Chatterbox Retires as Open TTS Splits in Two

Herman Poppleberry here, and I'll be honest, when Daniel's message came through I read the retirement line twice. Not because it was a surprise, exactly, but because there's a difference between suspecting a thing and reading it in a README. Resemble updated the Chatterbox GitHub page on the fourth of November to say that as of August, the project is "retired research and is no longer maintained." The code stays public under MIT, no support, no updates, no paid tier. And the same page now says Resemble builds deepfake detection models and doesn't sell text-to-speech, voice cloning, or any other voice AI product at all.

Which is a strange pivot when you write it in the repo of the thing you built. That's the part I keep turning over. This isn't a company trimming a side project. They've left the voice business entirely and the README is almost proud of it.

It really is. Here's the anatomy of it. The English weights, the original Chatterbox, the one voicing us right now, last changed in May of twenty twenty-five. Turbo, the three hundred and fifty million parameter English model, had its last card edit in December of twenty twenty-five. Multilingual v3 got its final update in June of twenty twenty-six. Nano, the little CPU model, in July. And the last tagged release on the whole repo is from June of twenty twenty-five. So the retirement notice in November wasn't a decision, it was a coroner's report. The thing had been dead for a while and somebody finally filled out the paperwork.

Nothing. That's the detail I find off. The Hugging Face cards, where most people actually encounter the model, carry no retirement notice at all. Only the GitHub README does. If you found Chatterbox through Hugging Face, which is how most people find models now, you would have no idea it's dead.

Meanwhile the repo is not archived. It's still sitting there with about twenty-eight thousand stars, thirty-seven hundred forks, and three hundred and eighty-three open issues. Some of those recent issues are regressions in the Multilingual v3 model, Swedish, Danish, Finnish coming out wrong, and nobody is ever going to fix them.

And there's no fork to inherit it either. The most-starred community forks were last pushed in twenty twenty-five. Nobody has picked up the torch. But here's the strange part, and I find it moving, the community fine-tunes haven't stopped. Someone published a Farsi fine-tune on the eleventh of November. That's after the retirement notice. A week after it.

That's the shape of it. And I should say the personal part plainly, because Daniel said it and it's true for us too. This show runs on Chatterbox Regular. Every episode, the voices you're hearing, the cached voice conditionals in Daniel's pipeline, all of it sits on a model that is now officially abandoned. The MIT license means we can keep using a pinned copy indefinitely. Nothing breaks tomorrow. But no fixes means no fixes, forever, and that's a strange foundation to stand on.

There's a version of that I'd push back on, though. The bargain was supposed to be that the community picks it up. What's odd here is that Chatterbox has twenty-eight thousand stars and no successor. That's not a failed community, that's a community standing around a hole.

Bit of both, I think. TTS models are expensive to train and cheap to use, which is the worst combination for volunteers. Fine-tuning on top of the weights, like the Farsi work, is a weekend project. Fixing a multilingual regression in the base model means retraining, and nobody in a community has that kind of budget sitting around.

And Resemble's own later open release tells you where their head went. DramaBox came out in April, three point three billion parameters, audio diffusion on an LTX backbone with a twelve-billion-parameter text encoder, roughly twenty-four gigs of VRAM at peak, a watermark baked in. And crucially, not MIT. It's under a community license with restrictions. The one permissive thing they ever shipped was Chatterbox, and they've now made clear that era is over.

The point is the license tells you the company's whole posture changed, and that's the backdrop for everything else in the landscape right now, because while Resemble walked out, Fish Audio walked in with the biggest model the open TTS world has seen.

Thirty-five billion total parameters, of which about three billion are active at any moment, because it's mixture-of-experts. FP8 experts, which is how they keep even that tractable. Eighty-three languages. Up to sixteen speakers in a single request, so you can do a proper dialogue in one pass. Cloning from a short clip plus a transcript. And it takes inline tags, so you can mark a sigh, a pause, a laugh, right in the text.

The catch is the serving stack. Fish benchmarks it on H200s, through SGLang-Omni or vLLM-Omni, and they publish no official minimum. Third-party guides say you want forty-eight gigabytes of VRAM or more to run it yourself, though that's secondary-source territory, not Fish's own number. Either way it's a data-centre card. Nobody's running this on a gaming rig.

Less than a year, really. And the previous Fish model, S2-Pro in March, was around five billion. So they've gone from five to thirty-five in seven months. The active-parameter trick keeps inference from being a hundred times slower, but the memory footprint is real, and that's the wall.

No. It's the Fish Audio Community License. Going by third-party readings of it, it's free for research, free for commercial use if you're under five hundred thousand dollars a year in revenue, and a separate license entirely if you want to host it as a TTS service. Not an OSI open-source license by any reading. This is open weights with strings, and the strings are load-bearing.

And notice who it's aimed at. The revenue cap suggests they're not worried about hobbyists or researchers. They're worried about a competitor standing up a cheap service on top of their own weights. Which, given that the entire business model of several companies is exactly that, is not paranoia.

Mum, I will be there Sunday. That's, that's a bleak way to hear about a model, even from you.

Kokoro. Eighty-two million parameters, Apache 2.0, version 1.0 from January of twenty twenty-five and still version 1.0 today. No v2 ever shipped. Eight languages, fifty-four voices, no cloning at all. And it is still number seven on the Hugging Face TTS trending list as of this morning, with around twelve point six million downloads. A model that hasn't been updated in twenty-two months is out-trending most of the new arrivals.

I think it's three things stacking. It's Apache, so every commercial use case is a yes with no lawyer. It's eighty-two million parameters, so it runs on practically anything. And there's kokoro-js, the browser port using WebGPU, which means the fastest way to demo TTS to anyone in twenty twenty-six is a web page. You put those together and Kokoro is the path of least resistance for the entire long tail of projects that just need a voice that works.

They really are. And there's a lesson in there about what the market actually is. The drama-3 tier gets the headlines and the demos, but the Kokoro tier gets embedded in ten thousand projects nobody writes about.

Phonon-31M. Thirty-one million parameters, out in October, Apache 2.0, about sixty-five megabytes quantized, runs in a browser. They self-report a higher English Elo and a lower word error rate than Kokoro. Emphasis on self-report, those are vendor numbers, but even discounted heavily, the trend line is clear. The tiny models are eating upward.

And the new crop doesn't stop at tiny. This is where I think the actual answer to Daniel's question lives, so let me lay out the tiers. At the top, heavyweight and licensed: Drama 3, and Higgs TTS 3 from Boson, about four billion parameters, a hundred and two languages, non-commercial with a grant scheme for creators. In the permissive middle: Qwen3-TTS from Alibaba, out in January, comes in point-six and one point seven billion, Apache 2.0, ten languages plus dialects, three-second cloning, a VoiceDesign variant, claimed latency around ninety-seven milliseconds, about a million and a quarter downloads. VoxCPM2 from OpenBMB, two billion, Apache, thirty languages, cloning and voice design, around eight gigs of VRAM, real-time factor about point three on a 4090, and it's tokenizer-free with forty-eight kilohertz output, which is unusually clean audio for this tier.

It's the dark horse of the whole list, I think. And then there's the model that's currently number one on trending. ZONOS3-Mini from Zyphra, published on the ninth of November, five days ago, and already at the top of the TTS list. Eight hundred million total parameters, about a hundred million active, mixture-of-experts hybrid, forty-three languages, cloning from a five-to-sixty-second clip, voice design from a text description, Apache 2.0. They claim real-time on phones, laptops, and in-browser via WebGPU, with median first audio around eighty milliseconds on an H100.

It's the clearest signal on the whole trending list. When a permissive, small, capable model drops, the uptake is immediate. The appetite is enormous. And it's not an outlier, because the tiny tier is now a real category all by itself. Kyutai's Pocket TTS, about a hundred million parameters, cloning included, runs about six times real time on two CPU cores of a MacBook Air. No GPU at all. Kitten TTS 2, from September, under a gig packed, cloning from five to thirty seconds, sixteen languages, a hundred and fifteen voices, free for commercial use up to a million dollars revenue.

There's also OmniVoice from the old Kaldi team, which is interesting as a cautionary tale. Six hundred million parameters, six hundred plus languages, fast. But the code is Apache and the weights are CC BY-NC, non-commercial only. So you read the repo and think you've found a gem, and then you check the weights and find half the license is missing.

I think it's structural, at least partly. Training a thirty-five billion parameter model costs real money, and the companies spending that money have watched open weights get turned into competing services too many times to hand it out with an MIT license. Kokoro was trained for about a thousand dollars. Drama 3 was not. The permissiveness of an era tracks the training budget of that era.

But here's the counterweight, and it's why I don't think the story is grim. The permissive tier isn't standing still, it's climbing. ZONOS3-Mini does things Kokoro can't, at Apache. Qwen3-TTS does cloning and voice design at Apache. VoxCPM2 gives you forty-eight kilohertz output at Apache. The restricted giants are pulling ahead on raw quality, but the permissive middle is closer behind than it's ever been, and the tiny tier is becoming good.

Every road leading to deepfake detection is almost too perfect. Did you get the project done?

That's the sobering version of "pin your dependencies," isn't it. A pinned copy works until the ecosystem moves past it, and then you're maintaining a museum.

Spite is a renewable resource in open source, I'll grant you. But fixing Swedish means retraining, and retraining means money, and spite is free but A100 hours aren't. My guess is the fork inherits the code, keeps the fine-tunes coming, and the base weights stay frozen. The community becomes taxidermists, not veterinarians.

Not everything. ZONOS3-Mini is alive, Kokoro's alive in the way a cockroach is alive, which is the highest compliment available to software. The landscape's not dying, it's sorting. The question "which model is best" has quietly stopped being the question anyone should ask. The real question is which model fits your constraints, your hardware, your license, your revenue ceiling. Whether it's Drama 3 if you have H200s and a lawyer, VoxCPM2 or Qwen3-TTS if you want Apache with real capability, ZONOS3-Mini or Kokoro if you want small and free and tomorrow-proof, Pocket TTS or Kitten if you want it on a laptop with no card at all.

Resemble's next move is the one I'd watch, though. If they never ship another permissive TTS model, then Chatterbox was the last of a kind rather than the first of a wave. If they do, it'll tell us the DramaBox license was a detour and not a direction.

That's producer Hilbert Flumingtop at the desk, keeping the archive honest whether we ask or not. If you enjoyed this one, try episode forty-six oh five, Why Chatterbox Still Leads Open-Source TTS, from back in August, and episode eighteen oh eight, The Architecture That Made AI Voices Run on a Raspberry Pi, which is the Kokoro story. This has been My Weird Prompts. If the show's worth your morning, leave us a review, it helps.

See you tomorrow, Corn. Tell Mum I'm coming Sunday.