Qwen3.8-27B Weights Are Out. That's the Local Path for Specs That Can't Leave the Building

Alibaba put Qwen3.8-27B weights on Hugging Face around 14 August: a 27-billion-parameter dense vision-language model, native 262K context, Apache 2.0. You can keep the files on a machine you control. That's the part that matters if your shop won't put a drawing, a J1939 excerpt, or a supplier PDF into a cloud chatbot.

The same week, Unsloth, Ollama, and OpenRouter talked up GLM-5.3 as if you could run it now. Unsloth's write-up is a local path on a 256 GB-class machine. That isn't the box most controls engineers have.

What shipped

Unsloth published local GGUFs for GLM-5.3, Z.ai's 744B-parameter model (40B active). Unsloth lists the 2-bit UD-IQ2_M quant at 239 GB on disk and says it wants 245 GB of memory, aimed at a 256 GB Mac Studio or combined RAM plus VRAM. Their post is that path. Those are Unsloth's published requirements, not a measurement in this post.

Ollama listed GLM-5.3 and GLM-5.3-Flash. The commands on those pages are ollama run glm-5.3:cloud and ollama run glm-5.3-flash:cloud. The :cloud tag is hosted inference in the US and EU, with zero data retention per Ollama's policy. That's a contract. It's not the PDF staying on the bench. Ollama's announcement is the cloud rollout.

OpenRouter has GLM-5.3 as z-ai/glm-5.3, aimed at software engineering and long-horizon agents. Fine if policy already allows a vendor API. Not fine if the rule is the file doesn't leave the building.

Flash is the smaller GLM, not a laptop GLM. Unsloth's Flash guide puts the 1-bit build at 100 GB of RAM plus VRAM. There isn't a smaller local Flash tag on Ollama's library pages. Only :cloud.

Published sizes for Qwen3.8-27B are the ones that line up with a 32–64 GB laptop or a workstation GPU in the 24 GB class — the usual Eaton-style box. Community write-ups have described a Q4 GGUF loading on a single 24 GB card. That's their report. This post doesn't include a VRAM result, a token rate, or a screenshot. Check the card on the machine you'll use.

How the local path is supposed to work

The recipe is simple on paper. Install a runner that keeps weights on disk. Pull a model that isn't tagged :cloud. Give it a file path, not a paste into a browser. Ask it for a table. Check that table against the page.

Ollama lists qwen3.8:27b at 18 GB on disk, text and image. That's Ollama's listed download size, not a RAM result. After a pull, ollama list is how you see what landed. Compare that figure to free RAM plus VRAM and leave a margin for the OS and for context. llama.cpp is the same idea if that's the runner already on the box.

The file-as-path habit showed up again this week in T3 Code nightly: attachments land as real filesystem paths the agent can parse. You don't need that app. You need the PDF, or one exported page, on the same machine as the model.

Ollama says that when you run locally, they don't collect or transmit prompts. That's the IT sentence for "drawings stay in-house." The next sentence is the one people skip: glm-5.3:cloud still leaves the box. Zero retention isn't an air gap.

The catch

A local model doesn't grant you a license. If policy forbids local models on ITAR, customer IP, or export-controlled files, you still don't run those files through it. Start with a public datasheet page.

Treat the output as a draft. Local models still invent identifiers that sound like they belong on a CAN map, skip the 8-point footnote, and flip a max with a typical while formatting the mistake cleanly. Those are failure modes to watch for in the experiment below, not results from a run on this site. Text extractors usually fail on scanned drawings; a page image into a vision-capable local build would be a second recipe, not a reason to upload the PDF to a cloud vision API.

Secrets don't belong in the prompt either, local or not. Transcripts sit on disk. If you also use Cursor, Claude Code, or Codex, ctxcop is the companion story this week: Webflow built it to redact AWS-style credentials before they hit a coding agent's context. It doesn't turn a cloud agent into an air gap.

Adding :cloud to the name does not make the file stay on your bench.

The Monday experiment

Here's the experiment — a recipe, not a completed trial. Pull a model that fits. Feed it one PDF page you're allowed to use. Ask for a table. Diff that table against the page by eye. The hour is done when you have something you could paste into a review after you've marked the mistakes.

Use a public datasheet (CAN transceiver, regulator, sensor), a dummy ECO, a redacted spec, or a J1939 excerpt you're already licensed to keep on that computer. Don't grab a customer drawing "just to try it."

Check memory first. Linux: free -h and nvidia-smi if you have NVIDIA. Mac: System Settings → About. Windows: Task Manager → Performance → Memory, plus nvidia-smi on a discrete GPU. Free RAM plus VRAM, not the sticker on the chassis.

On Linux, Ollama's docs are:

curl -fsSL https://ollama.com/install.sh | sh

If IT won't let you pipe a script to a shell, use the manual tarball on the same page. macOS and Windows: ollama.com/download. Confirm it's local with ollama -v. Then:

ollama pull qwen3.8:27b
ollama list

If the listed size is larger than free RAM plus VRAM, ollama rm qwen3.8:27b and pick a smaller local tag. Don't reach for :cloud to make it fit.

Put one page on disk:

pdftotext -layout -f 1 -l 1 datasheet.pdf page.txt

No pdftotext? Export or copy that page from a local viewer into page.txt. Keep it out of the cloud-sync folder if that's how files wander in your shop.

Then:

cat extract-prompt.txt page.txt | ollama run qwen3.8:27b

Tell it to read only the appended page, emit a markdown table (parameter, symbol if stated, min / typ / max, units, condition, source line), write UNKNOWN when the number isn't there, list footnotes it skipped, and name the three places it's least sure. For an ECO, swap the columns to was / becomes / zone / risk if missed.

Open the PDF and the table together. You're the checker.

What it isn't

It isn't a CAD copilot. It won't open Creo. It isn't a hardware standard for robots. It doesn't replace the person who knows which limit is load-bearing.

If the MATLAB pile is the next problem, MathWorks shipped the MATLAB Agentic Toolkit in April — MCP into the MATLAB install on your machine, plus skills so the agent stops guessing function names. Different hour. This one is the datasheet.

Unsloth's GLM-5.3 2-bit path is written for a 256 GB Mac. Ollama's GLM tags this week are :cloud. On a 32–64 GB laptop or a 24 GB-class workstation GPU, the published Qwen3.8-27B sizes are the ones that fit the catalog. Verify VRAM on the machine that will run it. Don't sign the first draft.

Sources

Qwen3.8-27B official weights: https://huggingface.co/Qwen/Qwen3.8-27B
Alibaba Cloud recap (Apache 2.0, mid-August 2026 open weights): https://www.alibabacloud.com/blog/alibaba-unveils-qwen3-8-27b-and-releases-weights-of-qwen3-8-flagship-model_603463
Ollama library, Qwen3.8 (qwen3.8:27b): https://ollama.com/library/qwen3.8
Unsloth, "GLM-5.3 - How to Run Locally": https://unsloth.ai/docs/models/glm-5.3
Unsloth, "GLM-5.3-Flash: How to Run Locally": https://unsloth.ai/docs/models/glm-5.3-flash
Unsloth on X (2-bit GGUF / 256 GB): https://x.com/UnslothAI/status/2093397494889890050
Ollama library, GLM-5.3 (:cloud): https://ollama.com/library/glm-5.3
Ollama library, GLM-5.3-Flash (:cloud): https://ollama.com/library/glm-5.3-flash:cloud
Ollama on X (cloud rollout): https://x.com/ollama/status/2093492205029908752
Ollama privacy policy: https://ollama.com/privacy
Ollama Linux install: https://docs.ollama.com/linux
OpenRouter, GLM-5.3: https://openrouter.ai/z-ai/glm-5.3
OpenRouter on X: https://x.com/OpenRouter/status/2093362477706969448
T3 Code / file-path attachments (pattern only): https://x.com/theo/status/2093622750510563606
Webflow, ctxcop: https://webflow.com/blog/securing-secrets
ctxcop repository: https://github.com/webflow/ctxcop
MATLAB Agentic Toolkit (next hour, not this one): https://blogs.mathworks.com/matlab/2026/04/13/introducing-the-matlab-agentic-toolkit/