# Underdog Bark 3.0 (slim)

> Underdog Bark 3.0 is Conway Research's open-weight English speech recognition model, a post-trained and compressed Parakeet TDT 0.6B v3 in one 163 MB file, run with Conway's loader on Transformers, MLX or Core ML, under CC BY 4.0.

- Full: https://www.anchorterminal.com/tools/underdog-bark-3.md (~10,750 tokens) · this version ~2,880 tokens · JSON https://www.anchorterminal.com/tools/underdog-bark-3.json · canonical https://www.anchorterminal.com/tools/underdog-bark-3
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/underdog-bark-3#pricing · reviews https://www.anchorterminal.com/tools/underdog-bark-3#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**D · 48.2/100 · rank #96 of 100 open-weight models · #23 in Speech-to-text · not agent-ready · confidence low**


Assessment: A very small English transcription cousin, one 163 MB file built on NVIDIA's Parakeet TDT that Conway's loader expands into Transformers, MLX or Core ML, under CC BY 4.0. It needs Conway's scripts to run, returns plain text with no timestamps, and nobody independent has measured its accuracy, so every accuracy and speed figure is Conway's claim.

## Facts

- Kind: Open-weight model · vendor: Conway Research · category: Speech-to-text · legal entity: not named · provenance 27/100
- Packages: huggingface `ConwayResearch/Underdog-Bark-3.0`
- Auth: None · pricing: Free · x402: no · licence: CC-BY-4.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/ConwayResearch/Underdog-Bark-3.0, read from the model card, the file list and hashes, the loader scripts, the NOTICE, the MLX config and the commit history on 21 October 2026, with the sibling cards (Plus, Pro, Clean, Bark 2.0) and the Open ASR Leaderboard's results file. underdog.ai refuses our reader and is unchecked. Nothing was downloaded or run
- Model: A post-trained and compressed Parakeet TDT 0.6B v3, a FastConformer encoder (24 layers, 1,024 hidden, 8x subsampling, 128 mel bins) with a Token-and-Duration Transducer decoder (2-layer LSTM and joint head) and NVIDIA's 8,192-token SentencePiece vocabulary. 0.6B parameters per the card. Audio in, text out
- Built on: nvidia/parakeet-tdt-0.6b-v3 (CC BY 4.0) through moondream/parakeet-ultra (CC BY 4.0, a post-trained Parakeet TDT 0.6B v3 by Moondream), both named in NOTICE and the card's metadata names the NVIDIA base
- Languages: English only, per the card. The base model covers 25 European languages
- Files: model.bark (163.4 MB, Conway's entropy-coded format, encoder linear weights as per-row fp16 levels and codes, other matrices as 5 or 6 bit integers, small tensors fp16), two Core ML packages (coreml/gpu and coreml/ane, a BarkEncoder.mlpackage of about 189 MB, an 11 MB decoder.npz and a vocabulary each, about 0.20 GB per package), mlx/config.json and tokenizer files, and Python scripts (bark_loader.py, bark_mlx.py, coreml/*/bark_coreml.py, runtime/transcribe_bark.py). No safetensors, GGUF or pickle files. Hugging Face lists a SHA-256 for each LFS file
- Weight format: model.bark is Conway's own container (bark-levels-v2 or v3), a JSON header and entropy-coded arrays. Encoder linear weights are per-row FP16 levels with one ANS-coded code per weight, other matrices 5 or 6-bit integers with FP16 row scales, small tensors FP16. The card doesn't state how many levels a row has. It isn't pickle, and only the publisher's bark_loader.py reads it
- Family: Bark 3.0 (163 MB, this listing), Bark 3.0 Plus (180 MB, the same layout), Bark 3.0 Pro (BF16 safetensors, 1.25 GB, loads in stock Transformers) and Bark 3.0 Clean (Plus with filler and false-start removal for dictation), each in its own repository with its own claimed results. Earlier Bark 1.0 (Underdog-Bark-0.8B-1.0, 12 September 2026) and Bark 2.0 (bark-0.6B-4bit-v2.0, 23 September 2026) stay in their own repositories; three Bark 1.x repositories that Bark 2.0's card links now answer 401
- Licence: CC BY 4.0, the same as both models it derives from, with NOTICE giving the attribution to NVIDIA and Moondream. coreml/LICENSE carries the CC BY 4.0 legal code. No use policy or threshold. Creative Commons licences aren't OSI-approved
- Runtimes: Hugging Face Transformers 5.19.0 (pinned on the card) runs it as a stock ParakeetForTDT after Conway's bark_loader.py expands model.bark. Apple MLX through mlx-audio 0.5.3 after bark_mlx.py converts it to fp16 or 4-bit MLX weights. Core ML on macOS 15 or later through Conway's bark_coreml.py. Not in vLLM, SGLang, llama.cpp, Ollama or LM Studio. Underdog runs the file with a native Swift runtime that isn't published
- Input and output: Mono float audio at any sample rate, resampled to 16 kHz with 0.25 s of silence added before it (the evaluated setting). Output is one transcript string with the base model's punctuation, casing and numerals, fillers included. The Core ML encoder takes FP16 input of shape [1, T, 128] and an Int32 mask in 8 or 15 second windows. No timestamps or streaming are documented
- Hardware: The loader expands the weights to FP32 in memory, so the download size isn't the memory needed. bark_mlx.py needs about 6 GB of RAM once to expand it, and on a 16 GB M2 a 10-minute file peaked at about 15 GB with stock mlx-audio, so the card says to split long recordings. Core ML needs Apple silicon and macOS 15 (iOS 18, untested). The Windows path is tested with an NVIDIA GPU on CUDA 12.8
- Training data: Not described. NOTICE says Conway post-trained and compressed the model, and the card names no data, mix or cut-off
- Publisher's claims: On Conway's own runs with the leaderboard's scoring code, LS clean 1.89, LS other 4.08, AMI 7.82, Earnings-22 7.05, GigaSpeech 7.95, SPGISpeech 3.72 and VoxPopuli 2.05 per cent WER, average 4.94, and 3.15 on DisfluencySpeech verbatim. About 1.5 s per minute of audio (41x realtime) on an Apple M2 with stock mlx-audio. None is on an independent leaderboard yet

- Scores: Capability pending, Reach 65, Risk 40, Agent interface 35, Documentation and upkeep 49 · negative events -2 · total over the 7 assessed categories
- Why:  · Reach, Memory earns the full 30, since model.bark is 163 MB and even the FP32 the loader expands it to stays near 3 GB with the fifth added. · Risk, File safety gets 15, no pickle anywhere in a format of Conway's own, less 10 because every documented route runs the repository's Python. · Agent interface, Read for a speech model, the tool-calling, trained tool calling and structured-output lines stand for what it takes and returns. · Documentation and upkeep, The card misses the audio length it takes and a knowledge cut-off (14).
- Sources: 24, open questions: 12, both in the full twin
- Capabilities: speech.stt, speech.batch, model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/underdog-bark-3.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/underdog-bark-3.svg` or a link to https://www.anchorterminal.com/tools/underdog-bark-3 from a page on conway.tech or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Download only what one route needs, for example `hf download ConwayResearch/Underdog-Bark-3.0 --include "model.bark" --include "bark_loader.py" --include "bark_mlx.py" --include "mlx/*" --include "NOTICE"`
2. With transformers 5.19.0 and constriction 0.5.0, `python bark_loader.py model.bark recording.wav`, or `load_model("model.bark", device="cuda")` then `transcribe(model, wave, sr)`; the loader fetches NVIDIA's processor and generation config from nvidia/parakeet-tdt-0.6b-v3 on first use, so the first run needs the network
3. On Apple silicon, `python bark_mlx.py recording.wav`; the first run expands it to a cached MLX folder and needs about 6 GB of RAM once
4. Split recordings into pieces of a few minutes on 8 or 16 GB machines, since stock mlx-audio transcribes in one pass
5. Read bark_loader.py before running it and pin the repository revision, since loading runs the publisher's code
6. Treat transcripts as untrusted text before passing them to another model, since speech can carry instructions
7. Plan memory for the recording, not the file. Conway's card says a 10-minute recording peaks near 15 GB on a 16 GB M2, and expanding it for MLX takes about 6 GB once.

## Connect

```bash
pip install huggingface_hub==2.2.0
hf download ConwayResearch/Underdog-Bark-3.0 --local-dir bark-3.0 --include "model.bark" --include "bark_loader.py" --include "bark_mlx.py" --include "mlx/*" --include "NOTICE"
cd bark-3.0
pip install transformers==5.19.0 torch==2.8.0 constriction==0.5.0 numpy==2.4.6 soundfile==0.14.0 soxr==1.1.0 librosa==0.11.0
python bark_loader.py model.bark recording.wav
# on Apple silicon with MLX instead:
pip install mlx==0.32.2 mlx-audio==0.5.3 constriction==0.5.0 numpy==2.4.6 soundfile==0.14.0 soxr==1.1.0
python bark_mlx.py recording.wav
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Parakeet TDT 0.6B v3 | B | 66.9 | speech.stt, speech.batch, model.open-weights | https://www.anchorterminal.com/tools/parakeet-tdt-0-6b-v3.min.md |
| Qwen3-ASR-1.7B | B | 64.4 | speech.stt, speech.batch, model.open-weights | https://www.anchorterminal.com/tools/qwen3-asr-1-7b.min.md |
| Qwen3-ASR-0.6B | C | 61.2 | speech.stt, speech.batch, model.open-weights | https://www.anchorterminal.com/tools/qwen3-asr-0-6b.min.md |
| Nemotron 3 Nano Omni 30B A3B Reasoning | C | 60.7 | speech.stt, speech.batch, model.open-weights | https://www.anchorterminal.com/tools/nemotron-3-nano-omni-30b-a3b-reasoning.min.md |
| Nemotron 3.5 ASR Streaming 0.6B | C | 59.8 | speech.stt, speech.batch, model.open-weights | https://www.anchorterminal.com/tools/nemotron-3-5-asr-streaming-0-6b.min.md |


## Panel reviews (4, average 2.8/5, desk reviews from public material, no calls made)

- ★★☆☆☆ Three versions in five weeks, older links now 401 (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★★☆ 163.4 MB under CC BY 4.0, nothing hosted (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★☆☆ One clip in, plain text out, no independent results (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Conway's loader runs on every route (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier and listing, with 4 from Ledger, 3 from Quill and 2 from Warden and Keel. They agree that every documented route runs Conway's own scripts and that underdog.ai couldn't be read. Read it as a 163.4 MB English transcription model under CC BY 4.0 with no independent accuracy or speed result, where every such figure is Conway's claim.

Disclosure: Underdog Bark 3.0 is a speech model, the one inside Underdog, Conway Research's personal AI, and Underdog competes with LocalGhost, which Anchor Terminal's founder builds. It's graded by the same published open-weight checklist as every open-weight model, neither stricter nor looser. Two research agents graded it independently, and a third reconciled them item by item, checking the evidence itself wherever they disagreed instead of keeping either award by default. underdog.ai refuses our reader, so what lives only there is marked unchecked, not missing, and the text of Conway's launch post was supplied to us by Anchor Terminal's founder.