Open Weights in Production: What Qwen, DeepSeek, Llama and GLM Really Cost

Chinese open weight models took about 61% of the tokens routed through OpenRouter by May 2026. A survey of 495 US enterprise decision makers put open source models at 11% of the enterprise LLM market. Any open source LLM comparison quoting one number and not the other is incomplete. Both are measured. They disagree because they count different buyers.

What open weights actually means, and what it does not

Open weights means the model parameters are published and you can download and run them. It does not mean the training data, the training code or the training process are public. Four families dominate the conversation in 2026: Qwen from Alibaba, DeepSeek, GLM from Z.ai, and Llama from Meta. Their licences differ more than their benchmarks do.

Open weights is not open source

The distinction is not pedantry. It decides what you can audit and what you must take on trust.

With published weights you can inspect behaviour, run the model inside your own network, and fine tune it. You cannot reproduce it. You cannot check what was in the training corpus. You cannot verify a claim about contamination or bias at source.

So the security case for open weights is narrower than the marketing suggests. It is a deployment control, not a provenance guarantee. That is still worth a great deal, and it is worth less than being able to see everything.

The four families you will be asked about

Qwen, from Alibaba, is the broadest family. It runs from small models you can host on one GPU up to a 2.4 trillion parameter mixture of experts. The 27 billion parameter version ships under Apache 2.0 with a 262,144 token native context.

DeepSeek ships fewer and larger models. DeepSeek V4 Pro is 1.6 trillion total parameters with 49 billion active per token, under the MIT licence, with a 1 million token context window.

GLM, from Z.ai, moved to MIT with GLM-5.2 in June 2026. That release put roughly 753 billion parameters and a 1 million token context under a licence with no user threshold and no regional restriction.

Llama is the one people assume is the safe Western default. On licence terms it is the most restrictive of the four, which is covered in detail below.

Two scoreboards, and they do not agree

Almost every argument about open weight adoption is two people quoting different instruments at each other. Both instruments work. They are pointed at different populations.

What the router counts

OpenRouter is a neutral router that developers point applications at. Its own study of more than 100 trillion tokens, covering roughly 13 months to November 2025, found proprietary models held about 70% of usage and open models about 30%. DeepSeek alone accounted for 14.37 trillion tokens, Qwen 5.59 trillion, Llama 3.96 trillion and Mistral 2.92 trillion.

The trend since then is steeper than the level. By May 2026, Chinese origin models were about 61% of tokens routed through the platform. Meta's Llama, the open weight leader two years ago, had dropped out of the top rankings.

This is real production traffic, not a survey. It is also self selected. Teams that route through OpenRouter are, by definition, teams that shop on price and switch models easily.

What the enterprise survey counts

Menlo Ventures surveyed 495 US enterprise AI decision makers between 7 and 25 November 2025. It found open source LLMs at 11% of the enterprise market, down from 19% the year before. Chinese open source models came in at roughly 1% of enterprise LLM API usage.

Its mid year update, drawn from a smaller sample of 150 technical decision makers, gave the reasons. A performance gap of 9 to 12 months, the technical complexity of deploying open models, and reluctance to send data to APIs run by Chinese companies.

Read those two paragraphs together and the picture resolves. Developers reach for open weights when nobody is watching the brand. Procurement does not sign for them.

Quality: where the four families actually land

Benchmark leadership now changes hands every few weeks, so treat any ranking as a snapshot with a date on it. Read on 28 August 2026, the Artificial Analysis Intelligence Index put the leading open weight models within a few points of the closed frontier.

The gap at the top is six points

Claude Opus 5 leads the index at 63. The best open weight entries sit at 60, 58 and 57. That is close enough that the ranking is no longer the deciding variable for most work.

What has not closed is the reliability gap around the edges. Long horizon agent runs, tool calling under load, and unusual output formats are where the frontier models still earn their price. If your workload is a single turn extraction or a classification, that premium buys you very little.

Index and price figures were read from Artificial Analysis on 28 August 2026 and will move. The Llama row is deliberately incomplete: no Llama model appeared in the top open weights view at the time of reading, and stating a rank we did not see would be inventing one.

Cost: the price gap is real, the saving is not automatic

The cost argument is the weakest of the three. It is also the one every vendor leads with. The list price gap is genuine, and what happens to it after deployment is where teams get caught.

The self hosting break even, worked

Take the standard unit of serious inference: an 8 GPU H100 node. Lambda's on demand list rate is $3.99 per GPU hour for an 8x H100 SXM instance. That is $31.92 an hour, or about $23,000 for a 30 day month running continuously.

Now divide that by the API price you would otherwise pay. The answer is the volume at which self hosting starts to win.

Our arithmetic, not a vendor's. Assumptions: $22,982 a month of GPU rental at Lambda list rates, 100% uptime, no engineering time, no storage, no networking, no idle capacity. Every one of those assumptions makes self hosting look better than it will be.

The bottom row is the only one most teams could reach. It also compares a self hosted open model against the most expensive closed model on the market. That is not a like for like comparison. It is a comparison of two different products.

The top row is the honest one. If a hosted API already sells you GLM-class quality at $0.10, the volume needed to beat it on rented hardware is roughly 88,700 tokens per second, sustained, forever. If you are not running at that scale, renting GPUs to serve open weights is a decision about control, not about cost.

There is a hardware wall underneath this too. Eight H100s carry 640 GB of memory between them. DeepSeek V4 Pro has 1.6 trillion parameters, which needs roughly 1.6 TB even at 8 bit precision. The largest open models do not fit on the node most people price them against, and the same effect shows up inside vendor pricing, as covered in the breakdown of what inference actually does to AI margins.

Reasoning tokens move the number after you sign

Blended prices assume a token mix. Reasoning models break that assumption, because they generate far more output tokens than the input suggests.

Artificial Analysis records Claude Opus 5 producing roughly 100 million tokens across its evaluation, against a median of 72 million for comparable models. Output tokens cost five times input on that model, at $25 against $5 per million. A 40% increase in verbosity is a 40% increase in the expensive half of your bill.

This is the mechanism behind bills rising while unit prices fall, which is examined in the piece on cheap tokens and rising AI bills. Before you compare two models on list price, compare them on tokens generated for the same task. Anything else is a cost model for a workload you do not have, and it will not survive contact with a real gross margin calculation.

Licences are not equal, and one of these is not open source

This is the section most comparisons skip. It is also the one that creates actual liability. Three of the four families ship under licences a lawyer will approve in an afternoon, and the fourth does not.

Qwen 3.8 is Apache 2.0. DeepSeek V4 and GLM-5.2 are MIT. Those are standard, permissive, decades old licences. No user threshold, no attribution duty beyond the notice, no naming rule, no field of use restriction.

What the Llama 4 licence asks that MIT does not

The Llama 4 Community License Agreement, effective 5 April 2025, is a bespoke commercial licence. Four clauses matter.

First, the user threshold. If your products had more than 700 million monthly active users in the calendar month before the release date, you must request a separate licence from Meta. Most companies are nowhere near it. The point is that a permissive licence contains no such clause at all.

Second, attribution. You must prominently display the words Built with Llama on a website, interface, blog post, about page or product documentation.

Third, naming. Any model you train, fine tune or improve using Llama materials, if distributed, must include Llama at the start of its name. That is a branding obligation attached to your own derivative work.

Fourth, jurisdiction. The agreement is governed by California law and California courts have exclusive jurisdiction over disputes. For a European or Indian buyer, that clause alone can be the one that stops the deal.

My position is straightforward. Llama should not be compared with MIT and Apache models as though the licences were interchangeable. Treat it as a commercial contract with a zero dollar fee, review it like one, and read it alongside the comparison of model provider terms before you standardise on it.

Sovereignty is the argument that actually holds

Strip out the cost claim and the benchmark claim, and one reason to run open weights survives every objection. The data never leaves your estate.

That reason is regularly confused with a different question, which is whether a given lab is safe to buy from. The two are not the same, and the confusion is expensive.

Italy's data protection authority ordered Hangzhou DeepSeek Artificial Intelligence and Beijing DeepSeek Artificial Intelligence to stop processing Italian users' personal data at the end of January 2025. The order followed an inquiry into what data the hosted service collected, where it was stored, and on what legal basis, as reported at the time by The Record.

Read what that order covers. It is about a hosted consumer service and its data processing. It is not a ruling about the MIT licensed weights on Hugging Face, which transmit nothing to anybody when you run them on your own hardware.

That distinction is the strongest argument in favour of open weights, and almost nobody makes it cleanly. A banned app is not a banned model. If your objection to a Chinese lab is data transfer, downloading the weights removes the objection entirely, and it is a different analysis from the one in the piece on data residency rules in the EU and India.

The regulatory relief is narrower than vendors imply, though. Article 53(2) of the EU AI Act exempts free and open source models from two obligations only: the technical documentation duty and the downstream information duty. The copyright policy requirement and the public training data summary still apply. And the exemption does not apply at all to models classified as carrying systemic risk, which is where the largest of these models sit. Anyone planning around that exemption should work through the transparency checklist first.

Where this argument is weakest

Three things in this post could be wrong, and it is better that you hear them from me.

The router data may not generalise. OpenRouter is a marketplace optimised for price sensitive switching, so a 61% share there tells you about developers shopping for cheap tokens. It says much less about regulated enterprises with procurement, indemnities and a security review. The author of that analysis says as much: first party consumer traffic dwarfs the router.

The survey data has the mirror problem. Menlo asked 495 US decision makers what they buy. Buying decisions lag engineering behaviour, so a survey taken in November 2025 cannot see a shift that happened in the first half of 2026. The honest reading is that the router is a leading indicator and the survey is a lagging one, and neither is the truth.

My cost model has soft assumptions. It uses list price GPU rental, which committed capacity beats by a wide margin. It assumes continuous utilisation, which almost nobody achieves. It ignores engineering time, which is usually the largest line. Those errors push in opposite directions, and I have not measured which one dominates for a given team. The break even table is a starting point for your own arithmetic, not a substitute for it.

One more concession, aimed at my own conclusion. The case that closed models are worth the premium is stronger than the price gap makes it look. Indemnities, support, retention controls and a named counterparty are things a board can act on when something goes wrong. No licence gives you that, however permissive it is.

A decision rule you can apply this quarter

Model choice is a portfolio decision, not a single bet. Here is the rule I would apply, which resolves most cases without a committee.

Row four is the one most teams get wrong. The evidence a regulator wants is your evaluation record, not your vendor's brand, which is the case for building an eval suite you own.

Frequently asked questions

Where to start this week

Two things, and neither needs a budget.

First, take your highest volume AI task and price it three ways: your current model, the cheapest open weight model that passes your evaluation, and a self hosted node at list rates. Use tokens generated for the same task, not list price per million. In most cases the middle column wins and the third column is not close.

Second, pull the licence for every model already in your stack and check three fields: user threshold, attribution duty and governing law. If any model in production ships under a bespoke community licence, that is a contract your legal team has not reviewed. Find out now, not during diligence.

Related reading