 960 GB/s |
| RX 7900 XTX | Yes | 79 | faster than you can read | 24 GB | 960 GB/s |
| RTX 6000 Ada | Yes | 79 | faster than you can read | 48 GB | 960 GB/s |
| RTX 3090 | Yes | 77 | faster than you can read | 24 GB | 936 GB/s |
| RTX 5070 Ti | Yes | 74 | faster than you can read | 16 GB | 896 GB/s |
| RTX PRO 4500 Blackwell | Yes | 74 | faster than you can read | 32 GB | 896 GB/s |
| L40S | Yes | 71 | faster than you can read | 48 GB | 864 GB/s |
| RX 7900 XT | Yes | 66 | faster than you can read | 20 GB | 800 GB/s |
| RTX A6000 | Yes | 63 | faster than you can read | 48 GB | 768 GB/s |
| RTX 4080 SUPER | Yes | 60 | faster than you can read | 16 GB | 736 GB/s |
| RTX 5070 | Tight | 55 | comfortable for chat | 12 GB | 672 GB/s |
| RTX PRO 4000 Blackwell | Yes | 55 | comfortable for chat | 24 GB | 672 GB/s |
| RX 9070 XT | Yes | 53 | comfortable for chat | 16 GB | 640 GB/s |
| Radeon AI PRO R9700 | Yes | 53 | comfortable for chat | 32 GB | 640 GB/s |
| Arc Pro B70 | Yes | 50 | comfortable for chat | 32 GB | 608 GB/s |
| RTX 4070 | Tight | 41 | comfortable for chat | 12 GB | 504 GB/s |
| Arc B580 | Tight | 37 | comfortable for chat | 12 GB | 456 GB/s |
| Arc Pro B60 | Yes | 37 | comfortable for chat | 24 GB | 456 GB/s |
| RTX 5060 Ti | Yes | 37 | comfortable for chat | 16 GB | 448 GB/s |
| RTX PRO 4000 Blackwell SFF | Yes | 35 | comfortable for chat | 24 GB | 432 GB/s |
| RTX 3060 | Tight | 30 | comfortable for chat | 12 GB | 360 GB/s |
| RX 9060 XT | Yes | 26 | comfortable for chat | 16 GB | 320 GB/s |
| RTX Spark | Yes | 25 | usable, a little slow | 128 GB (102 for the GPU) | 300 GB/s |
| RTX 4060 Ti | Yes | 24 | usable, a little slow | 16 GB | 288 GB/s |
| DGX Spark | Yes | 22 | usable, a little slow | 128 GB (126 for the GPU) | 273 GB/s |
| Ryzen AI Max+ 395 | Yes | 21 | usable, a little slow | 128 GB (96 for the GPU) | 256 GB/s |
Single-stream decode, from memory bandwidth at 70% efficiency. Prompt processing and batching not included. These are ceilings, not measurements: Phi-4 14B reads only 8.5 GB of weights per token at Q4, so little that the per-token costs the formula leaves out decide much of the real speed. The 5 figures marked ≤ are where a real runtime falls furthest below the number shown. See the speed estimator for other quantisations and Apple hardware, or every card compared if you are choosing hardware rather than a model.
Computed 16 October 2026Memory by quantisation and context
| Quant | 4,096 ctx | 8,192 ctx | 32,768 ctx |
|---|---|---|---|
| Q8_0 | 17.5 GB · fits 24 GB | 18.4 GB · fits 24 GB | beyond its 16,384-token window |
| Q5_K_M | 12.2 GB · fits 16 GB | 13 GB · fits 16 GB | beyond its 16,384-token window |
| Q4_K_M | 10.2 GB · fits 12 GB | 11 GB · fits 12 GB | beyond its 16,384-token window |
Weights at this quant: 8.5 GB at Q4. Every extra 1,000 tokens of context adds about 0.205 GB of KV cache at FP16. Try other settings in the VRAM calculator.
About the model
Phi-4 14B has 14.7B parameters, 40 layers and 10 KV heads of dimension 128. Microsoft's 14B model from December 2024, trained largely on synthetic data with an emphasis on reasoning and mathematics. Small, MIT-licensed and English-first.
Text only 16,384-token window MIT December 2024
Reasoning. Answers directly. Microsoft publishes separate Phi-4-reasoning variants.
Native context window: 16,384 tokens. Licence: MIT, a permissive licence. Weights published in December 2024.
Good at
- Short, self-contained reasoning and maths problems.
- Projects that need the most permissive licence there is.
Watch out for
- A 16,384-token window. That is the limit of the model, not of your card, and it rules out long documents and repository-scale coding.
- Primarily English.
- Every layer keeps a full-length cache, so long context costs several times what a 2026 hybrid model pays for the same conversation.
Similar models
The models that need about the same memory as Phi-4 14B at Q4_K_M and 8k, which makes them the real alternatives on whatever card you have. The context column is what each extra 1,000 tokens costs, and it is where models of the same size differ most.
| Model | Size | Needs | Per 1k context | Window | Released |
|---|---|---|---|---|---|
| Phi-4 14B | 14.7B | 11 GB | 0.205 GB | 16k | 2024-12 |
| Qwen3 14B | 14.8B | 10.8 GB | 0.164 GB | 40k | 2025-04 |
| Ministral 3 14B | 14B | 10.3 GB | 0.164 GB | 256k | 2025-12 |
| gpt-oss 20B | 21B · 3.6B active | 13.4 GB | 0.025 GB | 128k | 2025-08 |
| Gemma 3 12B | 12.2B | 8.7 GB | 0.066 GB | 128k | 2025-03 |
See every model compared, which models fit on 8 to 96 GB, or the cheapest hour of every cloud card.
Questions
- Can I run Phi-4 14B on a 24 GB card?
- Yes, at Q4_K_M and 8k context it needs about 11 GB, which fits a 24 GB card with room.
- How much VRAM does Phi-4 14B need at Q8?
- About 18.4 GB at 8k context. Q8 is near-lossless; use it when it fits.
- What is the context length of Phi-4 14B?
- 16,384 tokens natively. That is short by current standards, and it is a limit of the model rather than of your GPU: spare memory does not extend it.
- How fast is Phi-4 14B on an RTX 4090?
- At most about 83 tokens per second at Q4, single stream, which is faster than you can read. That is a ceiling worked out from the card's memory bandwidth, not a measurement.
- Architecture values from config.json in microsoft/phi-4 on Hugging Face, checked 16 October 2026. Every figure on this page is computed from those values; check them before buying hardware for this model.
- Estimates assume a single conversation on a card that is otherwise free. A desktop on the same GPU takes 0.5 to 2 GB.