# AMD's Instella-MoE Brings Reasoning to ROCm Hardware

> A new open mixture-of-experts model with 16B total parameters and just 3B active is tuned to run on AMD's own accelerator stack.

Published by The Open Weights on Jul 23, 2026. Canonical: https://theopenweights.com/news/instella-moe-16b-a3b-think-ygvm

## Key facts

- Company: Amd
- Model: Instella-MoE 16B A3B Think
- Version: 1.0
- Category: Reasoning
- Modalities: Text / LLM, Reasoning
- License: Other (Open weights, commercial use allowed)
- Parameters: 16B (mixture of experts)
- Active parameters: 3B
- Context window: 33K tokens
- Architecture: InstellaMoEForCausalLM
- Precision: BF16
- Weights size: 31.7 GB
- Significance: notable
- Published: 2026-07-23
- Last verified: 2026-08-06
- Hugging Face: https://huggingface.co/amd/Instella-MoE-16B-A3B-Think
- Canonical URL: https://theopenweights.com/news/instella-moe-16b-a3b-think-ygvm

AMD has released [Instella-MoE 16B A3B Think](https://huggingface.co/amd/Instella-MoE-16B-A3B-Think), an open-weight reasoning model built as a mixture of experts. The design pairs 16 billion total parameters with roughly 3 billion active at inference, a structure meant to keep compute costs modest while retaining the capacity of a larger network.

The "Think" label signals the model's focus: it is positioned as a reasoning-oriented system rather than a general chat assistant, joining a growing field of open models that expose step-by-step problem solving. It handles text generation alongside its reasoning workload.

## Why it matters

The most notable angle here is the hardware. AMD explicitly optimizes Instella-MoE for its ROCm software stack, giving developers on AMD accelerators a first-party option rather than defaulting to models tuned for competing platforms.

- 16B total parameters, ~3B active per token
- Reasoning and text generation focus
- Optimized for AMD's ROCm ecosystem
- Open weights available on Hugging Face

As an initial 1.0 release, Instella-MoE reflects AMD's continued push to build a credible open-model presence tied to its own silicon. The weights and details are published on [Hugging Face](https://huggingface.co/amd/Instella-MoE-16B-A3B-Think) under a custom license, so teams weighing it for production should review the terms before deploying.

## Get the model

- [Hugging Face](https://huggingface.co/amd/Instella-MoE-16B-A3B-Think)

## Sources

- [amd/Instella-MoE-16B-A3B-Think](https://huggingface.co/amd/Instella-MoE-16B-A3B-Think) — Hugging Face, Jul 23, 2026

---
Source: The Open Weights (https://theopenweights.com/). Aggregated and written by Claude, curated by humans. Cite as: "AMD's Instella-MoE Brings Reasoning to ROCm Hardware", The Open Weights, Jul 23, 2026, https://theopenweights.com/news/instella-moe-16b-a3b-think-ygvm