Muse-Spark-Open-816B-A42B-DFlash2 Vision-Language Model: 12GB+ VRAM

Muse-Spark-Open-816B-A42B-DFlash2 Vision-Language Model: 12GB+ VRAM

At a Glance

Item Value
Repository nvidia/Muse-Spark-Open-816B-A42B-DFlash2
Publisher guide NVIDIA: models and licenses
Published 2026-10-16
License nvidia-open-model-agreement-license
Formats safetensors
Paper arXiv:2602.06036
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

NVIDIA has released a DFlash 2 draft checkpoint, nvidia/Muse-Spark-Open-816B-A42B-DFlash2, to accelerate inference for the multimodal MoE (Mixture-of-Experts) causal language model Muse-Spark-Open-816B-A42B. Note that the target model Muse-Spark-Open-816B-A42B was not developed by NVIDIA and is built based on third-party requirements. This model is not intended to operate as a standalone target model, but rather to assist in serving the target model via speculative decoding using DFlash 2.

DFlash 2 is a technology leveraging the NVIDIA Model Optimizer, featuring a two-tap in-block convolution around each sub-layer and a pairwise path selector that follows a consistent path among top-k candidates at each position. This enables efficient draft generation. Technical details are described in the paper “DFlash: Block Diffusion for Flash Speculative Decoding“.

Specifications

  • Total Parameters: 4.38B (token embeddings and output head are shared with the target model)
  • Active Parameters: 4.38B (dense)
  • Architecture: Transformer (5 dense decoder layers, non-causal sliding-window attention, window 2048)
  • Context Length: Up to 1,000,000 tokens

Performance

According to measurements by the publishers, the model’s performance is evaluated using an Acceptance Length (AL) metric. This indicates how many tokens proposed by the draft model in a single step are accepted by the target model. These figures were measured using NVIDIA GB300 nodes under conditions of reasoning effort xhigh, 4,096 output tokens, and concurrency 32.

The results for greedy decoding (t=0) on MT-Bench are as follows:

Category AL, greedy (t=0) AL, sampling (t=1)
writing 3.60 2.95
roleplay 3.18 2.64
reasoning 3.52 2.83
math 4.47 3.51
coding 4.34 3.43
extraction 4.60 3.43
stem 3.23 2.73
humanities 2.92 2.55
Overall Average 3.73 3.01

The results on the qualitative split of SPEED-Bench are as follows:

Category AL, greedy (t=0) AL, sampling (t=1)
coding 4.15 3.31
humanities 3.23 2.61
math 3.95 3.13
multilingual 4.27 3.47
qa 3.43 2.76
rag 4.17 3.26
reasoning 3.85 3.09
roleplay 3.39 2.74
stem 3.42 2.83
summarization 4.04 3.24
writing 3.05 2.65
Overall Average 3.72 3.01

The results on the throughput split of SPEED-Bench are as follows:

Category AL, greedy (t=0) AL, sampling (t=1)
mixed 3.68 3.02
low_entropy 3.71 3.17
high_entropy 2.51 2.16
Overall Average 3.30 2.78

These tables show that the model exhibits high Acceptance Length in tasks such as extraction (4.60), math (4.47), and coding (4.34), indicating strong drafting capabilities in specific logical and structural tasks. High figures are also observed in multilingual (4.27) and rag (4.17). On the other hand, under high_entropy conditions, the Acceptance Length drops to 2.51 (t=0), showing that the acceptance rate by the target model tends to decrease in cases with high entropy (high prediction uncertainty).

Strengths and Use Cases

This model is a draft model designed to assist inference for the target model Muse-Spark-Open-816B-A42B, and its performance depends on the characteristics of the base model. Intended use cases for the base model include the evaluation and development of long-context and agentic text generation workflows, including tool-assisted coding.

Specifically, it features the following capabilities:

  • Multimodal Support: Supports text, image, video, and audio inputs.
  • Long Context: Handles context lengths up to 1,000,000 tokens combined for input and output.
  • Agentic and Tool Usage: Intended for tool-assisted coding and agentic workflows.

Due to its characteristics as a draft model, it is well-suited for efficiently predicting target model outputs and accelerating inference, particularly in tasks requiring logical or structural consistency such as math, coding, and extraction.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 4.4B parameters

Your VRAM Quantization File size Est. memory needed
12GB (RTX 4070 / 3060 12GB, etc.) BF16 8.2GB 9.8GB

Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: registered, vLLM: registered, MLX (mlx-lm): registered.

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Our Own Measurements

Measurements We Did Not Take

We have not confirmed that this site meets the commercial-use terms of this model’s license (nvidia-open-model-agreement-license). Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).

Recent Models in the Same Size Class

Models with 4–15B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.

→ Scroll horizontally to see all columns

Model Parameters Smallest VRAM tier License Our article
tencent/Youtu-Parsing-Omni 5.3B 12GB youtu-parsing Youtu-Parsing-Omni Omnimodal Analysis Model: 12GB+ VRAM (2026-11-17)
tencent/WeVisOne-VL-4B 4.8B 4GB apache-2.0 Tencent Releases WeVisOne-VL Vision Language Models (2026-11-17)
inclusionAI/ZwZ2-9B-GGUF 9.4B 8GB apache-2.0 ZwZ2-9B-GGUF Vision-Language Model: 8GB+ VRAM (2026-11-17)
inclusionAI/ZwZ2-4B-GGUF 4.5B 4GB apache-2.0 ZwZ2-4B-GGUF Vision-Language Model: 4GB+ VRAM (2026-11-17)
microsoft/Mage-VL-Cowork 4.7B 12GB apache-2.0 Mage-VL-Cowork Vision-Language Model: 12GB+ VRAM, GGUF Builds (2026-11-17)

How to Get It

  • Distribution format: safetensors
  • Supported engines: TRT-LLM, vLLM

When using vLLM alongside the target model, start it by specifying --speculative-config as shown in the following command:

vllm serve meta-models/Muse-Spark-Open-816B-A42B \
    --speculative-config '{"method": "dflash", "model": "nvidia/Muse-Spark-Open-816B-A42B-DFlash2", "num_speculative_tokens": 7, "draft_sample_method": "probabilistic"}' \
    --trust-remote-code

Use of this model is subject to the NVIDIA Open Model Agreement.

Quantized and Converted Variants

→ Scroll horizontally to see all columns

Added Publisher Format Repository Smallest VRAM tier (build, est. memory)
2026-10-16 ggml-org GGUF ggml-org/Muse-Spark-Open-816B-A42B-GGUF Q2_K 339.8GB (does not fit a single consumer GPU)

File sizes of each build:

  • Available builds in ggml-org/Muse-Spark-Open-816B-A42B-GGUF: Q2_K 283.2GB / Q4_K_S 437.8GB / Q8_0 811.8GB / BF16 1527.6GB

This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.

Related Articles

What to Read Next

Sources