Unsloth 1-bit V4.1 Flash GGUF：n-gram 表放 SSD，128GB 内存可跑
帖中顺带提到 Unsloth 的 1-bit DeepSeek V4.1 Flash-1017 GGUF 通过把 n-gram 表放 SSD、配合未合并的 llama.cpp PR，在 128GB 以上内存运行；主体是蚂蚁 Ling 开源语音模型。
Your voiceovers and transcripts can now stay on your own servers: Ant Ling just open-sourced 5 Ming-3.0 speech models under MIT. • Ming-3.0-tts: a 0.5B model and a 7.9B Tiny (1.3B active) clone or design voices in 7 languages, with laughs and sighs on cue. • Ming-3.0-audio and asr: a 126B Flash (6.5B active) for long-audio understanding, plus Tiny models for dialogue and dialect-aware ASR. • Day-0 serving: vLLM-Omni and SGLang-Omni support it, though Flash's reference setup is eight 80GB H800s. • Also today: Unsloth's 1-bit DeepSeek-V4.1-Flash-1017 GGUF runs on 128GB+ RAM by keeping its n-gram table on SSD, via an unmerged llama.cpp PR. Cloud voice API or your own speech stack: where does your next voiceover get made? https://huggingface.co/collections/inclusionAI/ming-30-6ad62ac04ca3a3f15e976f34