Open AI models (often called open-weight) let you download the software and run it on your own servers or a private cloud. Closed models stay behind a vendor API. Open wins on cost, privacy and control of drawings; closed frontier models still lead on the hardest multi-step reasoning. Most shops will use both.
If you own a fab shop, you have already heard two competing pitches. One side says you should live inside ChatGPT or Claude. The other says you should run “open source AI” on your own iron. Both can be right — for different jobs. Open vs closed AI models is not a religion. It is a make-vs-buy decision that looks a lot like buying a CNC versus leasing capacity at a job shop.
What is the difference between open and closed AI models?
A closed (or proprietary) model is a service. OpenAI, Anthropic, Google and similar labs keep the model weights — the giant file of numbers that is the model — on their side. You send text or files through an API or chat app and get answers back. You rent intelligence by the token, the way you might lease laser time.
An open or open-weight model is closer to software you can own. Labs such as Meta (Llama), Mistral, DeepSeek, Alibaba (Qwen) and Moonshot (Kimi) publish weights you can download. You, or a partner you trust, can run that model on servers you control. You still need GPUs, power and someone who knows how to keep the lights on — but the intelligence is not locked behind one company’s front door.
People say “open source AI” in conversation. Strictly speaking, many popular releases are open-weight: you get the model file and a license for commercial use, but not always the full training recipe. For a plant owner, the practical question is simpler: Can I run this where my data already lives?
Shop-floor analogy: closed AI is the service bureau that cuts parts from your DXF and ships them back. Open-weight AI is buying (or leasing) a machine that sits in your building, reading your nest files, with drawings that never leave the site.
Why is the market shifting toward open models?
Four forces matter for manufacturers — and they show up in real news, not just Twitter threads.
1. The quality gap on everyday work got small.
Through 2025 and into 2026, downloadable models from Meta, DeepSeek, Qwen, Mistral and others closed in on frontier closed systems for ordinary tasks: reading PDFs, drafting email, extracting fields, summarizing long specs. Meta released Llama 4 Scout and Maverick as openly available weights in April 2025. DeepSeek’s V3 and R1 releases forced every lab to rethink cost and openness (TechCrunch). You do not need the absolute best model to turn a mill cert into a spreadsheet row.
2. Cost.
Open weights, once hosted on your cloud or a specialist inference host, often cost a fraction of premium closed APIs for high-volume work. That matters when AI moves from “try it on one bid” to “every RFI, every week.” Menlo Ventures’ 2025 enterprise AI report lists the classic open-model advantages: customization, cost savings, and private-cloud or on-premises deploy.
3. Privacy and control of customer drawings.
Your GC’s structural drawings, coating specs and bid packages are not content for a consumer chatbot. Open weights (or a private-hosted model) keep inference inside a boundary you define — same instinct as locking print sets in the drawing room. Mistral notes many customers already run models in their own data centers and want regional control of inference. For Canadian fabricators with owner confidentiality requirements, that control is often the whole sale.
4. Running on your own servers (or a private cloud).
Once weights are downloadable, a systems integrator can put the model next to your file share, email archive or fab management database. You are not waiting on a public status page when a big bid is due Friday. You also avoid feeding a shared consumer product with proprietary takeoffs.
Honest caveat: that same Menlo survey found enterprise share of open-source LLMs fell from about 19% to 11% as buyers stayed cautious — even while startups raced toward cheaper open options like Qwen, DeepSeek and Kimi. The shift is real in capability, price and options. It is not yet “every plant ripped out Claude.” Treat open as a mature option on the menu, not a mandate.
What did the All-In podcast say about open source AI?
In July 2026, the All-In hosts spent a long stretch of episode 282, “The Fight Over Open Source AI…” on this exact debate — sparked by Moonshot AI’s open Kimi K3 release. (We are summarizing only points that appear in public show notes and transcripts.)
Takeaways that map onto a fab shop:
- Open means you can run it yourself. David Friedberg stressed the plain definition: a downloadable package you can put on your own servers, without staying tied to someone else’s chat window.
- Most jobs do not need the tip of the spear. Jason Calacanis and Chamath Palihapitiya argued that a large share of real workloads — Chamath framed it as roughly “95% of the tasks” — can be handled by many models, including cheaper open ones, while a smaller slice still deserves premium frontier systems.
- Control vs convenience. David Sacks noted open deployments are more controllable and customizable, and that enterprises often want models on their own infrastructure for data sovereignty — with the trade-off that open setups take more work than a polished API. He also argued both open and closed can win in a large market.
You do not need the Washington subplot to use the business lesson: route commodity document work to controlled, cheaper models; keep paying for closed frontier horsepower where judgment is expensive.
When should fabricators still use closed frontier models?
Closed models from the leading labs still earn their keep on the hardest problems:
- Complex, multi-step reasoning across messy bid packages with conflicting addenda, unusual alloys, and schedule logic that is not a simple checklist.
- Ambiguous change-order and claim language where tone, liability and commercial judgment matter.
- Agent-style workflows that chain many tools and decisions — closer to a junior coordinator who must improvise than a clerk filling a form. See what an AI agent is.
- Brand-new problem types your team has never seen, where general reasoning depth beats a smaller open model.
Menlo’s data still shows most enterprise API dollars with Anthropic, OpenAI and Google — buyers paying for reliability on hard reasoning. Your estimator on a one-of-a-kind stadium package may want that horsepower. Your overnight mill-cert matcher probably does not.
How should a shop choose between open and closed AI?
Use the same gut check you use for equipment:
| Question | Lean open / private | Lean closed frontier |
|---|---|---|
| Volume | Hundreds of similar docs a week | A few hard packages a month |
| Sensitivity | Customer drawings, pricing, owner marks | Public specs, generic Q&A |
| Latency / uptime | Must keep working if the public API hiccups | Occasional use is fine |
| IT appetite | You have (or will hire) someone to host and patch | You want a vendor login and a bill |
| Error cost | Low — human reviews every send | High — one wrong read costs a rework truck |
Most mid-market manufacturers will land on a hybrid: private or open-weight models for intake, logging and first drafts; closed frontier models for hard judgment. That matches how we talk about using AI in a steel fabrication shop — start with document-heavy workflows, keep people on price, stamps and safety.
Vendor checklist: questions to ask before you commit
Print this for your next sales call.
- Where do our files go? Which region, which cloud, who can see prompts and uploads?
- Can this run on our servers or a private tenancy? Or is it API-only forever?
- Is the underlying model open-weight, closed, or a mix? Can we switch models without rewriting the whole workflow?
- What is the unit cost at our volume? Ask for a quote at 10× today’s usage — surprise bills kill ROI.
- Who approves outputs that touch money or schedule? If the answer is “the model,” walk away.
- What is the rollback plan? If the vendor or model disappears Friday, what still runs Monday?
- License and provenance. Especially with open weights: commercial terms, attribution, and whether your counsel is comfortable with the origin of the weights.
Want help picking the right model stack for your shop?
Embedding AI builds AI employees for mid-market fabricators and manufacturers. We care less about brand-name models than about which jobs get done every day — bid intake, RFI chasing, mill certs, order-to-delivery follow-up — with your people still owning judgment. Book a free 30-minute consultation and we will map one workflow, show you where open vs closed actually matters for your data, and leave you with a practical plan. Prefer email? Write to info@embeddingai.ca.
Sources
- All-In Podcast, Episode 282: The Fight Over Open Source AI… (July 24–25, 2026); transcript.
- Menlo Ventures, 2025: The State of Generative AI in the Enterprise (December 9, 2025).
- Reuters, Meta releases new AI model Llama 4 (April 5, 2025).
- TechCrunch, DeepSeek: Everything you need to know (September 29, 2025).
- Mistral AI, In-region inference, open models, and new European infrastructure for sovereign AI (August 11, 2026).
Frequently asked questions
What is the difference between open and closed AI models?
Closed models (ChatGPT, Claude, Gemini and similar) are used through a vendor's service — you send a prompt and get an answer, not the model file. Open or open-weight models can be downloaded and run on infrastructure you control, so customer drawings and mill certs need not leave your environment.
Is open source AI good enough for a fabrication shop?
For many high-volume jobs — summarizing bid packages, drafting RFI replies, matching mill certs — strong open models are already useful and often much cheaper. Keep closed frontier models for the toughest judgment work, where a small mistake is expensive.
Can I run AI on my own servers so drawings stay in-house?
Yes. That is the main practical reason shops care about open weights. Running the model on your own hardware or a private cloud is closer to keeping fabrication drawings in the vault than faxing them to a service bureau. You still need access controls, logging and a human approval gate.
Why are companies talking more about open source AI lately?
Quality has closed the gap on everyday work, prices for open inference fell hard, and owners want control over sensitive data. Meta's Llama 4, DeepSeek, Qwen, Mistral and others made downloadable models a real option. Enterprise buyers are still cautious, but the strategic choice is no longer theoretical.
Should my shop pick only open or only closed AI?
Usually neither. Route routine, high-volume document work to a controlled open or private setup, and reserve premium closed models for hard reasoning. Ask every vendor where your data goes, whether the model can run privately, and who approves outputs that affect price or schedule.