Fish Audio S2 Pro

# Fix list: Fish Audio S2 Pro

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/fish-audio-s2-pro, the October 2026 research run, assessed 23 October 2026. Grade D, 52.5 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 S2 Pro: 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, 30 out of 100, up to 14 more on the total

Why it scored 30: File safety gets 5. The weights are safetensors but the codec is a pickle beside them (15), and prompts are built by Fish Audio's code rather than a template (10 off). Provenance gets the full 15 for files in Fish Audio's own organisation with hashes for every file. Licence risk keeps 5 after losing 5 for the incorporated use policy and the ban on training other models, and 5 for a revocable licence whose use policy Fish Audio controls. Agent guidance gets 5 for a general legal disclaimer. Safety evaluation and a security programme score 0, with no safety results and no route for reporting a flaw.

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, 61 out of 100, up to 13.7 more on the total

Why it scored 61: From 9 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, Intelligibility and speaker similarity 82, Speed 12. 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, 50 out of 100, up to 12.5 more on the total

Why it scored 50: Memory earns the full 30, since the BF16 files and the codec come to 11.01 GB, about 13.2 GB with the fifth added, though Fish Audio's own guide asks for a 24 GB GPU. Runtimes score 0, as no released version of the six named runtimes has its own code for this model; it runs through Fish Audio's fish-speech code, or SGLang-Omni and vLLM-Omni, which are separate projects. Builds score 0 because Fish Audio publishes no quantised S2 Pro. The research and non-commercial licence gives terms 5. The ungated repository gives download the full 10, and hosted too earns 5 because Fish Audio's API serves s2-pro and says it's the open-sourced model.

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, 56 out of 100, up to 4.4 more on the total

Why it scored 56: The card covers intended use, architecture, parameters and languages, and misses limitations, context length and a cut-off (11). Training data is described as over 10 million hours in about 80 languages with the filtering pipeline, but no sources or cut-off (5). The files haven't changed since March, but Fish Audio merged fixes to the code that runs the model on 10 October (20). A fishaudio member replied to 5 of 27 discussions, none since July (5). Earlier models keep their own repositories and the move to Drama 3 is dated (10). File health gets 5 for a template that isn't the prompt format, context lengths that disagree, an unanswered report about the vocoder's output, and a card that still promises fine-tuning code removed from the main branch on 13 October.

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. Fish Audio documents the server's endpoints, every request argument and the output formats in its own docs (25). One request can carry several speakers, each cloned from its own clip, with inline style tags (20). Output format and streaming are documented with how to ask for them (15). The chat template scores 10 because the file shipped is Qwen3's text template, and the real prompt is built in code. Reasoning control scores 0 with nothing to control, and context 0 because the report trained it at 16,384 tokens.

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.

- Unchecked, GitHub issues on fishaudio/fish-speech, since the GitHub API isn't enabled for this session; responsiveness rests on the 27 Hugging Face discussions
- Memory follows the checklist's rule (11.01 GB × 1.2 = 13.2 GB, 16 GB band, 30 points). Fish Audio's install guide says at least 24 GB of GPU memory, which would be the 25-point band if the publisher's figure were used
- The Acceptable Use Policy that the licence incorporates isn't linked or named by URL, and fish.audio pages beyond the terms and privacy policy weren't read because fish.audio's terms forbid crawling. If a reader counts an unlinked policy as a broken link, licence risk loses a further 5 and reaches 0
- The licensor is 39 AI, Inc., while fish.audio's terms name Hanabi AI Inc.; how the two companies relate wasn't established
- The Qwen3-4B base is Apache 2.0, which asks redistributors to keep its licence and notices; the S2 Pro repository ships no Qwen licence or NOTICE file
- fish-speech's Drama 3 update (13 October 2026) removed train.py, the LoRA code and the fine-tuning docs that the S2 Pro card says the release includes. Scored in file health; whether it's also a negative event (a feature removed while still advertised) is the founder's call
- sameModel is set to fish-audio-tts on Fish Audio's own words (changelog and models overview). Whether the API serves the same BF16 weights, rather than a converted copy on its serving stack, isn't stated

## Weaknesses

- Research and non-commercial use only; commercial use, internal business use and hosted services need a separate licence
- Not in any released version of Transformers, vLLM, SGLang, llama.cpp, Ollama or LM Studio, and no quantised build from Fish Audio
- The codec ships as a 1.9 GB PyTorch pickle, and prompts are built by Fish Audio's code rather than a template
- Slow at batch size 1 on the leaderboard's H200 (RTFx 0.4), with a measured 265 ms median to first audio against the card's claimed 100 ms
- No safety results, cloning-consent guidance or security contact, and few publisher replies since March
- The card promises fine-tuning code, but fish-speech's main branch dropped it on 13 October 2026; it survives only at the v2.0.0-beta tag

## 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 fish-speech from GitHub (`uv sync --python 3.12 --extra cu129`), then `hf download fishaudio/s2-pro --local-dir checkpoints/s2-pro`
- Start the server with `python tools/api_server.py --listen 127.0.0.1:8080 --llama-checkpoint-path checkpoints/s2-pro --decoder-checkpoint-path checkpoints/s2-pro/codec.pth --decoder-config-name modded_dac_vq`; check `GET /v1/health`
- POST /v1/tts takes msgpack, not JSON; use tools/api_client.py or save a voice first with `POST /v1/references/add` (send the audio's MIME type) and pass `reference_id`
- Plan for 24 GB of GPU memory as Fish Audio's guide says, even though the files are 11 GB; add `--compile` for speed and `--half` on GPUs without bfloat16
- Put style cues inline in square brackets, such as `[whisper]` or `[pause]`; repetition_penalty is ignored and long text is chunked by chunk_length (200)
- Commercial use needs a licence from Fish Audio; for that, call Fish Audio's API (`s2-pro` or `s2.1-pro`) instead

## What the review panel asked for

- codec in safetensors
- a security contact
- update the card for the v2.0.0-beta tag
- answer open discussions
- a JSON request body on /v1/tts
- a published commercial licence price

## 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.