FITMYLLM · JULY 28, 2026 · OPEN METHODOLOGY · COMMUNITY BENCHMARKS
LIVE · UPDATED DAILY
41.9B
Parameters (6.6B active)
Quantization Options Select your GPU for speed estimates Context length: 4K 8K 16K 32K 64K 128K
Quant Bits VRAM @ 16K Quality IQ3_XXS 3.25 19.0 GB
17.5 + 1.5 KV
low IQ3_XS 3.5 20.3 GB
18.8 + 1.5 KV
low Q3_K_S 3.64 21.1 GB
19.6 + 1.5 KV
low IQ3_M 3.76 21.7 GB
20.2 + 1.5 KV
low Q3_K_M 4 22.9 GB
21.4 + 1.5 KV
low Q3_K_L 4.3 24.5 GB
23.0 + 1.5 KV
moderate IQ4_XS 4.46 25.3 GB
23.8 + 1.5 KV
moderate Q4_K_S 4.67 26.4 GB
24.9 + 1.5 KV
moderate Q4_K_M 4.89 27.6 GB
26.1 + 1.5 KV
good Q5_K_S 5.57 31.2 GB
29.7 + 1.5 KV
good Q5_K_M 5.7 31.8 GB
30.3 + 1.5 KV
good Q6_K 6.56 36.3 GB
34.8 + 1.5 KV
excellent Q8_0 8.5 46.5 GB
45.0 + 1.5 KV
lossless FP16 16 85.8 GB
84.3 + 1.5 KV
lossless
Select your GPU above to see speed estimates and compatibility for each quantization.
Deploying for a team or in production? Size GPUs, cost & scaling in Enterprise → ▸ READY TO RUN THIS? RENT BY THE HOUR
RENT A GPU AND RUN PHI-3.5 MOE 42B NOW
Spin up an A100 / H100 / 4090 in ~60s. Pay by the second. Cancel anytime.
Community Ratings Chat Coding Reasoning Creative Vision Roleplay Agentic
Loading ratings...
Run this model IQ3_XXS — 17.5 GB VRAM IQ3_XS — 18.8 GB VRAM Q3_K_S — 19.6 GB VRAM IQ3_M — 20.2 GB VRAM Q3_K_M — 21.4 GB VRAM Q3_K_L — 23.0 GB VRAM IQ4_XS — 23.8 GB VRAM Q4_K_S — 24.9 GB VRAM Q4_K_M — 26.1 GB VRAM Q5_K_S — 29.7 GB VRAM Q5_K_M — 30.3 GB VRAM Q6_K — 34.8 GB VRAM Q8_0 — 45.0 GB VRAM FP16 — 84.3 GB VRAM
Ollama llama.cpp vLLM LM Studio KoboldCpp Jan Docker
▸ Easiest way to get started · Beginners
DOCS ↗ curl -fsSL https://ollama.com/install.sh | shCOPY
$ ollama run phi:42b-q4_K_MCOPY
Tag may need adjustment — check ollama.com/library/phi for available tags.
▸ SETUP GUIDE >_
Auto-setup with fitmyllm CLI Detects your GPU, recommends the best model, downloads it, and starts chatting — zero config. Benchmarks your speed and contributes anonymous data to improve predictions.
Auto-detect GPU Live tok/s in chat Speed benchmarks 9 inference engines
GPUs that can run this model At Q4_K_M quantization. Sorted by minimum VRAM.
Find the best GPU for Phi-3.5 MoE 42B
Build Hardware for Phi-3.5 MoE 42B Phi-3.5 MoE — runs at 7B speed but with 42B total capacity. Best of both worlds.
Read full model card ▸ COLOPHON FITMYLLM · INDEPENDENT · DATA-DRIVEN
FITMYLLM · EST. 2025 · © 2026
RECOMMENDATIONS FROM PUBLISHED MATH, CORRECTED BY THE COMMUNITY — 30.
▸ SPEC SHEET
Phi-3.5 MoE 42B — 41.9B MoE. ▸ SPECIFICATIONS
PARAMETERS 41.9B (6.6B active)
ARCHITECTURE Mixture of Experts
CONTEXT LENGTH 128K tokens
CAPABILITIES chat, coding, reasoning
RELEASE DATE 2024-08-20
PROVIDER Microsoft
FAMILY phi ▸ VRAM REQUIREMENTS
QUANT BPW VRAM QUALITY IQ3_XXS 3.25 17.5 GB 82% IQ3_XS 3.5 18.8 GB 84% Q3_K_S 3.64 19.6 GB 85% IQ3_M 3.76 20.2 GB 86% Q3_K_M 4 21.4 GB 88% Q3_K_L 4.3 23.0 GB 90% IQ4_XS 4.46 23.8 GB 92% Q4_K_S 4.67 24.9 GB 93% Q4_K_M 4.89 26.1 GB 94% Q5_K_S 5.57 29.7 GB 96% Q5_K_M 5.7 30.3 GB 96% Q6_K 6.56 34.8 GB 97% Q8_0 8.5 45.0 GB 100% FP16 16 84.3 GB 100%
§ 01 BENCHMARK SCORES
HumanEval 77.0
MMLU-PRO 60.8
MATH 66.0
IFEval 79.0
BBH 72.0
GPQA 43.0
MUSR 17.3
BigCodeBench 38.2
§ 03 COMPATIBLE GPUs
30 @ Q4_K_M Feedback 