Muse-Spark-Open-816B-A42B (GGUF UltraOptimised • Curated • Protium)

Official Solstice-AI Quantization Suite • Unsloth Dynamic i-Matrix Quants • Protium-1.0 Objective-Gated Cognitive Architecture • 1M Context

Base Model by Meta Superintelligence Labs • UD Quants & i-Matrix by Unsloth • Curation & Cognitive Architecture by Solstice-AI

---

Executive Summary

Solstice-AI/Muse-Spark-Open-816B-A42B-GGUF-UltraOptimised-Curated-Protium is the curated GGUF release of Meta's Muse Spark Open (meta-models/Muse-Spark-Open-816B-A42B) — an 816B total / 42B active Mixture-of-Experts with a Perception Encoder, a 1M-token context window, and native text + image + video + audio input.

This release ships three Pareto-optimal Unsloth Dynamic (UD) i-matrix checkpoints — no degraded extreme low-bits, no duplicate tiers — plus Solstice-AI's Protium-1.0 Objective-Gated Cognitive Architecture as a drop-in chat_template.jinja at the repo root.

Highlights:

1. Frontier Agentic Coding — Meta-reported: 76.4 Terminal-Bench 2.1, 60 DeepSWE 1.1, 89.2 MCP-Atlas, 80.4 OSWorld-Verified.
2. Native Multimodal Agentic Input — screenshots, charts, documents, audio, and video recordings inside long agent sessions; 1M context holds entire repositories.
3. Unsloth Dynamic i-Matrix Quantization — importance-matrix quants (imatrix_unsloth.gguf included) preserve outlier channels that matter for reasoning in MoE weights.
4. Protium-1.0 Cognitive Architecture — objective-gated, forward-only reasoning with no forced reasoning-strength cap: depth is set by the problem, not by a hardcoded token quota.

---

Curated Checkpoints

Every shard group is a UD (Unsloth Dynamic) tier: importance-matrix quantization calibrated on the actual activation distribution, sharded cleanly for HF download managers and multi-host serving:

Memory budgets are weights + modest KV headroom for tens-of-thousands of context tokens; the 1M window costs proportionally more KV cache, so scale context to your budget or use layer offload.

---

Protium-1.0 Objective-Gated Cognitive Architecture

Built directly into chat_template.jinja at the repo root. Full protocol documentation: Solstice-AI/Protium-1.0.

No Forced Reasoning Quota

Unlike stock, this template never injects a Reasoning strength: line on any default path — no hardcoded 512-token cap, no implicit quota. Reasoning depth is governed by Protium's objective-gated contract: anchor the termination criteria at step one, reason strictly forward, and terminate thinking the instant the solution passes active verification. Simple queries collapse to near-instant answers; hard proofs get full runway — elastic by construction, zero padding by rule.

A strength line renders only when you explicitly ask for one, and only as your own ceiling:

# Explicit native effort passthrough (low=8, medium=32, high=256, xhigh=512 tokens)
response = client.chat.completions.create(
    model="...",
    messages=[{"role": "user", "content": "Analyze system architecture"}],
    extra_body={"chat_template_kwargs": {"reasoning_effort": "xhigh"}}
)
# No reasoning_effort passed → no strength line at all. The model self-regulates.

A Reasoning strength: N. written in your own system prompt always wins and passes through verbatim.

The Cognitive Spectrum

Trigger with inline tags anywhere in your message (stripped before the prompt reaches the model) or via system prompt. Default: [L10] Proteus:

> [L7] Prove the Chapman–Kolmogorov equations hold for time-homogeneous chains.

---

Quickstart

Download

hf download Solstice-AI/Muse-Spark-Open-816B-A42B-GGUF-UltraOptimised-Curated-Protium \
  --include "UD-Q4_K_XL/*" --local-dir muse-spark-protium

Serve with llama.cpp

Important: the GGUF metadata carries the stock Muse template. To run Protium, pass the repo-root template explicitly:

llama-server \
  --model muse-spark-protium/UD-Q4_K_XL/Muse-Spark-Open-816B-A42B-UD-Q4_K_XL-00001-of-00012.gguf \
  --chat-template-file chat_template.jinja \
  --host 0.0.0.0 --port 8080 \
  --ctx-size 131072 -ngl 99

Omit --chat-template-file to get the stock Muse template baked into the GGUF metadata.

Vision note: no separate mmproj projector ships in this repository. Whether image/video/audio input works in your runtime depends on that runtime's Muse Spark support — verify it before relying on multimodal input.

Serve with vLLM / SGLang

vllm serve meta-models/Muse-Spark-Open-816B-A42B \
  --chat-template ./chat_template.jinja \
  --max-model-len 131072

Recommended Sampling (Meta)

For complex problem solving, coding, and agentic tasks Meta recommends native reasoning_effort of high or xhigh — pass it explicitly when you want the ceiling; leave it unset to let Protium self-regulate.

---

Model Overview

Meta-reported benchmarks (xhigh effort, temperature 1.0, top_p 1.0):

Full benchmark and evaluation methodology: Meta's evaluation report.

---

License & Responsible Use

The model weights, this quantization, and all files in this repository are governed by the Muse Model License, including its Usage Policy — not Apache/MIT. Review both before use.

Key points from Meta's model card:

• Deploy as part of a broader AI system with guardrails suited to your use case; the open-weights release carries additional safety post-training for standalone use.
• Avoid exposing raw chain-of-thought in user-facing applications; apply safety checks to any reasoning summaries, and protect internal assistant messages from manipulation.
• Not intended for use by individuals under 18.
• Quantized inference may show minor quality differences vs. full precision.
• Short/low-resolution video (a few seconds at ~100 px) may be rejected by the model; use ≥448 px on the short side for reliability.

---

Citations & Acknowledgments

• Meta Superintelligence Labs for the Muse Spark Open base model and open weights.
• Unsloth for the Dynamic i-matrix quantization and calibration data.
• Solstice-AI for curation (Pareto-optimal tier selection, shard packaging) and the Protium-1.0 Objective-Gated Cognitive Architecture.

@misc{meta2026musesparkopen,
  title        = {Muse Spark Open},
  author       = {Meta Superintelligence Labs},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B}},
}

---

Crafted with precision by Solstice-AI
Objective-Gated Cognition • Zero-Waste Reasoning • Ultra-Optimised Compute