Muse-Spark-Open-816B-A42B — 365GB (GGUF)

Mixed-precision quantized version of meta-models/Muse-Spark-Open-816B-A42B using a proprietary quantization method by baa.ai.

Per-tensor bit-width allocation via advanced sensitivity analysis and budget-constrained optimisation — no calibration data required. Built at the efficiency knee (best quality-per-GB).

Built with Muse Spark Open.

Metrics

Routed experts Q3_K; attention Q8_0 (12 attn_q at Q6_K, one attn_v at F16); attention output gate and expert latent projections Q8_0; output and token-embeddings Q6_K; router, norms and residual gates F32. The per-tensor spec is included as spec_muse_spark_knee.txt.

Files

Requirements

llama.cpp build ≥ b11671 is required — Muse Spark Open support (#30616) landed on 2026-10-15 and is not in Homebrew llama.cpp 0.6.0 or the v0.7.0 release. Build from source or use a newer release until downstream runtimes catch up.

Usage

git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
cmake -B build -DGGML_METAL=ON && cmake --build build -j --target llama-cli llama-server

hf download baa-ai/Muse-Spark-Open-816B-A42B-RAM-365GB-GGUF --include "*.gguf" --local-dir ./muse-spark-ram-365gb

# The model is split into 22 shards — point -m at the first; llama.cpp loads the rest automatically
./build/bin/llama-cli -m ./muse-spark-ram-365gb/Muse-Spark-Open-816B-A42B-RAM-local-knee-00001-of-00022.gguf -p "Hello!" -n 256 -ngl 99 --jinja

Or via llama-server for an OpenAI-compatible HTTP API (add --mmproj for image and audio input):

./build/bin/llama-server -m ./muse-spark-ram-365gb/Muse-Spark-Open-816B-A42B-RAM-local-knee-00001-of-00022.gguf \
    --mmproj ./muse-spark-ram-365gb/mmproj-Muse-Spark-Open-816B-A42B-F16.gguf \
    --port 8080 -ngl 99 --ctx-size 8192 --jinja

For fast inference use a host with ≥ ~370 GB RAM/VRAM; otherwise it runs via mmap paging. The chat template is embedded (the current upstream template, including tool namespaces); reasoning effort is set per request with "chat_template_kwargs": {"reasoning_effort": "low" | "medium" | "high" | "xhigh"} (default xhigh). No importance matrix (imatrix) was used — per-tensor sensitivity allocation provides the primary quality signal.

Smoke-tested on a 192 GB Apple M2 Ultra via mmap paging (CPU, -ngl 0): correct answers on factual, arithmetic and code prompts at ~1 tok/s generation and ~0.6 tok/s prompt processing. WikiText-2 perplexity was not measured — the model does not fit in that host's memory.

License

Use of this model is subject to the Muse Spark Open license and its Usage Policy, both included in this repository. Products or services using it must display "Built with Muse Spark Open", and Model-as-a-Service providers above the license's revenue threshold need a separate agreement with Meta.

---

Quantized by baa.ai

---

Black Sheep AI Products

Black Sheep AI builds private AI with the Bounded AI Architecture (BAA): three layers, each bound to something you control, on infrastructure you control. Nothing leaves your boundary.

Ram (Reasoning Layer) — Full-capability frontier models, shrunk to fit smaller GPUs with reasoning kept and certified. They rely strictly on your Data Modules for knowledge, so you swap the model as better ones ship.

Paddock (Knowledge Plane) — Binds your Data Modules directly to the reasoning layer: one per topic, period, person, permission or tenant. Permission, freeze, hand back or delete each one on its own.

Watchman (GRC Layer) — Governance, risk and compliance in a single, complete audit log: every module permission, every bind, every model's provenance. Five certificates before anything goes live.

Shepherd — Takes a frontier model, compresses it into a Ram reasoning model, certifies it with Watchman, binds it to your Paddock, and deploys the Bounded AI Architecture to your infrastructure: your cloud, a Mac Studio, or an air-gapped facility.

Learn more at baa.ai — Private AI for enterprises whose data cannot leave the building.