# The toll booth on the open road: grading August's open licenses

URL: https://www.thedeepfeed.ai/posts/2026-08-12-the-toll-booth-on-the-open-road/
Category: Models
Published: 2026-08-12
Author: the-deep-feed
Tags: open-weights, licensing, qwen, kimi-k3, meta, alibaba, revenue-share
Kind: deep

> Qwen3.8-Max's weights went live today under a license that charges companies above $50M in revenue. Kimi K3's asks up to 30% above $20M. Meta's Glimmer is Apache-clean. 'Open weight' just fractured into regimes — this is the map of who pays.

## TL;DR

- **Qwen3.8-Max's weights went live today** — 2.4T parameters across 213 shards — under a custom license, not Apache: companies offering model-as-a-service or AI assistants with **more than $50M annual revenue** need a separate commercial license, and Clause 2 names prohibited product categories.
- **Kimi K3's license seeks up to 30% revenue share above a $20M sales threshold**, per Reuters. The green flag we gave K3 on [the July openness map](/posts/2026-07-20-china-open-weight-surge/) now carries a price tag.
- **Meta's Muse Glimmer (Aug 10) is Apache 2.0, no strings** — and Alibaba has promised a 27B model 'within about a week' of Aug 3. 'Open weight' now spans clean-permissive to revenue-conditioned custom, and the difference is measured in basis points.
- **The contrarian read: this is not open-washing.** The revenue-share license is rational price discrimination — distribution stays free and buys doctrine points; competition gets charged and protects the API business. Hobbyists never pay, enterprises sometimes, rivals always.

The weights arrived today. Qwen3.8-2.4T-A95B — the 2.4-trillion-parameter flagship Alibaba shipped to its API on August 3 — landed on ModelScope and Hugging Face as 213 safetensors shards, a checkpoint so large that serving it at 4-bit quantization still wants roughly 1.2 terabytes of VRAM. The repo had been sitting empty since August 8; this morning it filled. On raw availability, Alibaba just did the thing it promised on [the July map](/posts/2026-07-20-china-open-weight-surge/) it hadn't yet done: the preview became a download.

But the file that matters most in that repo is not a shard. It is [the LICENSE file](https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/blob/main/LICENSE), and it is not Apache 2.0. Under its terms, companies offering model-as-a-service or AI assistants with more than $50 million in annual revenue need a separate commercial license from Alibaba. Clause 2 goes further and names prohibited product categories outright. Five days ago, [Reuters reported](https://www.reuters.com/business/retail-consumer/alibaba-plans-charge-big-users-its-next-open-source-ai-model-sources-say-2026-08-07/) that this was coming — and that Moonshot's Kimi K3, the 2.8-trillion-parameter model we mapped as the open frontier's ceiling in July, carries a license seeking revenue share of up to 30 percent above a $20 million sales threshold.

Two days before the Qwen weights landed, Meta went the other way: [Muse Glimmer](/posts/2026-08-10-zuckerbergs-6500-word-bet/), 30 billion parameters, Apache 2.0, no thresholds, no clauses, wrapped in a 6,500-word manifesto about openness as safety. Three releases inside ten days, all marketed with the same adjective, governed by licenses that could not be more different.

In July we traced [the giveaway becoming a doctrine](/posts/2026-07-18-the-giveaway-became-a-doctrine/). In August the doctrine got a rate card. This is the map of who pays — and the honest answer is more interesting than the "open-washing" complaint that will dominate the feeds. The toll booth on the open road is not hypocrisy. It is pricing. And the toll sits exactly where a competitor would drive.

![Schematic: open highway with tiered toll gates — free hobbyist lane, $20M Kimi 30% gate, $50M Qwen license gate, rivals lane blocked.](/post-images/2026-08-12-the-toll-booth-on-the-open-road/toll-road-tiers.jpg)

# Three regimes, one adjective

The July map graded openness on a spectrum from MIT to API-only. That spectrum still holds, but the interesting action has moved into the middle: the custom license is no longer a vague acceptable-use document you skim. It is a priced instrument with revenue thresholds, named product categories, and a negotiation on the other side. As of today, the "open" releases in play sort like this:

| Model | Maker | Params | License | Threshold / terms | Grade |
|---|---|---|---|---|---|
| **Muse Glimmer** (Aug 10) | Meta (US) | 30B dense | Apache 2.0 | None. Commercial use, modification, redistribution all clean | 🟢 Permissive |
| **LongCat-2.0** (Jul 4) | Meituan (CN) | ~1.6T MoE | MIT | None | 🟢 Permissive |
| **Qwen3.8-27B** (promised) | Alibaba (CN) | 27B dense | Not yet posted | Promised "within about a week" of Aug 3; no repo, no license text as of today | 🟡 A promise, graded as a promise |
| **GLM-5.2** (in market) | Zhipu / Z.ai (CN) | Frontier-class | Custom, bespoke | Acceptable-use and commercial clauses; read before building | 🟡 Conditional |
| **Kimi K3** (Jul 16) | Moonshot (CN) | ~2.8T MoE | Custom, revenue-conditioned | Revenue share of up to **30% above $20M sales** (per Reuters) | 🟡 Free until you're big |
| **Qwen3.8-Max** (weights today) | Alibaba (CN) | 2.4T MoE / 95B active | Custom, revenue-conditioned | Separate commercial license for model-as-a-service / AI assistant providers **above $50M annual revenue**; Clause 2 names prohibited categories | 🟡 Free until you compete |

Read the license column top to bottom and you are reading a strategy gradient. At the top, makers who monetize elsewhere and want maximum distribution. At the bottom, makers who run API businesses and want distribution *minus* the customers who would otherwise pay for the API. The middle rows are the promises and the bespoke terms — July's yellow flags, now with company.

One regrade is owed. On the July map, Kimi K3 earned a green flag as genuine open weights, with the asterisk that open meant downloadable, not reproducible. The weights are still downloadable — but the Reuters reporting on the license terms moves K3 into the conditional tier. A checkpoint you must share 30 percent of your revenue to commercialize at scale is not the same instrument as MIT, and the map has to say so. The [ceiling we measured in July](/posts/2026-07-16-kimi-k3-open-frontier-ceiling/) was about the economics of serving a 2.8T model. It turns out the license was part of the economics all along.

# The license file is the release

The Qwen3.8-Max license deserves a close reading, because its structure is the story. The trigger is not usage or scale of deployment. It is a *business description plus a revenue number*: per the license text, the separate commercial license applies to companies offering model-as-a-service or AI assistants with more than $50 million in annual revenue. [Plain AI's read of the document](https://plainai.tech/articles/qwen38-clause-two-named-products) walks through Clause 2, which enumerates prohibited product categories by name rather than leaving them to an acceptable-use policy's judgment.

Consider what that trigger does and does not catch. A startup fine-tuning Qwen3.8-Max for an internal tool: untouched. A university lab: untouched. A mid-size company embedding it in a product, under $50 million: untouched. A hyperscaler standing up a Qwen3.8-Max serving endpoint, or an assistant company past $50 million in revenue: caught, precisely. The license is not a fence around the commons. It is a customs checkpoint that waves through everyone except the trucks — and the trucks are, almost by definition, Alibaba Cloud's competitors. [SCMP's coverage](https://www.scmp.com/tech/tech-trends/article/3363927/alibaba-adds-commercial-restrictions-open-weight-qwen38-max-ai-model) calls these commercial restrictions on an open-weight model, which is accurate but undersells the precision: they target the model-serving business specifically.

The two-license structure — [one document for the crowd, one negotiation for the incumbents](https://sqmagazine.co.uk/qwen3-8-open-weights-two-licenses/) — resolves the contradiction Alibaba has carried since it previewed this model at WAIC while promising weights "soon." The promise is kept. The weights are here. And the fine print prices exactly the scenario Alibaba was always going to have a problem with: someone else selling access to Alibaba's flagship.

Kimi K3's version is blunter. Per Reuters, the license seeks up to 30 percent revenue share above a $20 million sales threshold. Thirty percent is not a royalty; it is app-store economics applied to a model checkpoint. The same Reuters story carries the detail that makes the design legible: Moonshot prices the K3 API at roughly a third of Anthropic's Fable. The license and the price list are one mechanism. Below $20 million, take the weights, build, evangelize. Above it, the arithmetic tilts hard toward Moonshot's API, which is priced to be the obvious choice. The revenue share is less a fee than a steering current.

# Four questions Apache never asks

Now go one level down, because "revenue share" is a headline, not a mechanism. A revenue-conditioned model license has to answer four questions Apache 2.0 never asks, and the answers determine whether the toll is a fee or a leash.

**First: revenue as measured how, and audited by whom?** A percentage of revenue is meaningless without a definition of the base (gross receipts, revenue attributable to the model, or revenue of the whole product the model sits inside) and an enforcement path. Commercial software royalties come with audit clauses for exactly this reason: the licensor gets the contractual right to inspect the licensee's books. Neither the public Qwen3.8-Max license text nor the Reuters account of the K3 terms discloses an audit mechanism — the real obligations live in the *separate commercial license* negotiated after crossing the threshold. The public document is the price sticker. The contract is behind the counter.

**Second: who counts as you?** The Qwen trigger is annual revenue of the *company*, not revenue from the model. A $60 million logistics firm that fine-tunes Qwen3.8-Max for an internal scheduling assistant is over the line the moment its assistant is deemed customer-facing. Does a subsidiary's revenue roll up? The text does not say, and ambiguity in a license always prices in the licensor's favor, because the cost of being wrong is an infringement claim against your core product.

**Third: what happens to derivative weights?** Fine-tune Kimi K3 on proprietary data and the resulting checkpoint is a derivative of the licensed weights — your data, their substrate. Under Apache 2.0 the derivative is unambiguously yours. Under a custom license, the derivative inherits the terms, which means the revenue-share follows your fine-tune into every downstream deployment. The license taxes the value *you* added.

**Fourth: can you pass the weights on?** Apache grants sublicensable rights; the custom regimes generally do not. A platform that redistributes Qwen3.8-Max to its own customers is not a user but a distributor, and distribution above the threshold is precisely the named prohibited activity. The license is viral in reverse: instead of freedom propagating downstream, the restriction does.

None of this is exotic. It is what commercial licensing has always looked like. What is new is finding it inside a repo tagged "open weights," one click from a download button that mentions none of it.

# We have run this experiment before, three times

The pattern (open until the economics bite, then a license that carves out the competitor) has a documented prior history in infrastructure software, and each episode predicts a piece of what happens next.

| Episode | Date | The move | The trigger | What happened |
|---|---|---|---|---|
| MySQL dual licensing | ~2000 onward | [GPL for the community, commercial license for embedders](https://lwn.net/Articles/172128/) | Shipping MySQL inside proprietary products | Became the canonical open-source business model; MySQL sold to Sun for $1B in 2008 |
| MongoDB → SSPL | [Oct 16, 2018](https://www.mongodb.com/company/newsroom/press-releases/mongodb-issues-new-server-side-public-license-for-mongodb-community-server) | Relicensed from AGPL to SSPL | Cloud providers "in Asia" selling hosted MongoDB without contributing | OSI never certified SSPL as open source; MongoDB kept growing anyway |
| Elastic → Elastic License / SSPL | [Jan 14, 2021](https://www.elastic.co/blog/licensing-change) | Elasticsearch off Apache 2.0 | AWS selling Elasticsearch as a service — Elastic's blog was titled ["Amazon: NOT OK"](https://www.elastic.co/blog/why-license-change-aws) | [AWS forked it](https://aws.amazon.com/blogs/opensource/stepping-up-for-a-truly-open-source-elasticsearch/) as OpenSearch within a week |

Every mechanism in August's model licenses has an ancestor in that table. MySQL invented the two-document structure Alibaba is using: one license for the crowd, one negotiation for companies whose usage looks like competition. MongoDB's SSPL invented the *business-description trigger* — it did not restrict use, it restricted *offering the software as a service*, which is Clause 2's named-product-category move with 2018 drafting. And Elastic supplied the cautionary ending: when the license moved against the hyperscaler, the hyperscaler simply forked the last permissive version and kept serving customers.

The fork threat is the piece that does not transfer cleanly to models, and the makers know it. You can fork Elasticsearch because the source compiles and the community can maintain it. You cannot fork Qwen3.8-Max in any meaningful sense: the "source" is a $100-million-plus training run, and the last permissive checkpoint ages out of relevance in months. Model weights are the first artifact in the history of open licensing where the licensor's position *strengthens* over time, because the escape hatch that disciplined MongoDB and Elastic (credible community forks) decays with every new training run. The database companies pivoted their licenses and paid in forks and trust. The model companies found the same pivot where the fork is not a threat, only the trust is.

There is also a definitional fight underneath, two years old. When the Open Source Initiative published [version 1.0 of its Open Source AI Definition on October 28, 2024](https://opensource.org/blog/the-open-source-initiative-announces-the-release-of-the-industrys-first-open-source-ai-definition), it drew the line well above *any* of August's releases: open-source AI requires rights to use, study, modify, and share for *any purpose*, plus enough training-data information to substantially recreate the system. [Meta's Llama failed the test](https://www.theverge.com/2024/10/28/24281820/open-source-initiative-definition-artificial-intelligence-meta-llama) on day one, and Meta [publicly disputed the definition](https://www.axios.com/2024/10/29/meta-osi-definition-open-source-ai-llama) rather than comply. Two years later the market settled the argument by ignoring it: nobody in August's discourse asked whether Qwen3.8-Max meets OSAID, because the operative question stopped being *is it open source* and became *what does it cost me at scale*. The OSI built a definition; the labs built a rate card. The rate card won.

![Three license-pivot tracks — MySQL, MongoDB SSPL, Elastic — converge on 2026 model licenses, where forks cannot rebuild weights](/post-images/2026-08-12-the-toll-booth-on-the-open-road/three-precedents.jpg)

# The toll at $20M, $50M, and $500M: a worked example

Abstract thresholds become concrete when you run a real revenue curve through them. Take the most public one in AI tooling: Cursor's maker Anysphere went from roughly [$1 million to $100 million ARR in twelve months](https://sacra.com/research/cursor-at-100m-arr/), hit [$200 million by March 2025](https://sacra.com/research/cursor-at-200m-arr/), and passed [$500 million by June 2025](https://techcrunch.com/2025/06/05/cursors-anysphere-nabs-9-9b-valuation-soars-past-500m-arr/). Now run that curve through each of August's licenses.

On **Glimmer or LongCat (Apache/MIT)**: nothing happens, ever.

On **Kimi K3's terms**: the $20 million line falls somewhere in month eight or nine of the growth year. From that point, up to 30 percent of revenue is on the table — at a $100 million run rate, as much as $30 million a year, which at typical AI-tooling gross margins is most or all of gross profit. No venture-stage board accepts that; the real effect of the clause is that the company migrates off the weights *before* crossing the line, which is the point. The license is not designed to collect 30 percent. It is designed so nobody ever pays it.

On **Qwen3.8-Max's terms**: the $50 million revenue line arrives mid-curve, and the trigger is the *business description* — an AI-assistant company past $50 million must negotiate a separate commercial license with terms that are not public. Unpublished terms mean unbudgetable costs, and unbudgetable costs at diligence time mean a valuation discount. The toll here is not the fee; it is the uncertainty premium applied to a company whose core input is licensed at a rate one phone call in Hangzhou can change.

Run the curve and the design is visible: the thresholds sit where venture-backed companies *live*. A $20 million trigger catches a Series A company mid-growth; $50 million catches a Series B. The licenses are not aimed at hobbyists, enterprises, or hyperscalers. They are aimed at the next Cursor — the startup that would otherwise turn free weights into someone else's $10 billion serving business.

# The toll is set where a competitor would drive

Here is where the map departs from the take you will read everywhere else. The reflexive criticism — this is open-washing, these licenses betray the word "open" — is true as far as it goes and lazy where it stops. Of course a revenue-conditioned license is not open source; the July map already had a rung for "custom / research-restricted." The interesting question is not whether the label is earned. It is why the toll sits at these particular numbers, and the answer is that the numbers are rational to the point of elegance.

An open-weight release does two things for its maker, and the licenses have learned to price them separately.

The first is **distribution**: mindshare among builders, a fine-tuning ecosystem, placement in every "best local model" thread, and — for the Chinese labs — alignment with a national posture that made open weights [a doctrine with state-level backing](/posts/2026-07-18-the-giveaway-became-a-doctrine/). Distribution is worth paying for, and the labs pay for it by giving the weights away. Every hobbyist, researcher, and sub-threshold startup is subsidized on purpose. They are the marketing budget.

The second is **competition**: the risk that the giveaway arms the companies that would otherwise be API customers, or worse, rival serving businesses. This is the scenario the thresholds carve out. The licenses do not charge for *use*. They charge for *becoming the thing the maker sells*.

Map the payers and the design is unmistakable:

| Who you are | Apache / MIT (Glimmer, LongCat) | Kimi K3 custom | Qwen3.8-Max custom |
|---|---|---|---|
| Hobbyist, researcher | Free | Free | Free on paper (see below) |
| Startup below threshold | Free | Free below $20M sales | Free below $50M revenue |
| Enterprise above threshold, internal use | Free | Negotiable terrain — read the license | Untouched unless you sell MaaS / assistants |
| Model-as-a-service or assistant business above threshold | Free | Up to 30% revenue share | Separate commercial license, terms unpublished |

Hobbyists never pay. Enterprises sometimes pay, depending on what they sell. Rivals always pay. That is not a betrayal of the giveaway; it is the giveaway with the free-rider problem solved. Meta could afford stringless openness in the Llama era because it monetizes attention, not tokens. Moonshot monetizes tokens. A license that gives Kimi K3 to ten thousand startups while taxing the three companies that would resell it is not confused about what "open" means. It knows exactly what it means, and it is charging rent on the difference.

![Schematic map of license regimes — Apache/MIT, revenue-share, and promised zones with Glimmer, LongCat, Kimi K3, Qwen3.8-Max plotted.](/post-images/2026-08-12-the-toll-booth-on-the-open-road/license-regime-map.jpg)

# 1.2 terabytes is the other clause

There is a second toll booth in the Qwen3.8-Max release, written in hardware rather than legal English. A 2.4-trillion-parameter checkpoint that needs on the order of 1.2 terabytes of VRAM at 4-bit quantization is "free to run" the way a container ship is free to park in your driveway. [The spec sheets](https://aireiter.com/blog/qwen-3-8-max-open-weights) make the point without editorializing: the self-hosting audience for this model is measured in dozens of organizations worldwide, nearly all of them clouds, labs, or enterprises comfortably past the $50 million line.

Which means the license threshold and the hardware threshold point at the same buyer. The population that can physically run Qwen3.8-Max and the population that must negotiate a commercial license are nearly the same set. The hobbyist tier of this release is notional — a gesture that costs Alibaba nothing because it cannot be redeemed on consumer hardware. The genuinely runnable gift is the promised 27B model, which as of today has no repo, no model card, and no license text. Alibaba said "within about a week" on August 3. The week is up. On the July map's rules, a promise grades as a promise, and the 27B stays yellow until the checkpoint downloads and the license reads clean.

That split — a flagship whose openness is legally and physically conditional, plus a small model whose openness is promised and would actually matter to individual builders — is the shape to watch. If the 27B lands permissive, Alibaba will have built a two-tier openness machine: real distribution at the bottom, controlled distribution at the top, one flag over both.

# Apache is a strategy, not a charity

Meta's position on this map deserves its own reading, because Glimmer's clean Apache 2.0 license is doing strategic work too — just different work. As [we wrote Monday](/posts/2026-08-10-zuckerbergs-6500-word-bet/), Glimmer is a 30B dense model built to run on a single consumer GPU, under 20GB quantized. It is the only model on this map that a person with a gaming PC can run *and* legally build a business on without reading past the first paragraph of the license.

Meta can afford that because Meta has no inference business to protect at this scale. There is no Meta API whose customers a permissive 30B model would cannibalize; there is only the ecosystem gravity Llama once had and lost. Apache 2.0 is how you buy that gravity back — the license is the product. And note what Meta did *not* do: it did not open its larger models, and the manifesto's promise of an open Muse Spark remains a promise. Meta's openness has a size threshold just as surely as Kimi's has a revenue threshold.

So the honest summary of the field on August 12 is not "the West is open and China closed," nor the reverse, nor "everyone is open-washing." Every lab has drawn its toll line where its own business ends: Meta below 30B parameters and at the edge of its unreleased flagships, Moonshot at $20 million in sales, Alibaba at $50 million and a list of named categories. The word "open" now tells you almost nothing. The threshold tells you everything.

# The benchmark trended; the clause didn't

The discourse this week engaged with everything about these releases except the licenses — which is itself the finding. Glimmer's arrival got real traction: Artificial Analysis's release-day scoring [drew about 780 likes and 113,000 impressions](https://x.com/ArtificialAnlys/status/2086916150278111551), leading with the Apache 2.0 detail. The sharpest read of the window came from Benchmark's Chetan Puttagunta, who saw the monetization layer coming before the Qwen license text confirmed it:

> Incredible move by Meta. It will be interesting to see which open source license they use for Spark 1.2. There are now many clear paths to monetize open weights should Meta choose to.
>
> — [@chetanp](https://x.com/chetanp/status/2086867083867111740), Aug 10

MIT's Christian Catalini put the structural version in front of 197,000 impressions the next day:

> Frontier labs argue distillation threatens R&D and US national security. Open-weights proponents counter that diffusion is essential to competition and innovation. Both believe theirs is the only safe path. Luckily, the economics is loyal to neither.
>
> — [@ccatalini](https://x.com/ccatalini/status/2087177319253459019), Aug 11

The builders, meanwhile, were tracking the promise, not the terms. One post from a small account captured the mood on the eve of the weights drop:

> Qwen3.8-27B: the promised week is here. Alibaba said open weights "the week of Aug 10". Still no Hugging Face repo, no model card, no license.
>
> — [@Adidotdev](https://x.com/Adidotdev/status/2086990982286127538), Aug 11

And when the Max weights did land this morning, the release-day cheer led with parameters and context length; Hugging Face's Adina Yakup [flagged the license as the last line of the spec sheet](https://x.com/AdinaYakup/status/2087579467682304039). Honest accounting: the license-specific conversation through today is thin — the Reuters revenue-share exclusive, arguably the most consequential open-model story of the month, drew a fraction of the engagement of any benchmark screenshot. The crowd's operative question is still the July map's binary, can I run it or not. Nobody is asking the second question, what happens when I succeed — precisely the question these licenses were drafted to answer.

# Exact change only

> **The Deep Feed's position:** stop grading these licenses against the word "open" and start grading them as pricing documents, because that is what they are. The revenue-share license is not hypocrisy; it is the giveaway maturing into a business model that separates distribution (free, strategic, subsidized on purpose) from competition (identified by threshold, charged by design). The right response for a builder is not outrage. It is arithmetic: find the threshold, project your revenue, and know the day your free model stops being free.

The July map warned that "open-weight soon" borrows the credibility of "open" for free. August's sequel: "open weight" itself now borrows credibility from a license regime that no longer exists as a single thing. Between Glimmer's Apache and Qwen's Clause 2 lies the entire distance between a gift and a customs form, and both shipped inside the same ten days under the same adjective.

The road is genuinely open. The weights genuinely download. But the booth at the end takes exact change — it knows your revenue, it knows your product category, and it was positioned, to the dollar, at the lane where you would start to matter.

## Sources

- [Alibaba Cloud — Qwen3.8-Max: A New Bar for Coding and Cowork (Aug 3, 2026)](https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421)
- [Qwen — Qwen3.8 release blog (Aug 3, 2026)](https://qwen.ai/blog?id=qwen3.8)
- [Reuters — Alibaba plans to charge big users of its next open-source AI model, sources say (Aug 7, 2026)](https://www.reuters.com/business/retail-consumer/alibaba-plans-charge-big-users-its-next-open-source-ai-model-sources-say-2026-08-07/)
- [Hugging Face — Qwen/Qwen3.8-2.4T-A95B LICENSE file](https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/blob/main/LICENSE)
- [SCMP — Alibaba adds commercial restrictions to open-weight Qwen3.8-Max AI model](https://www.scmp.com/tech/tech-trends/article/3363927/alibaba-adds-commercial-restrictions-open-weight-qwen38-max-ai-model)
- [Plain AI — Qwen3.8's Clause 2 and the named products](https://plainai.tech/articles/qwen38-clause-two-named-products)
- [SQ Magazine — Qwen3.8 open weights, two licenses](https://sqmagazine.co.uk/qwen3-8-open-weights-two-licenses/)
- [aireiter — Qwen 3.8-Max open weights: the specs](https://aireiter.com/blog/qwen-3-8-max-open-weights)
- [Meta Research — Introducing Muse Glimmer, an open agentic model (Aug 10, 2026)](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model)
- [Meta — The Future is for Everyone (Aug 10, 2026)](https://about.fb.com/news/2026/08/the-future-is-for-everyone/)
- [The Register — Zuck rekindles open-weights Llama drama with Muse Glimmer (Aug 10, 2026)](https://www.theregister.com/ai-and-ml/2026/08/10/zuck-rekindles-open-weights-llama-drama-with-muse-glimmer/5285666)
- [OSI — The Open Source Initiative Announces the Release of the Industry's First Open Source AI Definition (Oct 28, 2024)](https://opensource.org/blog/the-open-source-initiative-announces-the-release-of-the-industrys-first-open-source-ai-definition)
- [The Verge — Open-source AI must reveal its training data, per new OSI definition (Oct 28, 2024)](https://www.theverge.com/2024/10/28/24281820/open-source-initiative-definition-artificial-intelligence-meta-llama)
- [Axios — Meta, OSI tussle over definition of open source AI (Oct 29, 2024)](https://www.axios.com/2024/10/29/meta-osi-definition-open-source-ai-llama)
- [MongoDB — MongoDB Issues New Server Side Public License (Oct 16, 2018)](https://www.mongodb.com/company/newsroom/press-releases/mongodb-issues-new-server-side-public-license-for-mongodb-community-server)
- [Elastic — Doubling down on open, Part II (Jan 14, 2021)](https://www.elastic.co/blog/licensing-change)
- [Elastic — Amazon: NOT OK, why we had to change Elastic licensing (Jan 19, 2021)](https://www.elastic.co/blog/why-license-change-aws)
- [AWS Open Source Blog — Stepping up for a truly open source Elasticsearch (Jan 2021)](https://aws.amazon.com/blogs/opensource/stepping-up-for-a-truly-open-source-elasticsearch/)
- [LWN — On the dual-license model (Feb 15, 2006)](https://lwn.net/Articles/172128/)
- [Sacra — Cursor at $100M ARR](https://sacra.com/research/cursor-at-100m-arr/)
- [Sacra — Cursor at $200M ARR (Mar 2025)](https://sacra.com/research/cursor-at-200m-arr/)
- [TechCrunch — Cursor's Anysphere nabs $9.9B valuation, soars past $500M ARR (Jun 5, 2025)](https://techcrunch.com/2025/06/05/cursors-anysphere-nabs-9-9b-valuation-soars-past-500m-arr/)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-08-12-the-toll-booth-on-the-open-road/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt