Feature Request: Add Qwen3.8-27B to the Foundry Local catalog

[Neil-Godfrey (Neil-GS)]

Is your feature request related to a problem? Please describe.

Foundry Local's catalog currently tops out at the Qwen 3.5 series (and older Qwen2.5 / small Qwen3). Alibaba released Qwen3.8-27B this month — a dense 27B native vision-language model, Apache-2.0, 262K context, thinking + tool calling. It is already the model I run locally for coding and agent work, but I have to do that in Ollama because it is not in the Foundry catalog.

I would rather stay on Foundry Local: C# SDK, OpenAI-compatible endpoint, curated ONNX, Windows + Apple Silicon execution providers. The catalog gap is the only reason this model lives outside the Microsoft runtime.

Describe the solution you'd like

Add Qwen3.8-27B to the Foundry Local catalog, with ONNX builds for the execution providers you already ship for Qwen 3.5:

• Apple Silicon (CoreML / Metal) — this is the machine I can test on
• NVIDIA CUDA
• CPU fallback

Weights: https://huggingface.co/Qwen/Qwen3.8-27B
Ollama (for comparison, already running here): qwen3.8:27b (19 GB quantized)

There is a single deployment-sized dense checkpoint in this generation (Qwen3.8-27B). Qwen3.8-Max (2.4T MoE) is not a local-catalog candidate.

Describe alternatives you've considered

• Ollama / GGUF — works today on this machine (qwen3.8:27b on a MacBook Pro M5 Max, 128 GB). That is not the Foundry path, and GGUF does not load in Foundry Local.
• Olive compile from Hugging Face — official custom-model path. microsoft/olive-recipes has folders for Qwen3.5-27B and Qwen3.6-27B; there is no Qwen3.8 folder. Qwen3.8 uses a hybrid Gated DeltaNet + Gated Attention layout (same family as 3.5, new generation). A generic olive optimize against Llama-shaped recipes is likely to fail. I am willing to attempt a recipe PR against the 3.5-27B template and report the result, including full logs if it breaks.
• Wait for the next catalog drop — fine if that is the plan. A named tracking discussion helps.

Additional context

• Precedent: Discussion #514 requested Qwen 3.5; 0.8B / 2B / 4B / 9B landed in the catalog (June 2026), 27B was noted as in progress. Qwen3.8-27B is the same size class, current generation.
• Hardware I can validate on: MacBook Pro, Apple M5 Max, 128 GB unified memory, macOS. Happy to run a preview build and file results.
• Why this model specifically: it is the current open-weight bar for coding and long-horizon agent tasks at a size that actually fits a workstation. Catalog Qwen 3.5 is a generation behind what developers are already running locally.

Happy to follow up with an olive-recipes PR (info.yml + recipe cloned from Qwen-Qwen3.5-27B) if that is the useful next step versus waiting on an official ONNX build.