family member
Phi 4 logo

Phi 4 MINI

154 tokens a second on Windows · NVIDIA 12 GB · 16 CPUs · 16 GB RAM. No hosted API sells it, so this network is the place to run it.

21 builds of Phi 4 MINI, 29 signed runs on real machines: different packagers, precisions and engines, each a different file with its own numbers.

154tok/s, fastest machine
$1.66per M tokens · no API sells it
128Kcontext proved by recall
2.3 GBmemory, smallest build
5providers measured it
Use it in your own room Ask Evo in your gramx and it runs on a machine that proved it, at the metered $1.66 per million tokens. On your own machines it costs you nothing.Measure it on your machine A gram works through the catalog by itself and signs every result. 5 providers have measured Phi 4 MINI; every run is a point on the providers board.Compare every build Put it beside every other model in one table: speed, price, context and the API price, one row each.
best on ollama · 20 builds

trained 128K · proved 128K context, by maziyarpanahi Q3_K_L

154.3 tok/s peak, by maziyarpanahi Q2_K

best on vllm · 1 build

proved 4K context, by microsoft BF16

Speed against price, every machine

Up and left is faster and cheaper · the line joins machines no other beats on both · price on a log scale

$10$100$1000USD per million tokenstok/sWindows · NVIDIA 12 GB · 16 CPUs · 16 GB RAM · 2-bit by maziyarpanahi · 154.35 tok/s · $1.6602/MWindows · NVIDIA 12 GB · 16 CPUs · 16 GB RAM · 3-bit by maziyarpanahi · 134.8 tok/s · $1.901/MWindows · NVIDIA 12 GB · 16 CPUs · 16 GB RAM · 3-bit by maziyarpanahi · 110.09 tok/s · $2.3277/MWindows · NVIDIA 12 GB · 16 CPUs · 16 GB RAM · 4-bit by maziyarpanahi · 101.89 tok/s · $2.515/MWindows · NVIDIA 12 GB · 16 CPUs · 16 GB RAM · 4-bit by maziyarpanahi · 77.39 tok/s · $3.3111/MLinux · NVIDIA 15 GB · 4 CPUs · 31 GB RAM · 4-bit by unsloth · 74.2 tok/s · $1.8399/MLinux · NVIDIA 15 GB · 4 CPUs · 31 GB RAM · 2-bit by unsloth · 68.61 tok/s · $1.9806/MLinux · NVIDIA 15 GB · 4 CPUs · 31 GB RAM · 3-bit by unsloth · 66.57 tok/s · $2.0463/MLinux · NVIDIA 15 GB · 4 CPUs · 31 GB RAM · 3-bit by maziyarpanahi · 64.06 tok/s · $2.1009/MWindows · NVIDIA 12 GB · 16 CPUs · 16 GB RAM · 8-bit by maziyarpanahi · 39.21 tok/s · $6.5351/MMac · Apple silicon · 10 CPUs · 16 GB RAM · 4-bit by bartowski · 36.2 tok/s · $5.4463/MMac · Apple silicon · 10 CPUs · 16 GB RAM · 4-bit by bartowski · 33.74 tok/s · $5.843/MMac · Apple silicon · 10 CPUs · 16 GB RAM · 4-bit by bartowski · 33.7 tok/s · $5.8504/MMac · Apple silicon · 10 CPUs · 16 GB RAM · 2-bit by bartowski · 27.09 tok/s · $7.2781/MWindows · no GPU · 8 CPUs · 16 GB RAM · 2-bit by mungert · 4.41 tok/s · $16.1008/MMac · no GPU · 4 CPUs · 8 GB RAM · 3-bit by maziyarpanahi · 2.2 tok/s · $13.851/MMac · no GPU · 4 CPUs · 8 GB RAM · 4-bit by unsloth · 2.18 tok/s · $14.8637/MMac · no GPU · 4 CPUs · 8 GB RAM · 3-bit by maziyarpanahi · 2.17 tok/s · $14.0402/MMac · no GPU · 4 CPUs · 8 GB RAM · 2-bit by bartowski · 1.06 tok/s · $30.3882/MMac · Apple silicon · 10 CPUs · 16 GB RAM · 3-bit by bartowski · 0.69 tok/s · $286.1754/MWindows · NVIDIA 12 GB · 16 CPUs · 16 GB RAM · 16-bit (full precision) by unsloth · 0.08 tok/s · $3352.4625/M

Same build, different machines

Q3_K_L by maziyarpanahi on 3 machines: the fastest decodes 50× as fast as the slowest. One dot per machine, tokens per second.

  1. NVIDIA GeForce RTX 3080 Ti×1110 tok/s
  2. Tesla T4×164 tok/s
  3. apple×12.2 tok/s

What it fits on

The smallest build needs about 2.3 GB; 13 of 13 kinds of machine on the network can hold it. Bars: each kind's memory (its card's, else RAM) · ticks: every build, at the size it held plus a fifth for the prompt.

  1. NVIDIA GeForce RTX 3050 Laptop GPU4 GB on the card6 of 13
  2. NVIDIA GeForce GTX 16504 GB on the card6 of 13
  3. AMD Radeon(TM) Vega 8 Mobile Graphics7 GB RAM6 of 13
  4. Windows, CPU only · 12 CPUs7 GB RAM6 of 13
  5. apple8 GB RAM6 of 13
  6. Apple M18 GB RAM6 of 13
  7. NVIDIA GeForce RTX 3080 Ti12 GB on the card7 of 13
  8. Tesla T415 GB on the card7 of 13
  9. Intel(R) UHD Graphics 73016 GB RAM7 of 13
  10. Intel(R) Iris(R) Xe Graphics16 GB RAM7 of 13
  11. Apple M416 GB RAM7 of 13
  12. 2× Tesla T429 GB on the card12 of 13
  13. Linux, CPU only · 4 CPUs31 GB RAMevery build

How far back it remembers, machine by machine

Filled: the needle came back at that prompt size · outlined: asked and lost · blank: not asked · trained for 128K. 1 walk ended because the model forgot, 1 because the machine could not hold the prompt.

machine · build4K8K16K32K64K128Kwhy it stopped
Gram ·a5be46Tesla T4 · Q4_K_M by unsloth128K recalled at every size asked
Gram ·a5be46Tesla T4 · Q5_K_M by unsloth128K recalled at every size asked
Gram ·a5be46Tesla T4 · Q3_K_M by unsloth128K recalled at every size asked
Gram ·a5be46Tesla T4 · Q3_K_L by maziyarpanahi128K recalled at every size asked
Gram ·a5be46Tesla T4 · Q8_0 by unsloth128K recalled at every size asked
Gram ·a5be46Tesla T4 · BF16 by unsloth128K recalled at every size asked
nitin singhApple M4 · Q3_K_L by bartowski8K
nitin singhApple M4 · Q4_K_M by bartowski4K
nitin singhApple M4 · IQ2_M by bartowski4K
nitin singhApple M4 · IQ4_XS by bartowski4K
nitin singhApple M4 · Q4_0 by bartowski4K
nitin singhApple M4 · Q3_K_S by maziyarpanahi4K
nitin singhApple M4 · IQ3_XXS by bartowski4K
Gram ·a5be46Tesla T4 · BF16 by microsoft4K this machine could not serve at 8K

The 14 longest of 27 walks; every one is on its build's page.

Carbon per million tokens

Lowest build on each machine · 1 of 1 read off a power sensor, the rest modelled from the parts · fewer grams is better

  1. Gram ·a5be46333 g2.54 J/tok · sensor

Carbon is location-based: signed energy × the yearly average of the grid where the machine signed its place (world grid average (place not signed); Ember Global Electricity Review 2025 (2024 generation), location-based). Facility overhead is not included. A machine on its own solar is not credited until it can prove it.

Measured by

The people whose machines signed these numbers

proved by at least one build: Structured output · Sustained output · Writes code · Reasons about code · Synthesises sources · Follows instructions

Every build

Each its own file with its own page: the machines that ran it and their signatures.

buildenginewhose workmemorycontext provedbest tok/smachines
Q3_K_L by maziyarpanahihf.co/MaziyarPanahi/Phi-4-mini-instruct-GGUF:Q3_K_Lollamamaziyarpanahi24.6 GB128K110.13
Q4_K_M by unslothhf.co/unsloth/Phi-4-mini-instruct-GGUF:Q4_K_Mollamaquantizer · unsloth24.9 GB128K75.33
Q5_K_M by unslothhf.co/unsloth/Phi-4-mini-instruct-GGUF:Q5_K_Mollamaquantizer · unsloth25.3 GB128K68.71
Q3_K_M by unslothhf.co/unsloth/Phi-4-mini-instruct-GGUF:Q3_K_Mollamaquantizer · unsloth24.5 GB128K66.91
Q8_0 by unslothhf.co/unsloth/Phi-4-mini-instruct-GGUF:Q8_0ollamaquantizer · unsloth26.7 GB128K53.71
BF16 by unslothhf.co/unsloth/Phi-4-mini-instruct-GGUF:BF16ollamaquantizer · unsloth29.2 GB128K32.02
Q3_K_L by bartowskihf.co/bartowski/microsoft_Phi-4-mini-instruct-GGUF:Q3_K_Lollamaquantizer · bartowski—8K32.11
Q3_K_S by maziyarpanahihf.co/MaziyarPanahi/Phi-4-mini-instruct-GGUF:Q3_K_Sollamamaziyarpanahi—4K134.83
Q4_K_M by bartowskihf.co/bartowski/microsoft_Phi-4-mini-instruct-GGUF:Q4_K_Mollamaquantizer · bartowski3.4 GB4K40.61
IQ2_M by bartowskihf.co/bartowski/microsoft_Phi-4-mini-instruct-GGUF:IQ2_Mollamaquantizer · bartowski2.3 GB4K37.32
IQ4_XS by bartowskihf.co/bartowski/microsoft_Phi-4-mini-instruct-GGUF:IQ4_XSollamaquantizer · bartowski3.2 GB4K36.21
Q4_0 by bartowskihf.co/bartowski/microsoft_Phi-4-mini-instruct-GGUF:Q4_0ollamaquantizer · bartowski3.3 GB4K33.71
IQ3_XXS by bartowskihf.co/bartowski/microsoft_Phi-4-mini-instruct-GGUF:IQ3_XXSollamaquantizer · bartowski2.5 GB4K0.71
BF16 by microsoftmicrosoft/Phi-4-mini-instructvllmvendor · microsoft8.6 GB4K—1
Q2_K by maziyarpanahihf.co/MaziyarPanahi/Phi-4-mini-instruct-GGUF:Q2_Kollamamaziyarpanahi——154.31
Q4_K_M by maziyarpanahihf.co/MaziyarPanahi/Phi-4-mini-instruct-GGUF:Q4_K_Mollamamaziyarpanahi——101.91
Q4_K_S by maziyarpanahihf.co/MaziyarPanahi/Phi-4-mini-instruct-GGUF:Q4_K_Sollamamaziyarpanahi——77.41
Q2_K by unslothhf.co/unsloth/Phi-4-mini-instruct-GGUF:Q2_Kollamaquantizer · unsloth2.6 GB—71.01
Q8_0 by maziyarpanahihf.co/MaziyarPanahi/Phi-4-mini-instruct-GGUF:Q8_0ollamamaziyarpanahi——39.21
Q3_K_S by bartowskihf.co/bartowski/microsoft_Phi-4-mini-instruct-GGUF:Q3_K_Sollamaquantizer · bartowski——28.21
IQ2_XS by mungerthf.co/Mungert/Phi-4-mini-instruct.gguf:IQ2_XSollamamungert——8.51
derivatives — somebody else's change to these weights

Fine-tunes, merges and abliterations of this member, each its own model with its own page. None of their numbers are counted above.

phi4-clinical-mlxnot measuredPhi-4-mini-instructnot measured