Pinned post
AveLabs is an independent lab building open models you can run on your own hardware. Today we're releasing our first one.
pplx-decider-v1-27b can only pick options, it can't talk. Restored chat to it by grafting the LM head back.
Meet Qwen3.8-27B Perplexing-Pegasus: one model that makes typed decisions and holds normal conversations.
The finding
Perplexity's decider replaces Qwen3.8's language-model head with a 255-option readout. We found those readout rows are identical to Qwen3.8-27B's LM-head rows for the answer tokens (cosine 1.0000). The fine-tuned backbone never left Qwen's output space, so we put the full LM head and the MTP head back.
Decisions: unchanged
• 76.0% vs 76.0% on 2,196 decisions against the original, 96.7% identical answers
• JevBench public: 199–200 / 231
Chat: better than stock Qwen3.8-27B (same settings)
• IFEval 85.2 vs 84.3
• GSM8K 87.5 vs 86.0
• MMLU-Pro 81.6 vs 75.8
• Vision, tool calls and streaming reasoning all work
Runs where you are
• NVFP4 for Blackwell (SGLang, 262K context)
• GGUF for llama.cpp, Q8 down to IQ2_M: Q4_K_M fits a 24 GB card at the full 256K context, IQ2_M runs on 12 GB
Apache-2.0. Built on @perplexity_ai's pplx-decider and @Alibaba_Qwen's Qwen3.8 🙏
BF16: huggingface.co/AveLabs/Qwen3.…
NVFP4: huggingface.co/AveLabs/Qwen3.…
GGUF: huggingface.co/AveLabs/Qwen3.…
pplx-decider-v1-27b can only pick options, it can't talk. Restored chat to it by grafting the LM head back.
Meet Qwen3.8-27B Perplexing-Pegasus: one model that makes typed decisions and holds normal conversations.
The finding
Perplexity's decider replaces Qwen3.8's language-model head with a 255-option readout. We found those readout rows are identical to Qwen3.8-27B's LM-head rows for the answer tokens (cosine 1.0000). The fine-tuned backbone never left Qwen's output space, so we put the full LM head and the MTP head back.
Decisions: unchanged
• 76.0% vs 76.0% on 2,196 decisions against the original, 96.7% identical answers
• JevBench public: 199–200 / 231
Chat: better than stock Qwen3.8-27B (same settings)
• IFEval 85.2 vs 84.3
• GSM8K 87.5 vs 86.0
• MMLU-Pro 81.6 vs 75.8
• Vision, tool calls and streaming reasoning all work
Runs where you are
• NVFP4 for Blackwell (SGLang, 262K context)
• GGUF for llama.cpp, Q8 down to IQ2_M: Q4_K_M fits a 24 GB card at the full 256K context, IQ2_M runs on 12 GB
Apache-2.0. Built on @perplexity_ai's pplx-decider and @Alibaba_Qwen's Qwen3.8 🙏
BF16: huggingface.co/AveLabs/Qwen3.…
NVFP4: huggingface.co/AveLabs/Qwen3.…
GGUF: huggingface.co/AveLabs/Qwen3.…
huggingface.co AveLabs/Qwen3.8-27B-Perplexing-Pegasus · Hugging Face Image-Text-to-Text model with 28B parameters. Download, explore files and run it on Hugging Face. 
