Phi-4 reasoning 14B

Microsoft Phi-4 reasoning 14B, 4-bit GGUF on llama.cpp. Chat in your browser, plus an OpenAI-compatible API.

Engine: llama.cpp · Q4_K_M
Size: 9.1 GB
GPU memory: 11 GiB or more
Runs on

Pick a machine

CheapestBalancedFastest

Max

Picked Apple M4 Pro (20-core GPU): the lowest price that fits. No speed data for these GPUs yet.

3 machines can run this now · all new hosts1× Apple M4 Pro (20-core GPU)$0.12/hRecommended48 GB unified · 36 GB for AI 🇨🇭 Switzerland Dedicated 1× Apple M3 Max (30-core GPU)$0.13/h36 GB unified · 28 GB for AI 🇮🇳 India Dedicated 1× Apple M5 Max (40-core GPU)$0.26/h48 GB unified · 37 GB for AI 🇺🇸 United States Dedicated

On 1× Apple M4 Pro (20-core GPU). Runs 1 h, then stops · holds $0.115 · incl. 10% fee

Deploying accepts the rental terms. Unused budget comes back to your balance.

Customize

Sort by

GPU maker

GPU model

GPUs

VRAM per GPU

Region

Reliability

Availability

Disk

Verified hosts only

Options

Duration limit1 h

Duration in hoursHard cap: the machine stops at this limit even if you forget. Snaps to 30-minute steps up to a day, whole days after.

Budget cap

Held when it starts, never exceeded. Unused budget comes back.

Add SSH public keyOptional. You chat in the browser without one. Add it to also connect with SSH.

Hardware that can run this

M4 Pro · 24 GB unifiedQ4_K_M, Q8 · est. 18–26 tok/sDeployable now · $0.12/h

19 more devices could run it, none online

Hardware	Fits	Est. speed
RTX 3080 Ti · 12 GB	Q4_K_M	60–86 tok/s
M1 · 16 GB unified	Q4_K_M	Evidence unavailable: no cited memory bandwidth for this hardware
M1 Pro · 16 GB unified	Q4_K_M	13–19 tok/s
M2 · 16 GB unified	Q4_K_M	6–10 tok/s
M2 Pro · 16 GB unified	Q4_K_M	13–19 tok/s
M3 · 16 GB unified	Q4_K_M	6–10 tok/s
M4 · 16 GB unified	Q4_K_M	7–12 tok/s
RX 6800 XT · 16 GB	Q4_K_M	Evidence unavailable: no cited memory bandwidth for this hardware
M3 Pro · 18 GB unified	Q4_K_M, Q8	9–15 tok/s
RTX 3090 · 24 GB	Q4_K_M, Q8	62–88 tok/s
RTX 4090 · 24 GB	Q4_K_M, Q8	66–95 tok/s
RX 7900 XTX · 24 GB	Q4_K_M, Q8	Evidence unavailable: no cited memory bandwidth for this hardware
M1 Max · 32 GB unified	Q4_K_M, Q8	26–38 tok/s
M2 Max · 32 GB unified	Q4_K_M, Q8	26–38 tok/s
M3 Max · 48 GB unified	Q4_K_M, Q8, FP16	19–29 tok/s
M4 Max · 48 GB unified	Q4_K_M, Q8, FP16	27–39 tok/s
M4 Pro · 48 GB unified	Q4_K_M, Q8, FP16	18–26 tok/s
Instinct MI210 · 64 GB	Q4_K_M, Q8, FP16	Evidence unavailable: no cited memory bandwidth for this hardware
M4 Max · 64 GB unified	Q4_K_M, Q8, FP16	27–39 tok/s

Speeds are estimates from memory bandwidth and weight size, not measured benchmarks. A price is the cheapest listed offer for that hardware.

Template details

Rating: No ratings yet
Runs: 0
Min VRAM: 12 GB
Backend: Metal on Apple silicon
Weights: 9.1 GB
Version: llamacpp-phi-4-reasoning-v1.0.0
Image digest: 1f4b9cf58982…
Health check: SSH probe on the leased endpoint

We match GPUs on their measured runtime and dedicated VRAM. A hardware report doesn't qualify another vendor's backend.

#llm #chat #openai-api #gguf #reasoning

Reviews

No reviews yet.