Oriel — AI & technology worth noticing, every day.

unsloth/Qwen3.8-Flash-Next-GGUF

AI model · https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF

First seen by Oriel on .

AI summary

The Hugging Face page presents Qwen3.8-Flash-Next, an experimental open-weight architecture preview built for efficient long-context language and vision tasks, and links to Unsloth GGUF quantizations for local inference.

  • The model combines Gated DeltaNet with Qwen Sparse Attention, which selects micro-blocks to reduce long-context latency, alongside Gated Residual and n-gram embeddings.
  • It has 125B parameters with 6B activated, plus 51B n-gram embedding parameters and 4B MTP; its context is 262,144 tokens natively and can be extended to 1,000,000.
  • The model overview reports 512 experts, with 10 routed and one shared expert activated, and 20 million bigram/trigram embeddings.
  • Benchmark results show 62.5 on SWE-bench Pro, 73.5 on Toolathlon Verified, and 84.5 on AndroidWorld; scores vary across tasks and comparison models.
  • The page recommends llama.cpp or Unsloth Desktop for running the GGUF release and describes thinking-mode controls, sampling settings, and YaRN configuration for longer contexts.

#gguf #large language models #sparse attention #llm inference

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Likes

Total likes reported by the source.

  • Hugging Face Models: 1,835, latest observation ; +445 over observed period since

Downloads (30d)

Rolling 30-day downloads. A change in this value is not new downloads during the observed period.

  • Hugging Face Models: 2,673,551, latest observation ; +1,479,547 over observed period since

Aliases

SourceExternal IDURL
Hugging Face Modelsunsloth/Qwen3.8-Flash-Next-GGUFhttps://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF Read discussion