family member
Phi 4 logo

Phi 4 MINI-REASONING

189 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.

10 builds of Phi 4 MINI-REASONING, 35 signed runs on real machines: different packagers, precisions and engines, each a different file with its own numbers.

189tok/s, fastest machine
$1.36per M tokens · no API sells it
4Kcontext proved by recall
1.3 GBmemory, smallest build
8providers measured it
best on ollama · 10 builds

proved 4K context, by unsloth IQ3_XXS

189.4 tok/s peak, by unsloth IQ3_XXS

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

$1$10USD per million tokenstok/sWindows · NVIDIA 12 GB · 16 CPUs · 16 GB RAM · IQ3_XXS (3-bit) by unsloth · 189.43 tok/s · $1.3595/MWindows · NVIDIA 12 GB · 16 CPUs · 16 GB RAM · Q3_K_L (3-bit) by lmstudio-community · 89.92 tok/s · $2.8497/MWindows · NVIDIA 4 GB · 12 CPUs · 16 GB RAM · IQ3_XXS (3-bit) by unsloth · 41.79 tok/s · $5.1926/MMac · Apple silicon · 10 CPUs · 16 GB RAM · IQ3_XXS (3-bit) by unsloth · 25.64 tok/s · $7.6718/MMac · Apple silicon · 10 CPUs · 16 GB RAM · Q8_0 (8-bit) by lmstudio-community · 22.03 tok/s · $8.8724/MMac · Apple silicon · 10 CPUs · 16 GB RAM · IQ1_S (1-bit) by unsloth · 17.18 tok/s · $11.4266/MWindows · NVIDIA 4 GB · 12 CPUs · 16 GB RAM · IQ1_S (1-bit) by unsloth · 16.34 tok/s · $13.2553/MWindows · no GPU · 8 CPUs · 16 GB RAM · Q2_K (2-bit) by unsloth · 9.83 tok/s · $7.3078/MMac · Apple silicon · 8 CPUs · 8 GB RAM · Q4_K_M (4-bit) by lmstudio-community · 7.26 tok/s · $24.1396/MWindows · no GPU · 8 CPUs · 16 GB RAM · Q4_K_M (4-bit) by unsloth · 6.5 tok/s · $11.221/MWindows · Intel 2 GB · 8 CPUs · 16 GB RAM · IQ1_S (1-bit) by unsloth · 5.6 tok/s · $12.8228/MWindows · AMD 1 GB · 8 CPUs · 7 GB RAM · Q4_K_M (4-bit) by lmstudio-community · 4.75 tok/s · $36.8345/MWindows · AMD 1 GB · 8 CPUs · 7 GB RAM · Q3_K_L (3-bit) by lmstudio-community · 3.85 tok/s · $45.5148/MMac · Apple silicon · 8 CPUs · 8 GB RAM · Q3_K_L (3-bit) by lmstudio-community · 3.42 tok/s · $51.1518/MMac · no GPU · 4 CPUs · 8 GB RAM · Q4_K_M (4-bit) by unsloth · 1.85 tok/s · $17.4407/M

Same build, different machines

IQ1_S by unsloth on 6 machines: the fastest decodes 6.9× as fast as the slowest. One dot per machine, tokens per second.

  1. Apple M4×128 tok/s
  2. NVIDIA GeForce RTX 3050 Laptop GPU×120 tok/s
  3. Mac, CPU only · 8 CPUs×216 tok/s
  4. Intel(R) UHD Graphics 730×15.6 tok/s
  5. Intel(R) Iris(R) Xe Graphics×14.0 tok/s

What it fits on

The smallest build needs about 1.3 GB; 16 of 17 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. Linux, CPU only · 2 CPUs1 GB RAMnone
  2. NVIDIA GeForce RTX 3050 Laptop GPU4 GB on the card2 of 4
  3. NVIDIA GeForce GTX 16504 GB on the card2 of 4
  4. AMD Radeon(TM) Vega 8 Mobile Graphics7 GB RAM3 of 4
  5. Windows, CPU only · 12 CPUs7 GB RAM3 of 4
  6. Apple silicon8 GB RAM3 of 4
  7. Mac, CPU only · 8 CPUs8 GB RAM3 of 4
  8. NVIDIA GeForce RTX 3080 Ti12 GB on the cardevery build
  9. Tesla T415 GB on the cardevery build
  10. Intel(R) UHD Graphics 73016 GB RAMevery build
  11. Intel(R) Iris(R) Xe Graphics16 GB RAMevery build
  12. Apple M416 GB RAMevery build
  13. Apple M116 GB RAMevery build
  14. NVIDIA L422 GB on the cardevery build
  15. 2× Tesla T429 GB on the cardevery build
  16. Linux, CPU only · 4 CPUs31 GB RAMevery build
  17. NVIDIA RTX PRO 6000 Blackwell Server Edition95 GB on the cardevery 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. 0 walks ended because the model forgot, 0 because the machine could not hold the prompt.

machine · build4K8K16K32K64K128Kwhy it stopped
Kunal HindochaNVIDIA GeForce RTX 3080 Ti · IQ3_XXS by unsloth4K an older probe: where it stopped was not recorded
nitin singhApple M4 · IQ3_XXS by unsloth4K an older probe: where it stopped was not recorded
nitin singhApple M4 · Q8_0 by lmstudio-community4K an older probe: where it stopped was not recorded
Kunal HindochaNVIDIA GeForce RTX 3080 Ti · Q3_K_L by lmstudio-community— an older probe: where it stopped was not recorded
Gram ·53b229NVIDIA GeForce RTX 3050 Laptop GPU · IQ3_XXS by unsloth— an older probe: where it stopped was not recorded
nitin singhApple M4 · Q2_K by unsloth— an older probe: where it stopped was not recorded
nitin singhApple M4 · Q3_K_M by unsloth— an older probe: where it stopped was not recorded
nitin singhApple M4 · IQ1_S by unsloth— an older probe: where it stopped was not recorded
Gram ·53b229NVIDIA GeForce RTX 3050 Laptop GPU · IQ1_S by unsloth— an older probe: where it stopped was not recorded
Subrata MohapatraMac, CPU only · 8 CPUs · IQ1_S by unsloth— an older probe: where it stopped was not recorded
Amit KhetanIntel(R) UHD Graphics 730 · Q2_K by unsloth— an older probe: where it stopped was not recorded
Gram ·0548a8Mac, CPU only · 8 CPUs · Q4_K_M by lmstudio-community— an older probe: where it stopped was not recorded
Amit KhetanIntel(R) UHD Graphics 730 · Q2_K_L by unsloth— an older probe: where it stopped was not recorded
Amit KhetanIntel(R) UHD Graphics 730 · Q4_K_M by unsloth— an older probe: where it stopped was not recorded

The 14 longest of 30 walks; every one is on its build's page.

proved by at least one build: Structured output · Sustained output · Extended thinking · Reasons about code · Synthesises sources

Every build

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

buildenginewhose workmemorycontext provedbest tok/smachines
IQ3_XXS by unslothhf.co/unsloth/Phi-4-mini-reasoning-GGUF:IQ3_XXSollamaquantizer · unsloth11.5 GB4K189.46
Q8_0 by lmstudio-communityhf.co/lmstudio-community/Phi-4-mini-reasoning-GGUF:Q8_0ollamalmstudio-community5.2 GB4K22.21
IQ1_S by unslothhf.co/unsloth/Phi-4-mini-reasoning-GGUF:IQ1_Sollamaquantizer · unsloth1.3 GB—188.99
Q3_K_L by lmstudio-communityhf.co/lmstudio-community/Phi-4-mini-reasoning-GGUF:Q3_K_Lollamalmstudio-community——89.94
Q2_K by unslothhf.co/unsloth/Phi-4-mini-reasoning-GGUF:Q2_Kollamaquantizer · unsloth——37.73
Q3_K_M by unslothhf.co/unsloth/Phi-4-mini-reasoning-GGUF:Q3_K_Mollamaquantizer · unsloth——31.94
Q4_K_M by lmstudio-communityhf.co/lmstudio-community/Phi-4-mini-reasoning-GGUF:Q4_K_Mollamalmstudio-community——7.32
Q2_K_L by unslothhf.co/unsloth/Phi-4-mini-reasoning-GGUF:Q2_K_Lollamaquantizer · unsloth——6.83
Q4_K_M by unslothhf.co/unsloth/Phi-4-mini-reasoning-GGUF:Q4_K_Mollamaquantizer · unsloth2.8 GB—6.52
Q6_K by lmstudio-communityhf.co/lmstudio-community/Phi-4-mini-reasoning-GGUF:Q6_Kollamalmstudio-community———1