Seven open releases, two licence traps
What OpenAI, xAI, NVIDIA, Anthropic, Cloudflare, Tencent and Alibaba put on GitHub — and which ones you can actually bill for
Seven things the big labs open-sourced, all of them running on your own hardware rather than someone else’s API. Five are plain Apache 2.0 and you can put them straight into client work. The other two are where the actual question lives: one ships MIT code with separately licensed weights, and one excludes the EU, the UK and South Korea outright.
Every one of these is an official release from the company that built it, and every one runs on your own machine or your own cloud account — no per-seat API bill sitting between you and a client. That makes the licence the thing that decides whether a given one is a business or a toy, which is exactly the part these round-ups skip. Two of the seven carry real restrictions, and one of them is geographic.
The seven, verified against each repository on 20 September 2026
Licences read from the LICENSE file in each repo rather than taken from the write-up.
Tencent’s licence opens by stating in capitals that it does not apply in the European Union, the United Kingdom or South Korea, and it requires a separate licence from Tencent once a product built on it passes one million monthly active users. If the client is European or British, this one is off the table regardless of what you charge. NVIDIA’s split is gentler but still a split: the PersonaPlex code is MIT, while the weights are under the NVIDIA Open Model licence and gated — you accept the terms on the model page before Hugging Face will hand them over. Everything else here is plain Apache 2.0.
The usual reading is that negative signals are catastrophic — a report is weighted hundreds of times a like, so one report wipes out hundreds of likes. xAI’s README calls this out directly as a misconception. The weights scale the model’s predicted probability that *you specifically* would take that action, not raw counts of who did. So a high report weight does not mean one report cancels 468 likes; it means the model’s estimate of how likely you are to report something is worth a lot in your own score, and that estimate is driven largely by your own past behaviour. xAI says it added comments to the ranking code precisely so that people and LLMs reading it stop drawing the wrong conclusion — which is worth knowing before you sell strategy built on the wrong mental model.
What each one turned out to include
Symphony expects a codebase already set up for agents
The README is explicit that it works best in codebases that have adopted harness engineering — it is the step after managing coding agents, not a way to start. Proof of work is richer than the usual summary suggests: CI status, PR review feedback, complexity analysis and walkthrough videos. OpenAI labels it a low-key engineering preview for trusted environments, so every change still gets reviewed before it merges.
Two retrieval paths, then one model ranks both
In-network posts come from Thunder, which keeps recent posts from accounts you follow in memory. Out-of-network comes from Phoenix retrieval and SimClusters. Both streams are then ranked by the same model, which predicts a probability for each action you might take and blends them with weights held in the code. Whether a post can be shown at all is decided separately, by visibility filtering, from your own blocks and mutes and from labels other systems attach.
PersonaPlex is a fine-tune of Moshi, not a new architecture
The README says so plainly: it is based on the Moshi architecture and weights, benefiting from the Helium model underneath. That is not a criticism — it is why the full-duplex behaviour works — but it means the thing to read before betting on it is Moshi’s literature, and it explains why the server command is moshi.server rather than something NVIDIA-branded.
Six plugins, not four, and one of them is load-bearing
Alongside the pitch agent, GL reconciler and market researcher there are investment-banking and equity-research plugins. More importantly financial-analysis is the core: it carries the shared modelling skills and all eleven data connectors the others rely on, so it goes on first rather than as an afterthought.
Gatekeepers solve the approval problem, which is the actual idea
Described as guardrails, they are more specific than that. When an agent does something needing approval, the Gatekeeper simulates the result locally and tells the agent it succeeded, so the agent keeps queueing work instead of blocking. You approve or reject the queue later, in bulk. That is a direct answer to the choice between babysitting every action and giving up and passing --dangerously-skip-permissions. Worth knowing the repo is version 2, a complete rewrite, and Cloudflare calls the local mode a way to try it rather than to run on.
The hardware floor is exact, the speed claim is not
The README states 6 GB of VRAM for shape generation and 16 GB for shape plus texture — those numbers are verbatim, and they are the reason texture work needs a serious card. The widely quoted sub-second figure does not appear anywhere in the repository, so treat it as marketing until someone shows a benchmark; what the repo does document is the Turbo and FlashVDM distilled variants as the fast path.
Ten languages, and a three-second clone
The ten are Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian, listed per model rather than in aggregate. The base models do voice cloning from roughly three seconds of audio, the custom-voice models ship nine premium timbres, and everything streams. Sizes are 1.7B and 0.6B, so this is the lightest thing on the list.
PersonaPlex — the demo server
Needs an NVIDIA GPU and Linux. Accept the licence on the model page before the token will work.
Cloudflare OS — locally
The finance plugins
Hunyuan3D-2 — photo to model
Qwen3-TTS
What the source suggests charging
These are the write-up’s own suggested ranges, not observed market rates. Included because the licence question only matters if you are billing for it.