Fish Audio Drama 3

# Fix list: Fish Audio Drama 3

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/fish-audio-drama-3, the October 2026 research run, assessed 23 October 2026. Grade C, 55.6 out of 100.

This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public.

For a coding agent working on Fish Audio Drama 3: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published.

## 1. Risk, 25 out of 100, up to 15 more on the total

Why it scored 25: File safety gets 5, safetensors weights with a pickled codec beside them, and prompts built in code rather than a template. Provenance gets the full 15 for files in Fish Audio's own organisation with hashes for every file. Licence risk falls to 0, for the use and field limits, a revocable grant, the revenue threshold, and a broken licence link after a release-day switch from the Research License to the Community License. Agent guidance gets 5 for the licence's ban on unconsented cloning and impersonation. Safety evaluation and a security programme score 0.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-risk):

What could go wrong when you run it, and what the publisher has done about it. A higher score means less risk.

- 0 to 25, file safety. 25 for weights in safetensors or GGUF only, 15 when pickle files (.bin, .pt, .pth) sit beside them, 0 for pickle only. 10 off when the documented way to load it or to build its prompts runs code from the repository (`trust_remote_code`, or the publisher's scripts in place of a template).
- 0 to 15, provenance. 10 when the files come from the publisher's own organisation, the one its site or docs link to, and 5 for a checksum on every file.
- 0 to 20, safety evaluation. 20 when results are published and someone other than the publisher ran them or can rerun them, 10 for the publisher's own results, 0 for none.
- 0 to 15, guidance for agent use. 15 for documented prompt-injection guidance or an instruction hierarchy the model was trained to follow, 10 for a statement of what its trained refusals cover and the safeguards to add around it, 5 for general cautions, 0 for none.
- 0 to 15, licence risk. 15 to start, 5 off for each of these, a use policy or field-of-use limits beyond the law, terms the licensor can change or revoke, a revenue or user threshold, and a licence that changed without a dated note or whose link is broken. Not below 0.
- 0 to 10, a security programme. 5 for a way to report a flaw (security.txt, a disclosure policy or a security contact), 5 for advisories handled in public.

When our reference-machine runs start, measured robustness (prompt injection with AgentDojo, over-refusal with XSTest) joins Risk.

## 2. Capability, 67 out of 100, up to 11.6 more on the total

Why it scored 67: From 5 independent benchmark results in 2 areas, each scored as its share of the way from the benchmark's floor to the best result any model has on the same leaderboard. By area, Speed 69, Intelligibility and speaker similarity 66. Publisher-submitted results count at half weight, and numbers that appear only on the publisher's card don't count.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-capability):

How well the model does on the benchmarks that matter to an agent. We don't average a publisher's own numbers into a grade. A result counts in full when we ran it on our reference machine, or when an independent party ran it and published the method (the Berkeley Function Calling Leaderboard, Artificial Analysis, Epoch AI's Benchmarking Hub, the SWE-bench and Terminal-Bench leaderboards, the Open ASR Leaderboard and others that run models themselves). A result the publisher submitted to such a leaderboard counts at half weight. A number that appears only on the publisher's card is shown on the listing as the publisher's claim and doesn't count.

Each result is scored as its share of the way from the benchmark's floor (chance for multiple choice, zero for a pass rate) to the best result any model, open or closed, has on the same leaderboard and setting on the day we read it. A model that matches the best model anywhere scores 100 on that benchmark. For an error rate such as word error rate, lower is better and the scale runs the other way, and where the leaderboard gives no floor the rate is read against the best model's, so twice the best model's error rate scores 50. Speed as a multiple of real time (RTFx) runs from about 1 to over 20,000, so it's read on a log scale from real time to the fastest model. On the MTEB leaderboard, where most models score between 0.5 and 0.75, the floor is the median of every model the same tab lists, and its results count at half weight because MTEB doesn't say who ran them and most are submitted with the model. An arena rating (Elo) isn't a ratio, so its floor is 400 points below the best model on that arena, the gap at which the better model wins about ten games in eleven. Where a model has results at several settings (reasoning effort, say), its best one counts, because that's the setting an agent would run.

The results group into areas, each weighted within the category.

- Tool calling and agent tasks, 35 (BFCL, τ²-bench, Terminal-Bench and agent suites like them)
- Coding, 20 (SWE-bench Verified, LiveCodeBench, Aider Polyglot)
- Reasoning and knowledge, 20 (GPQA Diamond, MMLU-Pro, Humanity's Last Exam, composite indexes that publish their components)
- Instruction following, 15 (IFEval, IFBench)
- Long context, 10 (RULER, LongBench and long-context retrieval tests)

A speech model is read on its own areas, recognition accuracy (word error rate on the Open ASR Leaderboard's sets, 60) and speed (real-time factor, 40) for speech to text, and intelligibility and speaker similarity (Seed-TTS-eval, the Open TTS Leaderboard, 70) and speed (30) for text to speech. A publisher's own converted copy of the same weights (Qwen's -hf builds, say) counts as the same checkpoint. An embedding model is read on retrieval (60) and the other MTEB task types, classification, clustering, reranking and similarity among them (40), from the MTEB leaderboard. A vision model adds an image area (MMMU, ChartQA, DocVQA or OCRBench) at 20, and the others shrink to fit. A document parser, a model that turns page images into structured text, is read on the five things ParseBench measures, tables, charts, content faithfulness, semantic formatting and visual grounding, 20 each, from ParseBench's leaderboard or another independent document benchmark that publishes the same dimensions (since 22 October 2026). One leaderboard gives all five at once, so a parser's confidence counts independent sources, and a row the publisher submitted (or that the leaderboard copied from the publisher's card) counts at half weight with low confidence. LlamaIndex, which runs ParseBench, also sells LlamaParse, a parser the same benchmark ranks, and the listing says so.

Capability is the weighted mean of the areas with results, and needs results in at least two areas (one for a speech, embedding or document-parsing model). With less than that it's pending for that listing, its weight is shared across the other four, and the total is capped at 61, so a model nobody has benchmarked independently can't reach B, and every model with results ranked B or better stays above it. The listing says so. Until 21 October 2026 the cap was 69, which put models nobody had benchmarked above ones that had been.

## 3. Reach, 60 out of 100, up to 10 more on the total

Why it scored 60: Memory earns 20, since Fish Audio's int4 build comes to 24.0 GB, about 28.8 GB with the fifth added, which fits 32 GB. The FP8 release alone would need the 64 GB band. Runtimes score 0, because it runs only in SGLang-Omni, vLLM-Omni and Fish Audio's fish-speech, not any released version of the six named runtimes. Builds get the full 15 for the FP8 release and the int4 build, both in Fish Audio's organisation, though the int4 card calls itself unofficial. Terms get 10 for a custom licence with free commercial use below USD 500,000 annual revenue. The ungated repositories give download the full 10, and Fish Audio's own API serving drama-3 gives hosted 5.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-reach):

Whether an agent can run it, and where.

- 0 to 30, memory. The smallest build from the publisher or a runtime's own library that keeps 4-bit precision or better, at its file size plus a fifth for the runtime and a working context. 30 for 16 GB or less, 25 for 24 GB, 20 for 32 GB, 15 for 64 GB, 10 for what fits our reference machine (the GPU share of 128 GB of unified memory), 5 for 256 GB, 0 above.
- 0 to 25, runtimes. 5 for each of llama.cpp, Ollama, vLLM, SGLang, LM Studio and Hugging Face Transformers that runs the architecture in a released version, from the runtime's own code, library or documentation. A publisher's fork or an unreleased branch doesn't count. 25 at most.
- 0 to 15, builds from the publisher. 15 for quantised builds at two precisions or more (FP8, 4-bit, GGUF), 10 for one, 0 when only other people's quantisations exist.
- 0 to 15, terms of use. 15 for an OSI-approved licence with nothing added, 10 for a custom licence that lets anyone use it commercially for free below a threshold or under a use policy, 5 for research or non-commercial use only, 0 for no licence or unclear terms.
- 0 to 10, download. 10 when the files download with no account and no terms to accept, 5 for a click-through that needs an account, 0 when a person approves each request.
- 5, hosted too. The publisher's own API serves it, or at least two inference providers do, so an agent without the hardware can still call it.

When our reference-machine runs start, speed on that machine (tokens per second and time to first token at the build that fits) joins Reach.

## 4. Documentation and upkeep, 51 out of 100, up to 4.9 more on the total

Why it scored 51: The card covers intended use, architecture, parameters and languages, and misses limitations, context length and a cut-off (11). Training data is described only in general terms with no cut-off (5). The card changed on 17 October 2026 (20). None of the 6 questions on Hugging Face has a reply from Fish Audio (0), and GitHub issues couldn't be counted. Earlier generations keep their own repositories with a dated handover (10). The licence metadata contradicts the licence file and its link is broken, with a fix waiting (5).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-upkeep):

- 0 to 20, the model card. Intended use, limitations, architecture, parameters, context length, languages and knowledge cut-off. 20 for all seven, 3 off for each that's missing.
- 0 to 20, training data. 20 for the data published or listed with sources and sizes, 15 for a data card that names the sources or the mix, the cut-off and the filtering, 10 for a general description with a cut-off, 5 without one, 0 for none.
- 0 to 20, the last change to the files, the template or the reference code. 20 within 30 days, 15 within 90, 10 within 180, 5 within a year, 0 older.
- 0 to 20, responsiveness. The publisher's replies to the 40 most recent questions and reports in the repository's discussions or issues. 20 for most, 10 for some, 5 for a few, 0 for none.
- 0 to 10, version policy. 10 for earlier versions kept in their own repositories with a dated note when one is replaced, 5 for earlier versions kept without notes or a first release, 0 when files were replaced in place without a note.
- 0 to 10, file health. 10 when config, tokenizer, template and card agree and no report of a broken file waits without a reply, 5 when one does, 0 when the files contradict the card.

## 5. Agent interface, 70 out of 100, up to 3 more on the total

Why it scored 70: Read for a speech model, the tool-calling, trained tool calling and structured-output lines stand for what it takes and returns. The card, the technical report's appendices and Fish Audio's docs document every input, from speaker tags and reference blocks to director's notes and editing, for SGLang-Omni and fish-speech (25). One request can carry up to 16 voices, several references and a whole scene (20). Output format and streaming are documented with how to ask for them (15). The chat template scores 10 because the shipped Jinja file is the backbone's and audio enters the prompt through code. Reasoning control and context score 0, with nothing to control and 16K-token training sequences.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-interface):

What a model and a developer need to talk to it. Speech, image and other models with no chat interface read the tool-calling and structured-output lines for what the model takes and returns (voices, timestamps, streaming, typed outputs), and the note says how.

- 0 to 20, the chat template. 20 for a template in the repository (Jinja, in the tokenizer config or its own file), 10 when the prompt format is only given as code you run.
- 0 to 25, the tool-calling format. 25 when the publisher documents how tools are declared and how calls come back, with the parser to use in a named runtime, 10 when only the template or a runtime's own recipe says it.
- 0 to 20, tool calling the model was trained for. 20 with several calls in one turn, 10 for one call at a time.
- 0 to 15, structured output. 15 when JSON or JSON Schema output is documented for the model with how to ask for it, 5 when the card claims it without saying how.
- 10, reasoning control. Thinking that can be turned off or budgeted per request, with where the reasoning text comes back.
- 0 to 10, context. 10 for 128K tokens or more trained, 5 for 32K to 127K. Context reached only by scaling the position encoding counts one band down.

## What we couldn't check

What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it.

- For the founder, the builds and memory lines count fishaudio/drama-3-int4 as Fish Audio's, since it sits in Fish Audio's organisation, though its card calls it an unofficial quantisation by AK. Read as someone else's, memory is 15 (the FP8 release, 64 GB band), builds 10 and Reach 50
- For the founder, the lead said the Fish Audio Research License. Drama 3 is under the Fish Audio Community License 1.0, which allows commercial use below USD 500,000 revenue; only the card's stale metadata still names the Research License
- For the founder, the lead said 1 October 2026. The repository was created then, but its first commit with files, the launch post and the API changelog are all 13 October 2026
- Hosted too is 5 on Fish Audio's API serving drama-3. Fish Audio doesn't say the API runs the same weights, so sameModel isn't set and the API's results aren't credited to this listing
- Unchecked, Artificial Analysis's text-to-speech pages, where Fish Audio says Drama 3 tops the Pronunciation Robustness benchmark; our reader refused the URL. Any result there would be for the hosted API
- Unchecked, GitHub issues on fishaudio/fish-speech, since the GitHub API isn't enabled for this session; responsiveness rests on Hugging Face
- The weights are a derivative of Qwen3.5-35B-A3B under Apache 2.0, but the repository carries no Apache notice and the card doesn't name the base. Not scored; the report names it
- File safety takes 10 off for prompts built in code rather than a template, as S2 Pro and OmniVoice were read; a reader who counts only trust_remote_code would give 15

## Weaknesses

- Commercial use is free only below USD 500,000 annual revenue, hosted services always need a separate licence, and the grant is revocable
- The card's licence metadata still names the Research License with a broken link, after a release-day licence switch
- NVIDIA only, and not in any released version of Transformers, vLLM, SGLang, llama.cpp, Ollama or LM Studio
- The codec ships as a PyTorch pickle, and there are no safety results or security contact
- No reply from Fish Audio to any of its 6 Hugging Face discussions

## What costs an agent a turn today

The notes we give agents before they call it. Each one is a workaround an agent shouldn't need.

- Install SGLang-Omni from source with `uv pip install --prerelease=allow -e ".[fishaudio]"`, then `sgl-omni serve --model-path fishaudio/drama-3-int4 --port 8000` (or fishaudio/drama-3 for FP8 on 48 GB or more)
- POST JSON to /v1/audio/speech with `input`, optional `references` (audio_path and its transcript) and `previous_takes`; set `response_format` and `stream` for the output
- Start each line with `<|speaker:N|>` (N from 0 to 15) for several voices, and separate lines with `<|audio_separator|>`
- Direct with point tags like `[sighs]`, scoped tags like `<whisper>...</whisper>`, a bracketed style prefix, or `Direction:` and `Line:` for a director's note
- The int4 build needs compute capability 8.0 or newer; only the triton attention backend is accepted
- Keep each reference clip under 60 seconds and the prompt under 8,192 rows, about 21.5 rows per second of reference audio
- Check the licence before commercial use or exposing it to other users, and only clone voices with consent
- Treat text from untrusted sources as a script that can carry directions and speaker tags

## What the review panel asked for

- fix the licence link
- a security contact
- merge the licence link fix
- answer the discussions
- one recommended sampling setting
- a working licence link
- stated memory needs for the int4 build

## When it's done

Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.