Skip to content
World clockEU--:--UK--:--USA--:--CN--:--PLDEFRIT中文EN

portal about AI and technologyevents · analysis · interviews · technical background

Search
LIVE
›

Open weights, closed wallets: what the latest open source releases reveal about who pays

Aleph Alpha published Kolibri on 3 October, a 78B-parameter model with 3B active parameters under Apache 2.0. Its blog frames open weights as a route to regulated customers, not a giveaway. Cloudflare said it is moving more Enterprise-only features, including Logpush, to free and pay-as-you-go tiers.

BusinessAnalysisDr. Amara PatelPublished: 3 October 20266 min readSources 9
Open weights, closed wallets: what the latest open source releases reveal about who pays

Aleph Alpha published Kolibri on 3 October. It is an English-German mixture-of-experts model with 78B total parameters and 3B active, a 1M-token context window, and full weights downloadable from Hugging Face under Apache 2.0. The company's blog calls it a "sovereign open-weight model" built for public administration, industrials and aerospace.

The pitch is not generosity. It is a sales argument aimed at buyers who cannot send data to a US inference API. Aleph Alpha says customers get full supply-chain integrity and can run the model on-premise. It also says it tuned Kolibri for German, reasoning, math and agentic behaviour. The company claims Kolibri matches models with up to four times its active parameter count, naming Nemotron 3 Super, on math, coding, grounding and long-context tasks. That comparison is the company's own and has not been independently checked.

Why it matters: an open licence is now the entry ticket to regulated, sovereignty-minded buyers, not a statement about community. For a procurement team, the question shifts from licence terms to who operates the model and where. Neither the download numbers nor Aleph Alpha's revenue from Kolibri are public, so the commercial return on the release is unknown.

Cloudflare gives away the plumbing

Cloudflare's move is the other half of the pattern: give away the tool, charge for volume. In a blog post, the company said Logpush is now available to Free, Pro and Business customers on self-service, pay-as-you-go pricing. Logpush ships logs to storage and analytics destinations. Transformers, which filters and redacts log records with SQL, is generally available to all customers, and Custom Dashboards is open to everyone.

The post ties this to a pledge CTO Dane Knecht made a year earlier to make every Cloudflare feature available to everyone. Cloudflare is blunt about the history: a two-tier system developed where its most advanced features sat behind Enterprise contracts. "Where our plans differ, it's in how much you can use, not what you can use," the blog says. Nearly all products now carry role-based access control at account and zone level. Workers joined R2 and Access in offering roles at the individual resource level.

That is metering by consumption, dressed as democratisation. Small teams get the feature; heavy users still pay as they grow. Cloudflare has not said what the change does to Enterprise renewals.

The counter-example: paid weights, open data

Not every open-flavoured release follows the same logic. Vortech, a three-person team, published a CS:GO dataset for multiplayer world models under CC BY-NC 4.0. It holds 3,958 hours of 5v5 match state and actions, 3,283 hours of point-of-view video and 218,768 rounds. The site says the data lands on Hugging Face and that the licence covers non-commercial research and education only.

Vortech's reasoning is visible in its own numbers. It generated the data from Source-engine bots and GOTV recordings, running 33 servers on one map for five days. "Gameplay generation scales linearly with compute," the team writes. The dataset is the free sample; the trained multiplayer world model stays in closed beta, and Vortech says it will release it "when it is good enough".

Kitsuno's engineering write-up shows the same tension from the buyer's side. Kitsuno runs an open decision model, Laya, from ConvAI Innovations, under Apache 2.0, and wants to replace a closed model it calls Jev. Out of the box, Laya answered 44 percent of Kitsuno's test questions correctly; Jev answered 94 percent. After training, Kitsuno reports Laya improving, and cites a comparison by AJ Awan putting Laya's typed-decision score at 0.362 untrained and 0.766 after fine-tuning.

"For us, the training is the product," Kitsuno writes.

The money is not the motive. Kitsuno says Jev handled 35,812 logged calls since 23 September for USD 3.68 in total. It is switching for two reasons: its own principle of open-source models, and European data residency. Kitsuno also says TypeSafe's customer agreement, section 2.3(b) as updated on 23 September, bars using Jev's output to train a competing model. So it trained Laya without a single Jev answer. That reading is Kitsuno's own, not legal advice.

Open research, closed chequebooks

Meta's research arm took a different route. It published six papers produced with mathematicians working alongside its Muse Spark models in Thinking Mode through the ordinary meta.ai chat interface. Five present answers to previously open questions. Meta says each paper marks which passages researchers drafted and which the model drafted, and credits earlier work. It also discloses that outside teams independently announced solutions to some of the same problems, including Misiakiewicz and Wen on the Gaussian threshold.

None of that is a product. It is reputation spending. Publishing research openly costs little and buys credibility with the scientists Meta wants to hire.

The open-source tooling wave carries the same ambiguity. A framework called e2e says it raised $8m to build verification software for coding agents. Its site insists the framework "is open source and stays free", with users running it on their own machines and CI. The money, presumably, comes from somewhere other than the licence.

Then there is the hobbyist tier, where open source is a gift with no business model at all. Meta's own Muse gadget SDK is published on GitHub under Apache 2.0, with some third-party files under their upstream licences. It lets people flash an ESP32 board or a Raspberry Pi and pair it with the Muse app. The README warns of "bricked boards, voided warranties, brownouts, or bankruptcies". Smaller projects fill the same shelf. There is a native iOS push plugin for OpenClaw under MIT, with a paid Pro edition on Gumroad, and an Amiga optical-disc filesystem released as open source with benchmark claims against six older filesystems.

What connects these is not ideology but distribution cost. Code that runs on a user's own hardware costs the publisher almost nothing to hand out. It creates demand for something adjacent: a paid tier, a support contract, a token, a cloud account. Cloudflare's log pipeline is the cleanest example, because the free feature feeds a metered service.

What the model says about the money

For buyers, the practical test is no longer whether a licence says Apache 2.0 or MIT. It is whether the model, the weights or the service can be taken away. Kolibri's weights can be downloaded and kept. Cloudflare's Logpush cannot leave Cloudflare. Laya can be trained on a customer's own questions, which is exactly why Kitsuno wants it.

For competitors, the pressure runs the other way. A closed model priced per call has to beat a free download by enough to justify the invoice. Kitsuno's numbers suggest the gap can be closed with training data rather than licence fees. Aleph Alpha's claim to match larger models at 3B active parameters is the kind of assertion that, if it holds, makes expensive inference harder to sell in regulated sectors.

What remains unknown is the revenue. Aleph Alpha has not said what Kolibri earns. Cloudflare has not put a figure on the Enterprise features it just gave away. Vortech has not said when, or whether, its world model ships commercially. Meta's papers carry no price at all.

The pattern is consistent enough to name: open weights are becoming a customer acquisition channel for companies that sell something else. Does that sustain the companies doing it, or merely shift the bill to whoever needs scale? The next round of releases will answer.

Comments 0

Sources

9
  1. 01Kolibri Has Landed: A Sovereign Open-Weight ModelEN
  2. 02Updates on our pledge to make Cloudflare features accessible to everyoneEN
  3. 03Open 10-player CS:GO dataset for multiplayer world modelsEN
  4. 04Laya in shadow: training an open decision model to replace JevEN
  5. 05Solving Open Research Problems TogetherEN
  6. 06E2E: The open source AI testing frameworkEN
  7. 07Open source SDK to build Muse gadgetsEN
  8. 08Native iOS push notifications for OpenClaw agents (no PWA)EN
  9. 09A modern open-source Amiga CDFileSystem replacementEN

All figures and quotations in this text come from the sources listed below.

Content prepared by the editorial team with AI assistance.

Dr. Amara Patel

Dr. Amara Patel

Economy, business and world

Dr. Amara Patel covers business, world affairs and the economy for FLASH24, working from filings, central bank statements and trade data rather than press releases, and she does not let company spin stand in for numbers. She checks revenue recognition, debt covenants and currency effects line by line against audited reports and regulatory disclosures. Her week includes calls with analysts, logistics operators and trade lawyers, and she watches the calendar for rate decisions, earnings dates and port and freight updates, comparing each against prior quarters. Outside the desk she tracks tech-company accounts and rides cargo bikes, which keeps her close to both the balance sheets she reads and the supply chains she covers. She does not publish a figure she cannot trace to a primary document.

Newsroom →

Comments

0
  1. No comments yet — be the first.

Write a comment

Comments are public. We do not publish abuse, spam or advertising.