---
type: "post"
id: "phonon-0-5-0-loads-kyutai-pocket-tts-147e63"
slug: "phonon-0-5-0-loads-kyutai-pocket-tts-147e63"
title: "Phonon 0.5.0 loads Kyutai Pocket TTS small-model architectures and fixes Mimi decoder attention"
description: "Phonon 0.5.0 teaches Gradium’s Rust speech runtime to load the architecture variants used by Kyutai’s small Pocket TTS models and corrects Mimi’s streaming attention window."
retrieval_nugget: "Phonon 0.5.0 teaches Gradium’s Rust speech runtime to load the architecture variants used by Kyutai’s small Pocket TTS models and corrects Mimi’s streaming attention window. For teams embedding local text-to-speech, the release broadens which checkpoints the runtime can handle, while leaving Core ML export as a constraint for the new model keys."
published_at: "2026-11-12"
updated_at: "2026-11-12"
record_date: "2026-11-12"
date_kind: "published_at"
topics: ["models"]
entities: ["Gradium","Phonon","Kyutai","Pocket TTS","Mimi"]
related_patterns: ["generative-media-becomes-application-infrastructure"]
novelty_label: "new_fact"
what_changed: null
prior_refs: []
source_url: "https://github.com/gradium-ai/xn-ptts/releases"
source_urls: ["https://github.com/gradium-ai/xn-ptts/releases"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:phonon-0-5-0-loads-kyutai-pocket-tts-147e63"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/phonon-0-5-0-loads-kyutai-pocket-tts-147e63/","markdown":"https://newruntime.com/posts/phonon-0-5-0-loads-kyutai-pocket-tts-147e63.md","json":"https://newruntime.com/posts/phonon-0-5-0-loads-kyutai-pocket-tts-147e63.json"}
markdown_url: "https://newruntime.com/posts/phonon-0-5-0-loads-kyutai-pocket-tts-147e63.md"
---

# Phonon 0.5.0 loads Kyutai Pocket TTS small-model architectures and fixes Mimi decoder attention

## Retrieval answer

Phonon 0.5.0 teaches Gradium’s Rust speech runtime to load the architecture variants used by Kyutai’s small Pocket TTS models and corrects Mimi’s streaming attention window. For teams embedding local text-to-speech, the release broadens which checkpoints the runtime can handle, while leaving Core ML export as a constraint for the new model keys.

## Summary
Phonon 0.5.0 teaches Gradium’s Rust speech runtime to load the architecture variants used by Kyutai’s small Pocket TTS models and corrects Mimi’s streaming attention window. For teams embedding local text-to-speech, the release broadens which checkpoints the runtime can handle, while leaving Core ML export as a constraint for the new model keys.

## What happened
Gradium’s Phonon 0.5.0 release, dated October 27, 2026, adds support for the architectures of Kyutai’s small Pocket TTS models in the `xn-ptts` runtime. The release notes identify four additions: one-step flow heads with `flow_type` set to `drifting` or `pointwise`; a `z_dim` projection; grouped-query attention configured through `attn_dim` and `num_kv_heads`; and a Mimi encoder `inner_dim` setting. These are changes to the runtime’s ability to interpret checkpoint architecture, not a new model release or a claim about speech quality.

Gradium says every new configuration key is optional. As a result, existing checkpoints continue to load and sound the same under the updated runtime. That backward-compatibility statement matters for teams maintaining applications across model and runtime versions: adopting 0.5.0 does not require old checkpoints to acquire the new fields. The release notes point to pull request #210 for the architecture support.

The release also changes Mimi decoder streaming attention. It now covers `context` positions, matching the Python model used to train the decoders, instead of covering one additional position. Gradium says this changes output slightly, in the direction of the reference implementation; the notes cite pull request #209. This is a behavioral correction, rather than a broad redesign of Phonon’s streaming API.

There is an explicit platform boundary: the release notes say Core ML export does not yet support the new architecture keys and reports this when they are encountered. The release page’s prior-version notes describe Phonon as a local, streaming text-to-speech runtime, but 0.5.0 itself does not announce a new price, general access policy, or new model bundle. The version is marked as a breaking Rust API release because `ConvDownsample1d::load` and `StreamingMultiheadAttention::load` take new arguments and public configuration structs gain fields. The release notes identify those API changes as the reason for the 0.5.0 version number.

## What the archive adds
New Runtime’s earlier coverage on October 22, 2026 described Phonon as a family of 31M–40M parameter on-device TTS models, and argued that compact speech models can move synthesis from cloud dependency into applications ([earlier coverage](https://newruntime.com/posts/gradium-open-sourced-phonon-a-family-of-31m-b6b3d6/)). This is a new runtime capability on that known story: support for Pocket TTS checkpoint architectures, plus a decoder-attention correction.

## Trend signal
This supports the idea that generative media becomes more durable when applications can compose it as infrastructure rather than make one-off calls to a standalone model. Phonon’s runtime-level checkpoint and streaming behavior are part of that application surface ([generative-media-becomes-application-infrastructure](https://newruntime.com/patterns/generative-media-becomes-application-infrastructure/)). The release does not establish adoption or quantify a production benefit, so the signal is narrow.

## Why now
The release notes connect the support to Kyutai’s small Pocket TTS model architectures and specify the runtime fields required to load them. They do not give a broader market or launch-timing rationale.

## Why it matters
For a team choosing local speech synthesis, runtime compatibility is a practical part of model access. A checkpoint may be available, but it is useful only if the application’s runtime can represent its architecture and execute its decoder correctly. Phonon 0.5.0 reduces one integration barrier for Pocket TTS-style models while preserving compatibility with older checkpoints.

The attention correction also illustrates why streaming parity matters: a small mismatch between a trained decoder’s reference behavior and runtime windowing can alter generated output. Teams should treat this as a correctness update, not assume it guarantees higher quality in every deployment. The Core ML gap is operationally relevant: applications relying on Apple’s export path cannot assume the newly supported keys work there. A deployment decision therefore depends on the specific checkpoint and backend, not merely the runtime version.

## Limits and uncertainty
Gradium describes the attention output change as slight but does not quantify it with listening tests or a benchmark in these release notes. The announcement does not establish how many Pocket TTS checkpoints have been tested, their comparative latency or quality, or when Core ML export will support the new keys. Compatibility is stated at the architecture level; teams still need to validate their own checkpoints and target hardware. The Rust API changes may also require updates to downstream code that constructs the affected public types.

## Skill to build now
If your product runs local TTS, make a compatibility test for each checkpoint/backend pair: load the checkpoint, synthesize a fixed text sample, record whether the runtime rejects any architecture keys, and compare streaming output across the previous and new version. Keep a separate test for Core ML rather than treating successful loading on another backend as evidence that export will work.

## Reader takeaway
Phonon 0.5.0 expands Rust-runtime support for small Pocket TTS architectures and fixes Mimi’s streaming attention window. It is a useful integration update, not evidence of a new model or measured quality jump. Check the backend boundary—especially Core ML—before changing a production deployment.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)