When will new paramformer large model opensource?

[MonolithFoundation]

The current paraformer is good, but not good enough, any plan to opensource upgrade model?

[LauraGPT]

Short answer: there is still no public release date or committed ETA for a successor specifically named a new Paraformer-Large checkpoint.

As of 2026-07-16, the current FunASR model table still lists the public offline and streaming Paraformer models as the 220M paraformer-zh family. The releases published after this question was opened do not announce a successor Paraformer checkpoint, so it would be misleading to promise a date that has not been published. Current model table Releases

What has changed since this question was opened is the recommended model choice:

• Keep Paraformer-zh when you need the mature non-autoregressive Chinese/English path, character timestamps, hotwords, or the streaming Paraformer contract.
• On a GPU, start an accuracy comparison with Fun-ASR-Nano-2512 for difficult Chinese/English/Japanese audio, Chinese dialects/accents, context, and proper nouns.
• Use the separate Fun-ASR-MLT-Nano-2512 checkpoint when the requirement is the published 31-language scope.
• On CPU, or when emotion/audio-event output is part of the requirement, compare SenseVoiceSmall.

The Nano checkpoints are not drop-in replacements for every Paraformer output contract. In particular, do not migrate only because the model is newer if your application depends on Paraformer timestamps, hotwords, streaming behavior, latency, or memory usage. The current selection guide explicitly keeps Paraformer for Mandarin timestamp/hotword workloads and recommends benchmarking the alternatives on representative audio. Selection guide Benchmark checklist

Minimal GPU comparison:

from funasr import AutoModel

# Current Paraformer production contract
paraformer = AutoModel(
    model="paraformer-zh",
    vad_model="fsmn-vad",
    punc_model="ct-punc",
    device="cuda:0",
)

# Accuracy/context candidate for zh/en/ja + Chinese dialects/accents
nano = AutoModel(
    model="FunAudioLLM/Fun-ASR-Nano-2512",
    device="cuda:0",
)

# Separate multilingual checkpoint
mlt_nano = AutoModel(
    model="FunAudioLLM/Fun-ASR-MLT-Nano-2512",
    device="cuda:0",
)

Use the same 20-50 representative files and record CER/WER, latency, throughput, peak memory, model revision, and required output fields. That gives an actionable upgrade decision today; a future Paraformer release should only be treated as available after it appears in the official model table or release notes.