# Running 176B Qwen3.8-Flash-Next and 284B DeepSeek-V4 on 32GB/64GB CPU RAM (0 GPUs) with llama-modchip v3.0

> Source: <https://discuss.huggingface.co/t/running-176b-qwen3-8-flash-next-and-284b-deepseek-v4-on-32gb-64gb-cpu-ram-0-gpus-with-llama-modchip-v3-0/192008#post_1>
> Published: 2026-11-18 00:29:55+00:00

**Key takeaway:** llama-modchip v3.0 runs the 176.5B Qwen3.8-Flash-Next on 32-64 GB of CPU RAM with no GPU at 14-18.2 tok/s on a Ryzen 9 7950X/9950X, using the model's built-in blk.48 MTP draft head for speculative decoding.

We just shipped llama-modchip v3.0: 176.5B Qwen3.8-Flash-Next now runs on 32 GB and 64 GB CPU RAM with 0 GPUs (about 14-18.2 tok/s on a desktop Ryzen 9 7950X/9950X using its built-in blk.48 MTP draft head, <1s multi-turn TTFT in Claude Code).

We also updated our 284B DeepSeek-V4-Flash (fits in 64 GB RAM, ~10.1 tok/s on DDR5) and 30B Nemotron-3.5 (16 GB RAM, 30-60 tok/s) builds, plus full v3.0-r2 abliterated variants for local security/red-team testing.

• GitHub (code + 1-line installer): [GitHub - synchopate/llama-modchip: Run 284B DeepSeek-V4-Flash (<59GiB RAM: 41.1 cold pp, 9.1 N=1 tg, up to 25.3 spec tok/s) & 30B MoEs (22.4 N=1 / 111.6 spec tok/s) on 64GB CPU (0 GPUs). · GitHub](https://github.com/synchopate/llama-modchip)

• v3.0 Model Collection (16GB / 32GB / 64GB): [MaxQ v3.0 - a Synchopate Collection](https://huggingface.co/collections/Synchopate/maxq-v30-6afb6e2d63dfd573b4b1a95f)

If you test it on your CPU, let me know what hardware and tok/s you get!
