Edge0's 'Open' Speech Model Has a Non-Commercial Decoder Inside

Edge0-AI published Audio8 ASR Infinite on GitHub on Sept. 21, 2026, badging the 4-billion-parameter streaming speech-recognition model Apache 2.0. Five days later, a GitHub user found that roughly 3 billion of those parameters are a fine-tuned copy of Alibaba's Qwen2.5-3B-Instruct, which Alibaba licenses for non-commercial use only.

Audio8 ASR Infinite transcribes Chinese and English audio continuously using a rolling key-value cache, which Edge0 says lets it run indefinitely without the memory growth that limits most streaming transformers. It decodes up to 12.5 times a second and offers a selectable 80, 120 or 160 millisecond audio clock, according to the model's Hugging Face card.

What the licence issue found

GitHub user xocialize, evaluating the model for a commercial on-device dictation product, filed issue #2 on Sept. 26, 2026. The issue points to the model's own config.json, which lists text_config._name_or_path as Qwen/Qwen2.5-3B-Instruct, and to matching architecture details: 36 layers at 2,048 dimensions, grouped-query attention at 16 and 2 heads, and a chat template identical to Qwen2.5's.

Unlike Qwen2.5's 0.5B, 1.5B, 7B, 14B and 32B checkpoints, Qwen2.5-3B ships under the Qwen Research License Agreement rather than Apache 2.0. That agreement grants use "FOR NON-COMMERCIAL PURPOSES ONLY," says the licence is non-transferable, and requires a "Built with Qwen" notice on derivative works, none of which appears in Edge0's repository or model card, the issue says.

The filer proposed three fixes: swap in a permissively licensed decoder such as Qwen3-4B or Mistral's Ministral-3-3B, add a licence note naming the decoder's origin, or disclose whether Edge0 has a separate commercial arrangement with Alibaba. Edge0 had not replied to the issue as of Sept. 27, 2026.

A packaging bug on top of the licence gap

A day earlier, on Sept. 25, GitHub user scrappylabsai filed issue #1, reporting that the documented torch_streaming_decode.py load path throws a RuntimeError over missing weights. The cause: the Hugging Face repository ships the model's core weights in model.safetensors plus a separate semantic_vad_heads.safetensors, but the Transformers library loads only the first file and never reads the index that points to the second, so Edge0's own strict loading check fails. Edge0's vLLM serving path is unaffected, since it reads every file in the index, the issue says.

The same reporter benchmarked the model on an RTX 5090 laptop GPU against a 300-utterance LibriSpeech sample and reproduced Edge0's published accuracy claims.

Model	test-clean WER	test-other WER
Audio8 ASR Infinite (measured)	2.97%	6.96%
Audio8 ASR Infinite (README claim)	3.04%	6.81%
faster-whisper base.en	4.42%	11.61%

On a 10-minute real-time stream through Edge0's own vLLM server, the reporter measured a 3.68 percent word error rate with no drift and final text arriving 0.52 seconds after the audio ended. ScrappyLabs then published its own inference engine built on Edge0's weights, audio8-asr-lean, claiming 12.4 milliseconds per decoding step at 4-bit precision against roughly 16.6 milliseconds for Edge0's reference implementation.

Edge0 founder Samuel Zeng, who led machine-translation work at Alibaba before starting Audio8.ai, has not commented publicly on either issue. As of Sept. 27, 2026, Audio8 ASR Infinite's GitHub repository, its licence file and its Hugging Face card all still list Apache 2.0 with no mention of Qwen's terms. The pattern matches Alibaba's own Qwen-Image-2.1 release, which this site found had swapped Apache 2.0 for a non-commercial research licence without changing its "open weights" framing: a restriction on Alibaba's own model, inherited a second time by someone else's.