llama-modchip is a set of ggml/llama.cpp kernel patches, quantization transcoders and runtime tools for running mixture-of-experts models entirely in system RAM on CPUs, with an optional GPU for hybrid runs.
Follow llama-modchip to get an email when its listing changes, such as new pricing or features. One digest a day at most, and you can unsubscribe anytime.
Its Fractal-64 quantization fits the 284B DeepSeek-V4-Flash into 64 GB of RAM, and KV slot snapshots let coding-agent prompts restore in milliseconds, with a `modchip claude` command that wires Claude Code to the local server.
- 01
Fractal-64 and Fractal-MXFP4 quantization that fits 284B DeepSeek-V4-Flash in 64 GB RAM and 30B MoEs in 16 to 32 GB
- 02
Cache-tiled SIMD kernel patches for AVX-512, AVX2, ARM64 NEON/SVE and Apple Metal
- 03
Self-speculative n-gram drafter and `/dev/shm` KV slot snapshots for fast agent context restore
- 04
`modchip` CLI for setup, chat, web HUD, benchmarks, evals, quantization and `modchip claude` to launch Claude Code against the local 256K-context server
- 05
Hardware readiness check and a community CPU benchmark leaderboard
Free
llama-modchip is free and open-source under the Apache-2.0 license with no paid tiers. You supply your own hardware: 16 to 32 GB of RAM for 30B MoE models or 64 GB for the 284B DeepSeek-V4-Flash build.
Best for
Developers with a 64 GB desktop or workstation and no large GPU who want to run big open MoE models locally as a private backend for Claude Code and other coding agents.
Not ideal for
Users who want stock llama.cpp, Ollama or LM Studio compatibility (Fractal-64 GGUFs need the patched binaries) or a mature, multi-maintainer project, since it launched in late October 2026.
An ambitious CPU-first MoE runtime that makes very large open models usable on ordinary desktops for local agent work, best treated as an experimental fork until its patches settle.
v2.4.0 shipped on November 4, 2026 after v2.2.0 and v2.3.0 on November 2 and 3, 2026; commits on November 9, 2026 added macOS and Windows hardware detection and automatic Fractal-64 GGUF selection. The project reached the Hacker News front page in early November 2026.
As of
Is llama-modchip free?+
Yes - llama-modchip is Free. llama-modchip is free and open-source under the Apache-2.0 license with no paid tiers. You supply your own hardware: 16 to 32 GB of RAM for 30B MoE models or 64 GB for the 284B DeepSeek-V4-Flash build.
Is llama-modchip open source?+
Yes - llama-modchip is open source. llama-modchip is free and open-source under the Apache-2.0 license with no paid tiers. You supply your own hardware: 16 to 32 GB of RAM for 30B MoE models or 64 GB for the 284B DeepSeek-V4-Flash build.
Who is llama-modchip best for?+
Developers with a 64 GB desktop or workstation and no large GPU who want to run big open MoE models locally as a private backend for Claude Code and other coding agents.
Who is llama-modchip not ideal for?+
Users who want stock llama.cpp, Ollama or LM Studio compatibility (Fractal-64 GGUFs need the patched binaries) or a mature, multi-maintainer project, since it launched in late October 2026.
What are the best llama-modchip alternatives?+
The closest llama-modchip alternatives on ai.dosa.dev are Copperhead, NVIDIA Personal AI Router (PAIR), WebLLM - all listed under Hardware & Edge AI.
Who makes llama-modchip?+
llama-modchip is developed by Synchopate. It is listed in the Hardware & Edge AI category on ai.dosa.dev.
Favorite this tool to revisit it later, or Zap it to contribute to the public vote count.