Llama-3.1-405B-Instruct-MXFP4-Preview
Llama-3.1-405B-Instruct-MXFP4-Preview is a 405 billion parameter chat model from amd, released May 30, 2025. Llama-3.1-405B-Instruct-MXFP4-Preview is an open-weights chat model with roughly 405 billion parameters.
by amd · 405B parameters
Best for
- complex multi-step reasoning
- agent orchestration with tool use
- long-document analysis and summarisation
Ways to use Llama-3.1-405B-Instruct-MXFP4-Preview in osFoundry
Connect with your own key (BYOK)
Open the key dialog and paste your amd API key. osFoundry discovers Llama-3.1-405B-Instruct-MXFP4-Preview automatically — assign it to a Maestro role (router, direct, orchestrator, or fallback) in the Pipeline tab and it is live in every chat. Your key, your provider account — no token markup.
Deploy a dedicated endpoint
Llama-3.1-405B-Instruct-MXFP4-Preview is open-weights — run it locally for free, or deploy a dedicated GPU endpoint in your workspace for reserved capacity with no rate limits.
Use it in a Room App
Room Apps declare AI features in their manifest, then call them with invokeAI:
import { invokeAI } from '@osfoundry/app-sdk'
// 'summarize' is an AI feature declared in your app manifest.
const result = await invokeAI('summarize', userText)
Call it from your own apps
Once a model is wired into your workspace you can host it as an API and reach it from your own services, scripts, or CI — outside osFoundry.
What hardware can run Llama-3.1-405B-Instruct-MXFP4-Preview
Llama-3.1-405B-Instruct-MXFP4-Preview runs on a multi-GPU setup or H200 141GB at Q4 (~243 GB VRAM with KV-cache headroom). Full-precision inference requires multiple H100/H200 GPUs at FP16 (~972 GB).
Llama-3.1-405B-Instruct-MXFP4-Preview vs similar models
Licence
Unspecified — Licence terms not specified — verify the upstream model card before commercial use.
Check upstream documentation.
Frequently asked about Llama-3.1-405B-Instruct-MXFP4-Preview
Is Llama-3.1-405B-Instruct-MXFP4-Preview free to use?
Llama-3.1-405B-Instruct-MXFP4-Preview is free to run locally on your own hardware. Hosted access through osFoundry is metered (input Free (local), output Free (local)). You can switch between local and hosted at any time.
Can I use Llama-3.1-405B-Instruct-MXFP4-Preview commercially?
Commercial use is allowed with conditions. Licence terms not specified — verify the upstream model card before commercial use. Check upstream documentation.
How much VRAM does Llama-3.1-405B-Instruct-MXFP4-Preview need?
Approximately 243 GB at Q4 quantisation, or 972 GB at full FP16 precision. Requires multi-GPU at higher quantisation.
Can I run Llama-3.1-405B-Instruct-MXFP4-Preview locally?
Yes. Llama-3.1-405B-Instruct-MXFP4-Preview is open-weights and runs locally on a workstation GPU. osFoundry's local runtime handles model loading, quantisation, and routing.
What is Llama-3.1-405B-Instruct-MXFP4-Preview best at?
Llama-3.1-405B-Instruct-MXFP4-Preview is well-suited to complex multi-step reasoning, agent orchestration with tool use, long-document analysis and summarisation.
How do I use Llama-3.1-405B-Instruct-MXFP4-Preview in osFoundry?
Paste your amd API key in the key dialog (or deploy the open weights for self-hostable models), assign Llama-3.1-405B-Instruct-MXFP4-Preview to a Maestro role in the Pipeline tab, then use it in chat, Room Apps via invokeAI, or your own apps.
Published by amd on May 30, 2025. Source: https://huggingface.co/amd/Llama-3.1-405B-Instruct-MXFP4-Preview