Introduction

Awaken just crossed the “makers can actually run it” line. On 2026-11-04 (~15:25 UTC) openmartdev published open weights at openmartdev/awaken-full-duplex — a full-duplex spoken-dialogue adaptation of FunAudioLLM/Fun-Audio-Chat-8B. Today (2026-11-06) the missing piece landed: the installable code repo avera-labs/AwakenFullDuplex (GitHub created ~03:06 UTC, project page at avera-labs.github.io/AwakenFullDuplex).

Earlier soft-holds stalled on a dead install path (OpenmartAI/awake-audio-chat 404). That link is gone from the card; clone avera-labs instead and the HF checkpoint is usable.

What Shipped

  • Vendor / product: openmartdev / avera-labs — Awaken Full Duplex
  • Modality: open-weight full-duplex speech-to-speech / spoken dialogue (listen + speak with overlap)
  • Open vs closed: Apache-2.0 weights + code; base Fun-Audio-Chat-8B
  • Runnable today: git clone https://github.com/avera-labs/AwakenFullDuplex → CUDA pip install -e . → hf download openmartdev/awaken-full-duplex
  • Concrete fact: four sharded safetensors totaling ≈ 18.85 GB; a short GRPO stage lifts backchannel pass rate from 1% (SFT alone) to 63% while “taking the floor” on only 3% of items; groundedness probes 61% (base 66%); spoken logic puzzles 90%

Primary surfaces: the HF card (created Nov 4, lastModified Nov 6) and today’s GitHub + project page.

Why full-duplex matters here

Half-duplex chat models wait their turn. Full-duplex ones can acknowledge, interrupt, and hand over mid-utterance — but training that usually burns scarce natural conversation audio and can wreck reasoning. Awaken’s pitch is efficient adaptation: two supervised stages plus a short RL stage on composed conversations with explicit timing, overlap, and interruptions — no large-scale duplex pretraining.

Architecture-wise, user features, text embeddings, and audio-token embeddings share a fusion layer into a transformer that predicts the next text token and conditions the audio decoder (see the card’s architecture figure). The speech codec still pulls from FunAudioLLM’s CosyVoice stack at runtime.

Run It

git clone https://github.com/avera-labs/AwakenFullDuplex
cd AwakenFullDuplex

REM: pick the torch wheel for YOUR CUDA driver (e.g. cu128)
pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu128
pip install -e .            # Linux + CUDA
pip install -e ".[fast]"    # optional flash-attn
pip install -e ".[voice]"   # optional onnxruntime for voice prompts from a clip

pip install huggingface-hub
hf download openmartdev/awaken-full-duplex --local-dir ./pretrained_models/awaken-chat

Follow the repo’s Quick Start after the download — codec weights for CosyVoice fetch automatically from FunAudioLLM’s published cards.

Notes for makers

  • This is spoken dialogue, not a drop-in TTS API — plan GPU + the ~19 GB weight footprint
  • Distinct from ArtRealmAI’s earlier full-duplex coverage (MOSS-Live, Block Dessa preview, Realtime Venus) — Awaken is specifically a Fun-Audio-Chat-8B duplex adaptation with published GRPO backchannel numbers
  • Soft-hold cleared only because avera-labs/AwakenFullDuplex returns 200; if that repo moves again, re-verify before production installs
  • Card still lists evaluation figures for backchannels, groundedness, logic, and QA — treat them as author-reported, not third-party arena Elo
  • No Gen CTA: full-duplex voice chat does not map to the Flux / Grok / MiniMax-H3 / Seedream / Seedance Gen loop

Conclusion

If you want open full-duplex voice that keeps Fun-Audio-Chat’s head on its shoulders, Awaken is finally runnable today: Apache-2.0 HF shards, avera-labs install path, and a GRPO recipe that pushes backchannels to 63% without a duplex pretrain.

—Anabel ♡