← νŠΈλ Œλ”©

Tangerine IndexλŠ” ChatGPT 이후 λ‚˜μ˜¨ ν”„λ‘œλ•νŠΈμ™€ μ‚¬λžŒλ“€μ˜ λ°˜μ‘μ„ μ‹£μŠ΅λ‹ˆλ‹€.

llama-modchip

github.com

πŸ‹ Run 284B DeepSeek-V4-Flash (<59GiB RAM, 44.4 tok/s pp, 9.1-14.7 tok/s) & 30B MoEs (25-111 tok/s) on a pure 64GB CPU at 256K ctx (0 GPUs). SwiGLU-H2, Stockfish-PVS L3 & 13ms Claude Code restore.

μ˜€ν”ˆμ†ŒμŠ€

github.com λ°©λ¬Έμ™ΈλΆ€ μ‚¬μ΄νŠΈGitHub635 μŠ€νƒ€μ™ΈλΆ€ μ‚¬μ΄νŠΈ

λ°˜μ‘ (2)

λΉ„μŠ·ν•œ ν”„λ‘œλ•νŠΈ

μ„€λͺ…이 λΉ„μŠ·ν•œ 순

  1. I run 30B 22tok/s, 109tok/s not novel,6GB/16GB RAM overcoming llama.cpp

    Run a 110B on a 2016 PC with 16 GB RAM. Know your tok/s before you download. Placement beats budget: predicts speed + memory fit for any GGUF on your exact hardware, self-calibrates, emits the exact llama.cpp command β€” or 'quantprobe auto' does it all. Falsification-tested laws; misses published at full size. pip install quantprobe

    Hacker News5점GitHubμŠ€νƒ€ 93github.comμ˜€ν”ˆμ†ŒμŠ€λ°˜μ‘ 2개졜근 λ°˜μ‘

  2. I shrank DeepSeek V4 Flash to 57GB and it wrote a compiler on my Mac

    Text Generation model. Download, explore files and run it on Hugging Face.

    Hacker News21점huggingface.co졜근 λ°˜μ‘

  3. Llama 3.1 70B on a single RTX 3090 via NVMe-to-GPU bypassing the CPU

    High-efficiency LLM inference engine in C++/CUDA. Run Llama 70B on RTX 3090.

    Hacker News395점GitHubμŠ€νƒ€ 469github.comμ˜€ν”ˆμ†ŒμŠ€λ°˜μ‘ 2개졜근 λ°˜μ‘

  4. vLLM-Moet

    Frontier MoE models on RTX PRO 6000 / RTX 5090 (sm_120): DeepSeek-V4.1-Flash on 4x RTX PRO 6000, as a GPU+CPU hybrid on 4x RTX 5090, and 2-bit experts on vLLM v0.24.0

    GitHubμŠ€νƒ€ 571github.comμ˜€ν”ˆμ†ŒμŠ€μ΅œκ·Ό λ°˜μ‘

  5. Finetune Llama-3.1 2x faster in a Colab

    Just added Llama-3.1 support!

    Hacker News16점colab.research.google.com졜근 λ°˜μ‘