ibm-research / text-generation updated 1 year ago

PowerMoE-3b

Name: PowerMoE-3b
Author: ibm-research

PowerMoE-3B is a 3B sparse Mixture-of-Experts (sMoE) language model trained with the Power learning rate scheduler. It sparsely activates 800M parameters for each token. It is trained on a mix of open-source and proprietary datasets. PowerMoE-3B has shown promising results compared to other dense models with 2x activate parameters across various benchmarks, including natural language multi-choices, code generation, a...

Params

3.4 B

Context

4,096

Downloads 30d

1.8 M

Likes

Commercial use: allowed apache-2.0 Not gated SAFETENSORS View on Hugging Face ↗

Download history

daily snapshots · 10 days

1.8 M1.8 M

Jul 28Jul 31Aug 3Aug 6

Can you run it?

Estimated VRAM at 8K context unless noted. Pick your hardware to see the verdict per quantization.

File	Quant	Size	Est. VRAM	Verdict on RTX 4090 · 24 GB
model.safetensors	f32	13.5 GB	15.9 GB	✅ Runs comfortably

Estimate: file size × 1.1 + KV cache at 8K + 0.5 GB overhead. Not a benchmark — how we calculate this.

Run it

copy-paste, exact tags checked against the Hub

~ · curl · api/v1

$ curl -s https://aimodelscomparison.com/api/v1/models/powermoe-3b
{
  "hf_id": "ibm-research/PowerMoE-3b",
  "params_b": 3.37,
  "context_length": 4096,
  "license": { "id": "apache-2.0", "commercial": "yes" },
  "downloads_30d": 1843751,
  "vram_estimates": [
    { "quant": "f32", "gb": 15.9 }
  ],
  "updated_at": "2026-07-28T18:02:45Z"
}

est. VRAM —on RTX 4090 · 24 GBJSON API →

Specifications

Architecture: GraniteMoeForCausalLM
Parameters: 3.4 B
Tensor type: F32
Context length: 4,096
Vocabulary: 49,152
Layers / heads: 32 / 24
Licence: apache-2.0
First seen on the Hub: 2024-08-14
Training datasets: undisclosed
ARC (reported): 58.1
MBPP (reported): 32.4
PIQA (reported): 79.1
BoolQ (reported): 65
Hellaswag (reported): 71.5
humaneval (reported): 20.1
OpenBookQA (reported): 41
Winogrande (reported): 65
MMLU (5 shot) (reported): 42.8
math (4 shot) (reported): 14.8
GSM8k (5 shot) (reported): 25.9
Added to our catalog: 2026-07-28

Compare with

Qwen3-8B 16.3 M Qwen2.5-1.5B-Instruct 14.0 M Qwen2.5-7B-Instruct 12.1 M Qwen3-1.7B 7.8 M Qwen2.5-3B-Instruct 6.0 M Qwen3-4B 4.6 M

Appears in

Best open coding models Commercial-use text models Models that fit in 12 GB