SAVRN
Search Contact SAVRN

SAVRN Model Hub · Comparisons

PowerMoE-3b vs Qwen2.5-3B-Instruct

PowerMoE-3b has 3.4B parameters and Qwen2.5-3B-Instruct has 3.1B parameters; PowerMoE-3b is released under Apache License 2.0 and Qwen2.5-3B-Instruct under other; at 16-bit, PowerMoE-3b needs about 8.1 GB (1x MI300X from $1.85 an hour) and Qwen2.5-3B-Instruct about 7.4 GB (1x MI300X from $1.85 an hour).

Published metadata for 2 models, each read from its own repository.
Field PowerMoE-3b
ibm-research/PowerMoE-3b
Qwen2.5-3B-Instruct
Qwen/Qwen2.5-3B-Instruct
Publisher IBM Research Qwen
Task Text generation Text generation
Modality Text Text
Parameters, as reported 3.4B parameters 3.1B parameters
Architecture GraniteMoeForCausalLM Qwen2ForCausalLM
Library transformers transformers
Context length 4,096 tokens 32,768 tokens
Repository size 13.5 GB 6.2 GB
Artifact formats safetensors safetensors
License apache-2.0 other
Access Open weights, no gate Open weights, no gate
Memory at 16-bit (weights and margin) 8.1 GB 7.4 GB
Cheapest GPUs at 16-bit, per hour 1x MI300X, $1.85 1x MI300X, $1.85
Memory at 4-bit (weights and margin) 2 GB 1.9 GB
Cheapest GPUs at 4-bit, per hour 1x MI300X, $1.85 1x MI300X, $1.85
Revision viewed 13fcb5a98001 aa8e72537993
Downloads reported by the hub 1.3M 5.1M
Last observed 2026-09-18 2026-09-18

An evaluation row appears only where at least two of these models report the same benchmark with the same stated configuration, metric, unit and setup. Different evaluators stay named in each cell. Values are shown as reported: no unit conversion, no ranking.

Other Reported Results

These results are listed for each model on its own, because the conditions needed to compare them are not stated or do not match. Two results that leave a condition blank are not assumed to share it.

PowerMoE-3b

BenchmarkConditionsResultReported byRevisionDate
ARC Task text-generationMetric accuracy-normComparison conditions not established 58.1 ibm-research
Publisher reported
Evaluated revision not stated
BoolQ Task text-generationMetric accuracyComparison conditions not established 65 ibm-research
Publisher reported
Evaluated revision not stated
GSM8k (5 shot) Task text-generationMetric accuracyComparison conditions not established 25.9 ibm-research
Publisher reported
Evaluated revision not stated
Hellaswag Task text-generationMetric accuracy-normComparison conditions not established 71.5 ibm-research
Publisher reported
Evaluated revision not stated
MBPP Task text-generationMetric pass@1Comparison conditions not established 32.4 ibm-research
Publisher reported
Evaluated revision not stated
MMLU (5 shot) Task text-generationMetric accuracyComparison conditions not established 42.8 ibm-research
Publisher reported
Evaluated revision not stated
OpenBookQA Task text-generationMetric accuracy-normComparison conditions not established 41 ibm-research
Publisher reported
Evaluated revision not stated
PIQA Task text-generationMetric accuracy-normComparison conditions not established 79.1 ibm-research
Publisher reported
Evaluated revision not stated
Winogrande Task text-generationMetric accuracy-normComparison conditions not established 65 ibm-research
Publisher reported
Evaluated revision not stated
humaneval Task text-generationMetric pass@1Comparison conditions not established 20.1 ibm-research
Publisher reported
Evaluated revision not stated
math (4 shot) Task text-generationMetric accuracyComparison conditions not established 14.8 ibm-research
Publisher reported
Evaluated revision not stated

SAVRN's Notes on PowerMoE-3b

Forty experts, eight awake for each token: that is how IBM Research's PowerMoE-3b holds 3.4B parameters while activating 800M per token. Memory does not care which experts are asleep, though. All 40 must be resident, so 16-bit needs 8.1 GB, 8-bit 4.0 GB and 4-bit 2.0 GB. Each lands on one MI300X with 192 GB at $1.85 an hour on demand and leaves the card mostly empty, so share it. The download is 13.5 GB, twice the 16-bit footprint.

Apache 2.0 makes commercial use, modification and redistribution straightforward; keep the notices and record your changes. Two checks. The context window is 4,096 tokens, so long documents will not fit in one pass; this model is for short prompts, code and math. And the training mix includes proprietary datasets, with every published number reported by the publisher itself, so evaluate on your own workload first.

SAVRN's Notes on Qwen2.5-3B-Instruct

At 16-bit, 7.4 GB is all this one asks for. Qwen2.5-3B-Instruct carries 3.1 billion parameters and a 32,768-token window, aimed at instruction following, structured data, JSON output and long text past 8K tokens. The cheapest Index setup is one MI300X with 192 GB at $1.85 an hour; 8-bit needs 3.7 GB and 4-bit 1.9 GB. Nobody sizes a 192 GB card for a 7.4 GB model, so it shares the card, beside a larger model or as several copies.

The weights are open access, so nothing gates the files, but the license field reads 'other' and carries no summary; the terms live in the publisher's license file, and nothing about commercial use can be assumed until that file has been read. Then check the lineage: this build derives from the Qwen2.5-3B base, where any fine-tune would start, and its paper is the Qwen2 technical report, arXiv:2407.10671.

Questions

Which is larger, PowerMoE-3b or Qwen2.5-3B-Instruct?

PowerMoE-3b (3.4B parameters) is larger than Qwen2.5-3B-Instruct (3.1B parameters), by the parameter counts their publishers report.

Which is cheaper to run, PowerMoE-3b or Qwen2.5-3B-Instruct?

At 4-bit, PowerMoE-3b fits on 1x MI300X from $1.85 an hour and Qwen2.5-3B-Instruct on 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.

Can I use PowerMoE-3b commercially?

Yes. PowerMoE-3b is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.

Related Comparisons