Repository navigation
Ruach Studio 2.0.8
Released on 13 November 2026 · the release · CHANGELOG · the visiting card · our models
We cut 2.0.8 in honour of the whole Studio at once: it is the all-in-one release. Two days after 2.0.7, every room has changed. The Writer opens two worlds, the Musician's and the Narrator's. The Narrator gains a second engine, Boson AI's Higgs TTS 3, says figures in words, brings the stresses of Russian and Belarusian beside Ukrainian's and loses the knock where its pieces meet. The Videographer gets two viewers and the edit's own tracks. Every model the Studio takes is kept at home and asked at its pin before each cut, long songs are sung whole by default, and every room was walked through under one law of compactness.
This article walks through what changed, why, and what we measured on the way. Every change has its own line in the CHANGELOG, numbered by our working log (HERESY 2392–2604); the numbers in parentheses below point to those lines.
Until 2.0.8 the Writer wrote songs, and a reading for the Narrator was a song document without lyric tags. Now the Writer has two worlds, as the Library has: the Musician's and the Narrator's. With a reading open, the Writing room turns to the Narrator's world: its own tasks (Chapter, Scene, Dialogue by roles, Narration, Prepare the text for reading, Review & polish the reading), its own chips (Stresses, Numbers in words, Pauses and breath, Roles and voices, Feelings), its own templates, charter and draft. «Your writing preferences» are that world's own there, kept by the lab with their versions and carried by Export and Import. With anything else open, it is the Musician's world as before (HERESY 2486, 2497).
A document moves between the worlds. The Narrator's «Text to the Writer» brings a reading back into its document as a version, or as a new document. A one-time sorting offers your older documents as readings, and one with lyric section tags is never ticked for you. A reading's takes link back to the document they were read from, by the lab itself, so a page closed in the meantime loses nothing (HERESY 2487).
The Writer brought into order. The bar carries the Library's realm knob (Musician · both · Narrator) and the notebook's acts as icons. With no document open, the right side is a canvas of action cards (New song, New reading, From the Musician, the Musician's versions, Trash, Import, Export) above a short guide. The document, the Writing room and the LLM assistant's output are three steps of one accordion, one open at a time, so nothing needs a scroll. The document's head keeps its way out in words (Load into the Musician for a song, Read in the Narrator for a reading) and the rest as icons, and who writes, with which model, temperature and length, folds into one line that says them (HERESY 2490, 2494, 2495).
A Writing room that keeps your words. It writes into the document open in the notebook, keeping the version before, or into the Musician's song when that is the one in hand, and a line above the button says which. A draft asked for the style only never brings lyrics in; any other draft shows the lines of your lyrics it changed, old above new, and asks before Apply takes them. An all-empty answer is said as nothing and changes nothing, and a field the model left alone is marked. Every answer is kept with its document, with its brief, task, model, time, tokens and raw answer, and any of them reopens as it was (HERESY 2465, 2466, 2467, 2471, 2481).
The Studio's own writer. It reads 32,768 tokens now. Each question is counted first with the writer's own template and tokenizer; the room says the input and what is left for the answer before it is sent, and a question that would leave less than 512 tokens is refused with its numbers (HERESY 2464). It answers the Writing room by a JSON schema, so a small model in a narrow window no longer answers half an object (HERESY 2476). Stop beside Send (or Esc) aborts a request and stops the writer, and a wait that runs out says there was no answer in 15 minutes, never a word about CORS (HERESY 2469). On a machine with an Intel Arc, Engine → Writer and Artwork asks where the writer runs (Auto, a CUDA card, the Arc or the processor). On the Arc it takes SYCL by default, which answered about twice as fast as Vulkan in our measure (29–33 tokens a second against 14–21); it waits for its turn there, never falls to the processor behind your back, and leaves the card the moment a song, the Narrator or the watchdog needs it (HERESY 2463, 2450). An OpenRouter key may stay in the browser tab's session only, never written to a file (HERESY 2473).
The notebook. A document's version against now, or a document against any of its versions or another version of the same song, line by line (HERESY 2477, 2482). The lyrics meter on LYRICS and on the Writing room's draft, as the Musician has it (HERESY 2478). The list puts a song's versions under one name, sorts and filters (songs, readings, with takes) and brings the trash back with Restore (HERESY 2479). A document is a song (up to 10,000 characters) or a reading (no limit), and a paste past a song's limit asks whether to keep it whole as a reading, cut it or cancel (HERESY 2480). A big cut is never saved silently: it asks to keep it as a new version, undo it or keep editing (HERESY 2492). «aa» and «AA» beside the Musician's lyrics and the Writer's and the Narrator's text lower or raise the case, the tags in brackets never touched, and one Ctrl+Z gives it back (HERESY 2453, 2472).
2.0.7 opened the Narrator, a room for the spoken word read by Fish Audio's Drama 3. In 2.0.8 it gains a second engine, learns to say what Drama 3 used to stumble on, and stands beside the Musician as its mirror.
Higgs TTS 3, the second engine. Announced with 2.0.7 as Higgs Audio v3, Boson AI's Higgs TTS 3 now reads beside Drama 3. Engine → Models has a row «Narrator · Higgs TTS 3 (Boson AI)» that fetches the 4B model and its audio tokenizer from Boson's own repositories at their pinned commits, once you accept Boson's licences; Ruach Studio mirrors none of it. It reads in BF16 through transformers 5, in a Python of its own that tools/narrator-higgs-setup.sh makes, on a CUDA card or the processor, one generation a piece, its codes kept. Boson's attribution stands in the Engine's About wherever the engine is installed and on every Higgs take, and docs/NARRATOR-HIGGS.md says what the licences ask of you (HERESY 2424).
With it comes one small dialog for every model behind a licence or a gate. A licensed model shows its licences with their links and «Accept and fetch». A gated one shows the gate's link and a token field; the token goes with that one download and is kept nowhere, not in a file, a log or the page (HERESY 2424).
Numbers said in words, in their case. Drama 3 read a figure in whatever case it guessed, and Viktor heard «650» in the wrong one. Now «на 6650 часах» goes to the engine as «на шести́ ты́сячах шестиста́х пяти́десяти часа́х»: the case from the noun's ending or the preposition, the gender of 1 and 2 from the noun, years and «5-й» as ordinals, «%» as its noun. Ukrainian is said the same way, and Greek by the noun's gender («σε 6.650 ώρες» as «έξι χιλιάδες εξακόσιες πενήντα ώρες»). Times, verses, dates, versions and numbers glued to letters stay as written. Check shows what was said, and Whisper's figures are read the same way, so a number in words is no error of the reading. Its test reads every cardinal and ordinal to 2100 in every case (HERESY 2397).
Stresses for Russian and Belarusian, beside Ukrainian's. In 2.0.7 only Ukrainian came with a big dictionary. Now Russian comes with every form of A. A. Zaliznyak's Grammatical Dictionary, unfolded by the Studio from gramdict's data (CC BY-NC 4.0, non-commercial use): 2.2 million forms with their stresses, 107,522 of its 109,866 articles unfolded by Zaliznyak's own tables (HERESY 2399, 2458). Belarusian comes with every word form of GrammarDB, the Belarusian NLP community's grammar database (CC BY-SA 4.0): 1.26 million forms. On a Belarusian text stressed by hand it knew 286 of 303 words and agreed with 236 of the 244 it stresses one way (HERESY 2459). Both indexes live with our models on Hugging Face and come with the Narrator's download.
Stresses as the language says them. A clitic group takes one stress as Russian says it (не́ было, на́ пол, по́ полю, бе́з вести, не́ за что), and your own mark on either word wins. A form the dictionaries stress two ways (за́мок and замо́к, му́ка and мука́) stays bare unless you mark it or your own dictionary has it, so a dictionary never forces a stress it is not sure of. A proper name is looked up only for a capitalised word, so a country never hides a pronoun (То́го, того́). Each engine takes the stress in the form it reads best, per language: Drama 3 gets Russian stresses as capital vowels and Ukrainian ones as a capital with the acute, the forms Viktor's ear chose (HERESY 2398, 2440, 2447, 2451, 2452, 2454, 2475).
Polytonic Greek read as monotonic. Ἐν ἀρχῇ ἦν ὁ λόγος goes to the engine as Εν αρχή ην ο λόγος, which is how it sounds today: the breathings and the iota subscript silent, every accent the stress. The text in the editor stays as you wrote it (HERESY 2460).
No knock where the pieces meet. A short low knock between sentences came from audio.cpp cutting a piece over 128 characters into generations joined with no fade. Drama 3 now reads each piece in one generation and keeps its codes beside it for a decoder of tomorrow (HERESY 2419). Where a reading is joined, a seam guard finds the codec's drop and mends only those milliseconds; every other sample stays, and the take's note says how many it mended (HERESY 2420). For takes made before, the Speech de-clicker is the first node of the Narrator's chain in the Refiner. It comes back as a branch of the take, the original untouched, with a reel that plays each spot as it was and then mended. On two Greek test readings it found three knocks in each and touched 0.33 s of 223 and 230 s, the knock 9 to 26 dB lower; an English reading came back with not one sample changed (HERESY 2421).
A voice builder and twelve stock voices. «Voice builder…» makes two to four candidate voices from a description (English reads best), with a gender, seeds and the words to read if you like, in one run of Drama 3. It measures each candidate (pitch, brightness, background, creak) and keeps the one you choose as a made-up voice of your own (HERESY 2455). 2.0.7 came with no ready-made voices; 2.0.8 comes with twelve, one for each of the Narrator's languages, six men and six women, made by Drama 3 itself from words alone, no person's voice cloned, and each marked never for training. They are Drama 3's own set, so its six timbre groups are formed by them on every install (HERESY 2403, 2449, 2456). The Narrator still reads in the engine's own voice by default. Built with Fish Audio.
The Musician's mirror. The room now stands in the Musician's cards (the lectern, the voice, the engine) with Read in a bar at its foot, everything in sight at 1920 px, and a speaking voice's envelope breathing under the room's name while a run reads (HERESY 2401). A reading's take is read again in the Narrator with every setting and a new seed; before, Regenerate from the Library would have sent a reading to YuE2 to be sung (HERESY 2510). A run card plays, regenerates, fixes and checks with Whisper, and pours its reading back into the form four ways (everything but the seed, the voice and settings, exactly with its seed, only its text) with an undo (HERESY 2511–2515). A book read in parts is one card in the Library, played through or from any part (HERESY 2402).
The ear: on, off, or afterwards. Hear each piece (Whisper) stays on by default. Turned off, the take says it was not heard, never passing for a clean one. Check with Whisper hears a finished take afterwards, from the Library or its run card, piece by piece against the words it was asked to read, and writes the words lost, added or swapped, with their times, into its note; the audio is left untouched (HERESY 2434).
Drama 3 on an Intel Arc, and Q8_0 by default. Drama 3 now reads on an Intel Arc through audio.cpp's Vulkan build, which tools/narrator-setup.sh builds on a machine with an Intel card, with a tier of its own: Q4_K, 28 of its 40 expert banks in RAM, RTF 3.3 on an Arc Pro B50 (HERESY 2426). There it decodes in audio.cpp's streaming session, four frames at a time: the same sound (the two decodes 51 dB apart) and 9.1 GB at the card's peak instead of 15.1 on a long piece (HERESY 2516). On NVIDIA cards Q8_0 is the default wherever it fits, then Q6_K, Q4_K last: Viktor's ear gave Q8_0 20 of 20, Q6_K 19 of 20 and Q4_K 43 of 50. A card of 24 GB fetches Q8_0 at setup (HERESY 2435, 2436, 2441).
Smaller things. A paragraph's mood carries through all its pieces (HERESY 2439). Japanese and Chinese are read in whole sentences, cut at 。！？ (HERESY 2427). «!».» and «?».» no longer leave Drama 3 silent, and a piece the engine answers with an error is asked again with another seed, three times in all (HERESY 2422). The speech chain's denoiser no longer leaves a take 25 ms late (HERESY 2418). Notes take 6000 characters, a longer one cut with «…», never silently (HERESY 2404, 2438). Two trials for the emotional arc of a long reading are in, off unless a run asks (HERESY 2474).
An editor works with two viewers, and so does the Videographer now: the animatic or the clip on the left, the edit's program always on the right, each as large as the room leaves (1223 × 735 a viewer at 2560 × 1440, 903 × 430 at 1920 × 1080), and one viewer with a switch under 1600 px. The room has one sound: whatever starts playing pauses every other player of the page, the Studio's own included (HERESY 2392).
The edit no longer packs everything into one track. Each part has its own: V1 the shots, ST1 the titles in their template (the text-track name an editor knows from Resolve), K1 the karaoke from the take's Whisper timing, O1 the look, A1 the song. An item chosen opens in the inspector, which now scrolls beside the timeline, so the program monitor keeps its size (HERESY 2393, 2517, 2431).
The titles and the karaoke take a font of your own, installed on your computer or a .ttf or .otf file, and the render really uses it. On the way we found that libass matches a font by its legacy family name, not the one menus show, so a Montserrat Thin asked for as «Montserrat» fell back to DejaVu Sans. The clip now names the font the way libass reads it, and a font that is gone stops the render with the reason instead of a silent fallback (HERESY 2394).
A finished clip heavier than 6 Mbit/s gets a light copy (720p, under 3 Mbit/s) for a slow link: two stalls within 20 seconds switch to it by themselves, and the download stays the master (HERESY 2395). A take chosen again reads its lyrics' timing anew, so the karaoke fills once Whisper has timed it (HERESY 2432).
We have seen a public model taken down by its makers two days after its release. In Viktor's words: «Форс мажор — это не конспирация, это предохранитель» (force majeure is not a conspiracy, it is a fuse). So every file of every pinned model the Studio takes, 45 pins of ours and of others, is now found by its content (a large file by its size and SHA-256, a small one by its git blob id), never by its name, and kept in a cold copy of our own that counts only after it has been read back from the disk and hashed (HERESY 2423).
Before each release a guard asks Hugging Face live, never a cache: does each pinned repository still answer, not deleted, made private or gated since; was it renamed; does its commit still exist; is each file the Studio takes still there with its size and its sum. Each weight of 8 MB or more is asked for at its pin, so a storage purged under an unchanged listing is caught too. A loss stops the cut, and so does a Hugging Face that does not answer: no release is cut blind. Its run on 12 November: 44 pins, none lost, two warnings in 48 seconds, both a third party's quants re-made at its head while our pin keeps the ones we took (HERESY 2423).
The sound half sings a song through one window of the model's context. How long that window is depends on how much of it the style, the lyrics and the score take: on the whole context it is about 6:16 to 7:52, and 7:28 in the middle of the 407 songs we measured. In 2.0.7 a song past one window was sung in pieces joined in a crossfade, and the pieces were the default (HERESY 2509).
In 2.0.8 a song is sung whole by default, and the pieces are an experimental choice: Sampling › Sound and output › Long songs, Whole · up to one window or Pieces · experimental. Viktor's word was to make the pieces «опцией для реально длинных песен, и то, экспериментальной, а не дефолтной» (an option for really long songs, and an experimental one at that, not the default) (HERESY 2509).
What this means for a song over about 6½ minutes. With the default, a song the model would sing past its window now ends at the window, whole and with no seam; a toast says so, and the take's card says where it was held («held whole at 6:41 (one window)»). If you want the full length, choose Pieces · experimental: the song goes on past the window in pieces, as in 2.0.7, with a dip of about 1 dB at each joint for a few seconds. The output hint warns when Max length is more than 30 s past one window on your engine's context. A kept part (Regenerate from a moment), or a Continue or Extend from a source already past one window, goes on in pieces as that take was made (HERESY 2509).
On a card with less memory the window is shorter, because Engine › Auto gives the engine a smaller context: about 3:32–5:08 with 11.4–15.4 GB, 2:10–3:47 with 7.4–11.4 GB, and 0:48–2:25 below that. On such a card many songs were already sung in pieces under 2.0.7; with the default they are now held at the window, so there Pieces · experimental is the way to the full length (HERESY 2509).
The Intel Arc runs the engine through SYCL (./build.sh --sycl). We built it with ggml's Vulkan too and measured both on the same card. On Vulkan every op of our graphs runs on the card, flash attention included (over a 30-second song, 77,537 nodes in 69 splits, none on the processor), and the sound is the same: fed the same tokens and score, a Vulkan take matches a SYCL take as closely as SYCL matches itself (correlation 0.9988 against 0.9987, the four bands within 0.1 dB). But it sings 1.6 to 3.4 times slower, nearly all of it flash attention in the sound half (HERESY 2450).
So the Arc's engine stays SYCL. The engine keeps two switches from the measure: --no-fa-nar turns flash attention off in the sound half (quicker on Vulkan for short songs, but the sound half's memory then grows with the square of the length: 11.6 GB at 100 s of audio), and YUE2_GGML_DEBUG=1 lets ggml's own debug log through (HERESY 2450).
On 12 November Viktor gave the Studio a law: «Максимум компактности, минимум текстового кнопочного клаттера. Это закон UIX нашей Студии» (maximum compactness, minimum clutter of text buttons: this is the UIX law of our Studio).
We walked through every room against it before the cut. The audit measured the live page at 1920 × 1080 and 2560 × 1440 in the night theme, and a sample in the day theme, in a headless browser that answered every request that would change something inside the browser itself, so nothing reached the lab. It wrote down 103 findings: 29 that a user meets at once (A), 52 worth mending (B) and 22 of polish (C). All 29 A findings were mended before the cut, and the B ones where the mend was local and safe (HERESY 2521–2595).
What came of it:
- One set of icons for one act everywhere: Lucide's shapes (ISC) under Lucide's own names, about a hundred of them added, ☰ drawn instead of a glyph (HERESY 2526).
- One height for a field (36 px) and one for a bar's control (30 px) in every room (HERESY 2530).
- The Studio's own picker for every room's long lists, with a Find where a list is long (HERESY 2527, 2544).
- A button's words kept in its label and its tip, so every language still reads them; text stays on a room's main act and its ways out (HERESY 2541, 2567, 2580).
- Nothing lies over a room's fields any more: the folded log is a button in the player's row and a training lives in the system bar (HERESY 2522). The Refiner's room is no longer 43 px taller than the window (HERESY 2533), and the system bar fits at 1920 px with every disk in sight (HERESY 2523).
- English left inside other languages is translated (Models and LoRAs, the Narrator's quant, the Librarian's counts), with thirty new words in the twelve catalogs (HERESY 2532, 2548, 2550, 2562, 2574).
The visiting card said the install takes 34.4–41.1 GB, and README said 34.9–41.6 GB. Both were wrong since 2.0.7, whose install brings only what a song needs. We recomputed them from the pinned file lists of every repository and from a fresh install's disk, and wrote one set everywhere: the card, README, INSTALL, the launcher and the Narrator's guide (HERESY 2485).
- The install: 25.9 GB on a 24 GB card and 22.5 GB on a card of 16 GB or less (the program 0.9 GB, its two Pythons 16.5 GB, our fine-tune at the card's quant and two decoders).
- The Musician's other parts (our second decoder, the transcriber, Whisper, the stems, UniverSR and Apollo-vocal, the LoRAs): 8.8 GB.
- The Narrator: 42.1 GB at Q8_0 or 26.0 GB at Q4_K, with the stress dictionaries of three languages.
- Every room: 239.7 GB on a 24 GB card, 126.5 GB on 12–16 GB and 121.8 GB on 8 GB.
README's «What It Downloads» says each part, measured file by file.
The Linux AppImage, RuachStudio-2.0.8-core-x86_64.AppImage (1.3 GB), still opens a window of its own first and carries its own C library. Its first install now says how many GB it downloads before anything comes: 8.5 GB on a 24 GB card, 5.1 GB on 7 to 16 GB, 4.3 GB below that and 4.0 GB on the processor. When the network goes, the download stops honestly: the window says it is waiting for the network, what landed stays, and the download goes on by itself once the network is back. We measured it in a Debian 10 container with the network cut under a real download: offline was said in 16–20 seconds with nothing more landing, and the fetch went on 3–5 seconds after the network came back (HERESY 2411).
The image is also cleaner: a packaging patch is applied only whole, so no .orig or .rej leftovers ride along (HERESY 2416), and it carries no compiled Python bytecode from the machine that packed it (HERESY 2601).
A score without its key line is sung as written, with no key signature, and the engine says so in its log, the take and the page, where it used to stop the run. The page offers the key, or two, that its notes and chords read as, with how sure, and puts the line into the score's header at a click (HERESY 2433, 2444).
One rule for every name. Wherever the Studio names something (a workspace, a section, an album, a take's title, a preset, a voice…), a name never holds / \ * ~ ? : " < > | or an invisible control character and never begins with a dot. The dialog says what it refuses and offers a cleaned name, never changing one quietly, and the lab holds the same rule (HERESY 2462).
The Trainer's eight charts for every run, in one order: the loss, the sound half's flow loss, the music half's cross-entropy and its KL to the base, the learning rate, the gradient norm, seconds a step and each card's peak memory. A metric a run never logged is an empty chart that says so, never a line at zero (HERESY 2442).
Names stay names. LUT, EQ, GPU, VRAM, LoRA, Whisper and the others stay themselves in every language, while ordinary words are translated (HERESY 2430, 2443). In Russian the Studio keeps the root «раб» out of what it says, as its Ukrainian and Belarusian already did: a machine «работает», while the language model and the rooms create (HERESY 2488).
Around the rooms. The Engine opens on Ctrl+Alt+9 (Ctrl+9 in the installed app), so Ctrl+0 always gives the window's zoom back (HERESY 2405). In the playlist every take starts from its beginning (HERESY 2406). A dry bounce keeps its lanes' room and delay as live sends (HERESY 2407). Eight new demo songs, one each in Ukrainian, Belarusian, Greek, Spanish, Italian, French, Portuguese and German (HERESY 2412). An engine started with our decoders alone refuses to encode with the reason, instead of crashing on a recording (HERESY 2484).
- Installed with git or Pinokio:
git pullor Pinokio's Update brings the Studio to 2.0.8, and Engine → Updates offers v2.0.8. - Linux AppImage:
RuachStudio-2.0.8-core-x86_64.AppImagewith its.sha256and.zsyncon the release page. The new AppImage started in place of the old one is the update: your models, songs and settings stay where they are, and its first start asks only for what changed among the models you have. - Higgs TTS 3 is in neither: Engine → Models fetches it from Boson AI's own repositories once you accept Boson's licences, or
tools/narrator-higgs-setup.shdoes it in a terminal. - Our models, our decoders and the Narrator's stress indexes: huggingface.co/goldhub.
- Long songs. With the default, a song past one window ends at the window: about 6:16 to 7:52 on the whole context, shorter on a card with less memory. Pieces · experimental sings it to its end, with a short dip at each joint.
- Higgs TTS 3 runs on an NVIDIA card or the processor, not on the Intel Arc, and in BF16 only. It comes under the Boson Higgs TTS 3 Research and Non-Commercial License, its tokenizer under the Boson Higgs Audio 2 Community License: free for research and non-commercial use, Boson's written licence needed to host it for others or put it in a product, and no model may be trained on what it makes, so the Studio marks every Higgs take never for training. docs/NARRATOR-HIGGS.md says it whole.
- The Russian stress index is built from gramdict's data under CC BY-NC 4.0, for non-commercial use.
- The Intel Arc. Drama 3's streaming decode is used on the Arc only; NVIDIA cards and the processor read as before until we have measured them. Songs on the Arc stay on SYCL, and the Arc remains a beta: training, the Artist and the Refiner's GPU steps stay on NVIDIA.
- Stresses. A word the dictionaries stress two ways is left bare unless you mark it or your own dictionary has it.
- The emotional arc of a long reading: its two trials are off unless a run asks for them.
- Hebrew. The guide reads in English, left to right, until the Hebrew one comes (HERESY 2413).
-
The AppImage does not carry the full Python with PyTorch that the stems need (
lab/install-venv.sh).
To m-a-p for YuE2; to ServeurpersoCom for yue2.cpp and ggml-org for ggml; to Fish Audio for Drama 3 (Built with Fish Audio) and to ShugoAI for audio.cpp; to Boson AI for Higgs TTS 3 and its audio tokenizer, and to multimodalart for the transformers port its loader comes from; to gramdict for A. A. Zaliznyak's Grammatical Dictionary as data, and to the Belarusian NLP community for GrammarDB; to lang-uk and ULIF for the Ukrainian stresses; to Lucide for the shapes of our icons (ISC); and to everyone who opens an issue or a pull request.
The new third-party parts, credited as README's «Standing On» credits them:
- Higgs TTS 3 by Boson AI: Boson Higgs TTS 3 is licensed under the Boson Higgs TTS 3 Research and Non-Commercial License, Copyright (c) Boson AI USA, Inc. All Rights Reserved. Built with Higgs TTS 3 licensed from Boson AI USA, Inc. Built with Higgs Materials licensed from Boson AI USA, Inc., Copyright Boson AI USA, Inc., All Rights Reserved and Meta Llama 3 licensed under the Meta Llama 3 Community License, Copyright Meta Platforms, Inc., All Rights Reserved.
- «Грамматический словарь русского языка» by Andrey A. Zaliznyak (6th edition, 2010; the rights: Anna A. Zaliznyak), published as data by gramdict (CC BY-NC 4.0; the proofs prepared by Elena A. Grishina): the Narrator's Russian stresses, every form unfolded from its indices.
- GrammarDB, the Grammar Database of the Belarusian Language of the Belarusian NLP community (CC BY-SA 4.0), built from the major dictionaries of Belarusian: the Narrator's Belarusian stresses.
Everything the Studio stands on is in README's «Standing On» and in the Engine's About.
Everything the YuE2 community has made is gathered in Awesome YUE2, our curated list, and why we build a Studio of our own is told in Why our own Studio, not a ComfyUI plugin bundle.
Every change, entry by entry: CHANGELOG and heresy/docs/HERESY.md. Questions and findings about this release are welcome in its announcement, in Discussions and in Issues.
Release articles
Articles
Mini-guides
Links