Muse Spark Open versus Nemotron: memory requirements and the limits of task-cost claims

Bolt’s 24GB label meets a 28GB deployment recipe

NVIDIA’s selected Nemotron 3.5 Bolt artifact carries a 24GB label. Its serving recipes budget 28GB and 32GB. The 28GB recipe targets an RTX Spark with 64GB or a discrete GPU with at least 32GB. The architecture-search target is therefore an incomplete guide to server capacity. Developers considering a 24GB GPU should check the full serving configuration before buying hardware. Sources: NVIDIA

Meta promotes Muse Spark Open on two M5 Ultra Mac Studios. Hugging Face reports a separate single-M3-Ultra experiment. Super VL provides a DGX Spark quick start. Published results do not provide a matched comparison of startup, memory growth and latency across these configurations. They also do not compare repeated task completion and electricity use. The table below separates deployment recipes from the reported Muse measurement. These differences support a deployment shortlist, but not a purchase recommendation or task-cost ranking. Sources: NVIDIA, Hugging Face, Meta AI Research

Muse’s single-Mac result uses a 2-bit llama.cpp configuration

Meta describes Muse Spark Open as an 816-billion-parameter model with 42 billion active parameters. Its maximum context is one million tokens. Its promoted setup uses two M5 Ultra Mac Studios. Each has up to 512GB of unified memory and quoted bandwidth of 1.2TB/s. Hugging Face instead used one M3 Ultra with 512GB for a tested 2-bit llama.cpp configuration. Results from that machine should not be transferred to the paired setup. Sources: Hugging Face, Meta AI Research

Hugging Face reports memory use of 304GB and throughput of 18 tokens/s for the tested 2-bit llama.cpp configuration. Sources: Hugging Face

Hugging Face separately documents oMLX and mlx-lm availability as other deployment paths. It says the described oMLX path cannot disable reasoning. That oMLX path also maps Auto to xhigh reasoning, a setting to preserve when comparing latency and token use. Sources: Hugging Face

Hugging Face’s NVFP4 reference example uses SGLang with four GB300 GPUs. Engine support should therefore be checked for the specific checkpoint, precision, parser and hardware combination. For developers with a 512GB M3 Ultra, the tested 2-bit llama.cpp configuration offers a starting point for a local trial. Sources: Hugging Face

Nemotron’s recipes expose additional memory requirements

The selected Bolt artifact has 74 billion parameters, 9 billion active parameters and IQ2 quantization. Its card names llama.cpp, Windows, Linux, Ada and Blackwell support. It recommends the latest source rather than a fixed runtime commit. Both recipes add a separate drafter and multimodal projector. The 28GB recipe sets 65,536 tokens, one sequence, six DFlash draft tokens and a q8_0 key-value cache. The 32GB recipe sets 262,144 tokens, one sequence and seven draft tokens. It targets RTX Spark 128GB or a discrete GPU with at least 48GB. Sources: NVIDIA

Bolt’s introductory command sets a third context limit: 98,304 tokens. These differences make the exact recipe more useful than the product label for capacity planning. NVIDIA’s announcement promotes a 24GB fit. Bolt’s architecture search targets below 24GB at batch size 1 and 128K context. That target accounts for quantization and runtime-state precision. The card separately lists 28GB and 32GB serving recipes. Reducing context, disabling speculation or adding offload would create a different configuration to evaluate. Sources: NVIDIA

The selected Super VL artifact has 120 billion parameters, 12 billion active parameters and mixed NVFP4 precision. Its maximum context is one million tokens. Its DGX Spark quick start specifies vLLM v0.32.0 and one sequence. It sets GPU memory utilization to 0.75 and allows 8,192 batched tokens. It adds a LiLiCorr drafter with seven speculative tokens and dflash. It uses qwen3_coder tool parsing and nemotron_v3 reasoning parsing. The deployment table instead describes five multi-token prediction draft tokens for Spark. Developers should preserve one consistent recipe rather than combine these settings. Sources: NVIDIA

Agent scores and throughput answer different questions

NVIDIA reports Super VL NVFP4 scores of 71.07 on SWE-bench Verified and 62.78 on Terminal-Bench 2.1. Bolt’s card reports 50.98 on Terminal-Bench 2.1 and 87.82 on PinchBench. These agent benchmarks characterize performance under their evaluation conditions, rather than establish local task costs. NVIDIA also claims up to fivefold throughput against models in Super VL’s class. Throughput alone cannot show how quickly an agent completes acceptable coding work. Sources: NVIDIA

Quantization also complicates a single-score ranking. NVIDIA says its Super VL BF16 and NVFP4 comparisons hold harnesses and other settings constant. WebVoyager falls from 77.72 to 66.31, while Terminal-Bench 2.1 rises from 60.67 to 62.78. Bolt’s card lists a Qwen comparator at 88.07 on PinchBench. That exceeds Bolt’s 87.82. NVIDIA’s announcement nevertheless calls Bolt the leading open model in its class. That class distinction is material to the claim. Sources: NVIDIA

Meta demonstrates Muse Code changing and building the Ice Cubes iOS application and operating DaVinci Resolve while reviewing footage. A useful deployment comparison would turn such workflows into repeated tasks with fixed time, token and tool-call budgets. Bug fixes, refactors and test repairs should face independent acceptance tests and regression checks. Completion counts should distinguish autonomous success from retries and human assistance, while recording invalid tool calls and parser failures. Sources: Meta AI Research

Commercial terms can decide the shortlist

Muse’s license permits use, modification and distribution subject to conditions, including retention of the license, copyright notice and usage policy. It requires prominent Built with Muse Spark Open attribution in a permitted location and defines derivatives to include models improved using outputs. Its revenue provision applies to model-as-a-service businesses above a $20 million threshold. This counts aggregate gross revenue of the licensee and affiliates during any consecutive 12 months. Those businesses must notify Meta and obtain a separate agreement before offering Muse or derivatives in that business. Sources: Meta

Internal use, research, non-model-service commercial activities and qualifying product features are expressly exempt from Muse’s revenue provision. Both selected Nemotron artifacts identify OpenMDW-1.1 as their license. It grants free rights in provided model materials and requires retaining the agreement and applicable origin notices upon distribution. It imposes no obligations on generated outputs. It has no comparable revenue threshold or product attribution requirement. Litigation-triggered termination, warranty exclusions and responsibility for third-party rights still apply. Sources: NVIDIA, Meta, OpenMDW

Compare the cost of accepted work

OpenRouter’s October 16, 2026 listing priced Muse at $1.25 per million input tokens and $4.25 per million output tokens. Cache reads cost $0.15 per million tokens. At those rates, 100,000 uncached input tokens plus 20,000 output tokens would cost $0.21, excluding tools and other charges. That illustrative token bill does not price successful completion. The table below provides a dated reference for a local-versus-hosted comparison. Sources: OpenRouter

For local inference cost, total allocated hardware, electricity, labor and tool charges across all attempts. Divide that total by independently verified successful tasks. Separate marginal operating cost from amortized ownership cost. If nothing succeeds, report no successful completions.

Preserve checkpoint hashes, quantization details, drafter revisions, runtime and container versions, and harness and task commits. Also record parsers, context settings, hardware, operating system and drivers. Measure setup time and cold loading separately. Then increase context while recording host and device memory, swap, offload and cache growth. Track first-token latency, prefill, generation and full-task latency alongside metered electricity and purchase price. Start with hardware the team can access and define acceptance tests before running the agent.