---
type: "post"
id: "microsoft-foundry-local-ships-qwen3-8-27b-stack-ff29eb"
slug: "microsoft-foundry-local-ships-qwen3-8-27b-stack-ff29eb"
title: "Microsoft Foundry Local ships Qwen3.8-27B stack beating llama.cpp on RTX 5090 local coding"
description: "Microsoft says Foundry Local CLI 0.13.0 and later can run Qwen3.8-27B locally and connect it to GitHub Copilot through BYOK."
retrieval_nugget: "Microsoft says Foundry Local CLI 0.13.0 and later can run Qwen3.8-27B locally and connect it to GitHub Copilot through BYOK. The release gives developers a concrete on-device coding path, with stated GPU and context thresholds, but its speed comparison with llama.cpp is Microsoft’s own report rather than independent validation."
published_at: "2026-11-05"
updated_at: "2026-11-05"
record_date: "2026-11-05"
date_kind: "published_at"
topics: ["github","local-models"]
entities: ["Microsoft","Foundry Local","Qwen3.8-27B","GitHub Copilot","ONNX Runtime","llama.cpp"]
related_patterns: ["harness-architecture-outlives-model-choice"]
novelty_label: "new_fact"
what_changed: "Microsoft now offers a separate Foundry Local Qwen3.8-27B coding stack integrated with GitHub Copilot, claiming it beats llama.cpp on RTX 5090."
prior_refs: ["https://newruntime.com/posts/betelgeuse-llama-cpp-fork-pages-moe-experts-from-fb5996/"]
source_url: "https://devblogs.microsoft.com/foundry/foundry-local-agentic-coding-qwen3-8-27b"
source_urls: ["https://devblogs.microsoft.com/foundry/foundry-local-agentic-coding-qwen3-8-27b"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:microsoft-foundry-local-ships-qwen3-8-27b-stack-ff29eb"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/microsoft-foundry-local-ships-qwen3-8-27b-stack-ff29eb/","markdown":"https://newruntime.com/posts/microsoft-foundry-local-ships-qwen3-8-27b-stack-ff29eb.md","json":"https://newruntime.com/posts/microsoft-foundry-local-ships-qwen3-8-27b-stack-ff29eb.json"}
markdown_url: "https://newruntime.com/posts/microsoft-foundry-local-ships-qwen3-8-27b-stack-ff29eb.md"
---

# Microsoft Foundry Local ships Qwen3.8-27B stack beating llama.cpp on RTX 5090 local coding

## Retrieval answer

Microsoft says Foundry Local CLI 0.13.0 and later can run Qwen3.8-27B locally and connect it to GitHub Copilot through BYOK. The release gives developers a concrete on-device coding path, with stated GPU and context thresholds, but its speed comparison with llama.cpp is Microsoft’s own report rather than independent validation.

## Summary
Microsoft says Foundry Local CLI 0.13.0 and later can run Qwen3.8-27B locally and connect it to GitHub Copilot through BYOK. The release gives developers a concrete on-device coding path, with stated GPU and context thresholds, but its speed comparison with llama.cpp is Microsoft’s own report rather than independent validation.

## What happened
In a November 5, 2026 post, Microsoft’s Foundry team presents Qwen3.8-27B as a model sized for high-performance local deployment and shows how to use it with GitHub Copilot. The setup connects Copilot to a Foundry Local API endpoint on the developer’s machine, so the model can be selected as a provider in the Copilot app. Microsoft says the workflow requires Foundry Local CLI 0.13.0 or later. Users load the model with the CLI; if it is not already present, that command downloads it. NVIDIA GPU users must separately install compatible CUDA and cuDNN runtimes, which Foundry Local does not download.

Microsoft says its ONNX Runtime-based stack tunes the full inference pipeline for NVIDIA GPUs, including prefill, decoding, memory management, long-context processing and concurrent requests. The listed techniques include DFlash2 speculative decoding, CUDA graph fusions, kernel optimizations for NVIDIA’s latest SM121 architecture, INT4 weight handling, KV-cache quantization, embedding memory sharing and CPU offloading. Prefix caching reuses previously computed context, while continuous batching handles concurrent workloads. The post frames these as ways to reduce compute and memory bottlenecks, not as a guarantee of a particular speed on every machine.

With 4-bit quantization, Microsoft puts the model’s size at approximately 18 GB. It says a consumer GPU with 24 GB of VRAM can run it with context windows up to 64K tokens; GPUs with 32 GB or more can support context up to the model’s stated maximum of 262K tokens. The company compared Foundry Local with llama.cpp using the 4-bit model on an NVIDIA GeForce RTX 5090 with 32 GB VRAM and an NVIDIA DGX Spark with 128 GB unified memory, using a dataset of real GitHub issues. Microsoft reports higher generation speed at every tested prompt length, with the advantage widening beyond 131K tokens. The post does not provide independent validation of that result.

The suggested flow is to install the CLI, load `qwen3.8-27b`, retrieve the local API endpoint and add Foundry Local as a provider in Copilot settings. Microsoft describes local models as suited to everyday coding and targeted subtasks, while noting that smaller models may have limited reasoning compared with advanced cloud models. It says it plans to extend support to Macs and AMD GPUs, add more coding-optimized models and deepen the Copilot integration; these are future plans, not current availability.

## What the archive adds
This is a new release in a familiar local-inference thread. Our earlier note on [Betelgeuse’s llama.cpp fork](https://newruntime.com/posts/betelgeuse-llama-cpp-fork-pages-moe-experts-from-fb5996/) described NVMe-paged MoE inference aimed at running a larger Qwen model on a 12 GB GPU. The present item is not a restatement: Microsoft offers its own optimized runtime and Copilot connection, and directly compares performance with llama.cpp.

## Trend signal
This supports the assumption that Microsoft Foundry keeps models swappable within developer workflows: [the owned-learning-loop shift](https://newruntime.com/shifts/owned-learning-loop/) describes platforms separating model choice from customer workflows. Here the model can be selected through a local Foundry endpoint in Copilot. It is a narrow implementation example, not evidence about evaluation rubrics or trace portability.

## Why now
The post argues that cloud-only coding assistants can add costs, send sensitive context over the network and stop working when connectivity fails. Its proposed response is a local option for developers whose hardware can run the model, alongside cloud models for harder reasoning work.

## Why it matters
For a team considering local coding assistance, this release makes the hardware and integration requirements more concrete than a general claim that a model can run locally. A 24 GB NVIDIA GPU is the stated entry point, while the longer context claim requires 32 GB or more. Copilot BYOK provides a route to try the model within an existing coding interface, and local execution may be useful where keeping code on-device or working offline matters.

The trade-off is that teams should compare real tasks on their own machines rather than infer productivity from Microsoft’s throughput claim. The post itself says smaller local models may be weaker at complex reasoning. A hybrid arrangement—local inference for routine work and cloud models for difficult tasks—is the workflow Microsoft recommends, but cost, quality and privacy outcomes will depend on a team’s workload and configuration.

## Limits and uncertainty
The source is a Microsoft product post and does not provide independent benchmark results, the full dataset, raw measurements, hardware configuration details beyond the named systems, or enough information to reproduce the comparison. It reports higher generation speed at all tested prompt lengths, but does not establish that this translates into faster completed coding tasks. The post also gives no price comparison or measured privacy assessment. Its Mac and AMD support statements are roadmap plans without dates. The 24 GB and 32 GB thresholds are presented for particular quantization and context conditions, not as universal performance guarantees.

## Reader takeaway
Foundry Local offers a documented way to use Qwen3.8-27B locally through Copilot, provided the developer has a suitable NVIDIA GPU and installs the required runtime dependencies. Treat the llama.cpp speed advantage as a vendor claim to test against your own coding tasks, and keep cloud models available for work that exceeds the local model’s reasoning capacity.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Github](https://newruntime.com/topics/github.md): Explore the github topic hub.
- topic hub: [Local Models](https://newruntime.com/topics/local-models.md): Explore the local-models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)