PowerMoE 3B — Hardware Requirements & GPU Compatibility
Chat
PowerMoE-3B is IBM Research's 3.4-billion-parameter sparse Mixture-of-Experts language model, built for general text generation, reasoning, and code rather than a specific vision or audio modality. It routes each token to 8 of 40 experts, activating roughly 800 million parameters per token according to IBM, keeping decoding fast even though the full set of experts still has to be held in memory. At this size it fits comfortably on a single mainstream consumer GPU once quantized. Context length is limited to 4,096 tokens. It is released under the Apache 2.0 license, permitting unrestricted commercial and research use, and was published in August 2024. Its defining feature is the Power learning-rate scheduler IBM used to train it, which the team reports lets an 800-million-active-parameter model match dense models with roughly twice as many active parameters.
Runs great on a 24 GB GPU
Runs from 8 GB (Q4_K_M · 4K)Best on 24 GB (Q4_K_M · ~354 tok/s · 4K)Macs from 8 GB (Q4_K_M · ~62 tok/s)
Apache 2.0
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run PowerMoE 3B
One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.
Recommended
ollama run hf.co/mradermacher/PowerMoE-3b-GGUF:Q4_K_MQ4_K_M fits fully on the RTX 4090 · 21.4 GB free · 4K context
Download: PowerMoE-3b.Q4_K_M.gguf2.06 GBmradermacher
Ways to runOther ways
each opens step-by-stepRun in cloud
Runs on RTX 3060 12GB (9 GB headroom) · Q4_K_M
- Generation speed
- ~211 tok/s
- generation speed
- Cost per 1M output tokens
- $0.081
- per 1M output tokens
How Much VRAM Does PowerMoE 3B Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 1.9 GB | — | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Fits: 4K context · ~389 tok/s | |
| 1.9 GB | — | 3-bit small — 3-bit small quantization | Fits: 4K context · ~386 tok/s | |
| 2.1 GB | — | 3-bit medium — 3-bit medium quantization | Fits: 4K context · ~376 tok/s | |
| 2.1 GB | — | 4-bit · legacy — 4-bit legacy quantization | Fits: 4K context · ~373 tok/s | |
| 2.5 GB | — | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Fits: 4K context · ~354 tok/s | |
| 2.8 GB | — | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Fits: 4K context · ~334 tok/s | |
| 3.2 GB | — | 6-bit, very good — 6-bit quantization, very good quality | Fits: 4K context · ~317 tok/s | |
| 3.8 GB | — | 8-bit, near-lossless — 8-bit quantization, near-lossless | Fits: 4K context · ~293 tok/s |
Selected Q4_K_M
4-bit medium quantization2.6GBat 4K context
- Weights
- 2.0 GB
- KV cache
- 0.3 GB
- Overhead
- 0.3 GB
Fits RTX 4090 (reference) · 21.4 GB free · ~354 tok/s
≥50 tok/s needs ≥59 GB/s · RTX 5090 ~424, RTX 3090 Ti ~354
Which Hardware Can Run PowerMoE 3B?
Q4_K_M · 2.5 GBPowerMoE 3B (Q4_K_M) requires 2.5 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 4+ GB is recommended. Using the full 4K context window can add up to 0.1 GB, bringing total usage to 2.6 GB. 52 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti. 163 devices with unified memory can run PowerMoE 3B, including NVIDIA DGX Spark, NVIDIA Jetson AGX Thor Developer Kit.
≤ 11 GB
≤ 11 GB tier →fits · 5.4 GB free · up to 4K ctx
16 GPUs · 1 Mac class · 2 device families
RTX 3080~315Simulate output on RTX 3080RTX 2080 Ti~285Simulate output on RTX 2080 TiRTX 3070 Ti~284Simulate output on RTX 3070 TiM3*8 GB~62Simulate output on iMac 24" M3 (8 GB)Orin Nano~81Simulate output on NVIDIA Jetson Orin Nano 8GB (Super)iPhone~48Simulate output on iPhone 1720 GB
fits · 17.4 GB free · up to 4K ctx
1 GPU
RX 7900 XT~127Simulate output on RX 7900 XT32–40 GB
fits · 29.4 GB free · up to 4K ctx
1 GPU · 3 Mac classes · 4 device families
RTX 5090~424Simulate output on RTX 5090M5 Max36 GB~117Simulate output on Mac Studio M5 Max (36 GB)M532 GB~78Simulate output on MacBook Air 13" M5 (32 GB)Orin~143Simulate output on NVIDIA Jetson AGX Orin 32GBSnapdragon X Elite (X1E-80-100)~90Simulate output on Snapdragon X Elite Copilot+ PC
By VRAM tier · typical GPUs and Macs at Q4_K_M
Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| Granite 4.0 MicroIBM | — | 3.7 GB | ~261 | same size |
| Deepseek Vl2 TinyDeepSeek | — | 9.4 GB | ~152 | same size |
| Llama 3.2 3B InstructMeta | — | 4.1 GB | ~309 | same size |
| Tiny Aya GlobalCohere | — | 9.4 GB | ~89 | same size |
| AI21 Jamba Reasoning 3BAI21 Labs | — | 2.3 GB | ~291 | same size |
| MiniMax M3 EAGLE3Inferact | — | 2.7 GB | ~284 | same size |
Specifications
- Publisher
- ibm-research
- Parameters
- 3.4B
- Active parameters
- 1B per token
- Architecture
- GraniteMoeForCausalLM
- Context Length
- 4,096 tokens
- Vocabulary Size
- 49,152
- Release Date
- 2024-08-14
- License
- Apache 2.0
Frequently Asked Questions
How much VRAM does PowerMoE 3B need?
PowerMoE 3B requires 2.5 GB of VRAM at Q4_K_M, or 7.2 GB at BF16.
VRAM = Weights + KV Cache + Overhead
Weights = 3.4B × 4.8 bits ÷ 8 = 2 GB
KV Cache + Overhead ≈ 0.5 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
KV Cache + Overhead ≈ 0.6 GB (at full 4K context)
VRAM usage by quantization
What's the best quantization for PowerMoE 3B?
For PowerMoE 3B, Q4_K_M (2.5 GB) offers the best balance of quality and VRAM usage. Q5_K_S (2.8 GB) provides better quality if you have the VRAM. The smallest option is IQ3_XXS at 1.7 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run PowerMoE 3B on a Mac?
Yes — MacBook Air 13" M3 (8 GB) and 145 other Macs can run PowerMoE 3B. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 2.5 GB of usable unified memory (RAM minus macOS overhead).
Can I run PowerMoE 3B locally?
Yes — PowerMoE 3B can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is PowerMoE 3B?
At Q4_K_M, PowerMoE 3B can reach ~153 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~354 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)
Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.
Example: NVIDIA B200 → 0.7 GB active ÷ (8000 × 0.65) = 0.134 ms, plus 32 layers × 0.055 ms = 1.760 ms, so 1000 ÷ 1.894 ms = ~528 tok/s
Estimated speed at Q4_K_M (2.5 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of PowerMoE 3B?
At Q4_K_M, the download is about 2.02 GB. The full-precision BF16 version is 6.75 GB. The smallest option (IQ3_XXS) is 1.31 GB.
Which GPUs can run PowerMoE 3B?
52 consumer GPUs can run PowerMoE 3B at Q4_K_M. Top options include AMD Radeon RX 6700 XT, AMD Radeon RX 6800, AMD Radeon RX 6800 XT. 52 GPUs have plenty of headroom for comfortable inference.
Which devices can run PowerMoE 3B?
163 devices with unified memory can run PowerMoE 3B at Q4_K_M (2.5 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.