Muse Spark Open 816B: How Much VRAM, and Does It Beat DeepSeek V4 Flash?

On 15 October Meta released Muse Spark Open, an 816-billion-parameter mixture-of-experts model with 42B active per token: open weights, a 1M-token context, and text, image, video and audio input. It shares its name with the closed Muse Spark we covered in April, the first model from Meta Superintelligence Labs; Meta's model card does not say how the two relate, and it is the largest open model Meta has published. Together AI serves it free of charge until 18 October at 20:00 Belgrade time, so for three days anyone can put it next to whatever they run at home.

We run DeepSeek-V4-Flash, a 304B MoE, on two GB10 boxes joined by one 200 Gb cable. The question every owner of a box like that asks about a frontier release is the same one: does it make what I own look slow, or stupid? We measured both halves. The memory answer is that Muse Spark Open does not come near a pair of DGX Sparks at any precision Meta publishes. The quality answer is that on every test we own it scores the same as the 304B model that fits.

Test rig. The "pair" is DeepSeek-V4-Flash-0731 at NVFP4 with its DSpark drafter on two GB10 workstations, tensor parallel over one 200 Gb cable; its numbers are our runs of 25 September, the ones that scored the V4.1 decision. Muse Spark Open was reached from a desktop through Together AI's API on 16 October 2026, same prompts, same scoring, reasoning off unless stated. Every number below is ours unless it sits in the section marked as Meta's.

What Meta released

The licence is not Apache or MIT, and Hugging Face files it under "other". The clause that matters: a company with more than $20 million in annual revenue that offers the model as a service has to notify Meta and sign a separate agreement first. Internal use, and products that are not model-as-a-service, are exempt. For most readers that is a free licence; for a hosting company it is a phone call.

How much memory: 496 GB at NVFP4, and why not 417

The pair's GPU allocation is 208 GiB: two 128 GB boxes, minus what the system keeps, is what vLLM can take. Both Muse sizes below are summed from the Hugging Face file listings on 16 October, not estimated.

Meta's model card puts the floor plainly: 700 GB of VRAM for NVFP4, 2 TB for bf16.

Why 496 and not the 417 GiB that simple scaling gives? V4-Flash at NVFP4 is 0.511 GiB per billion parameters; 816 × 0.511 is 417. Meta's NVFP4 build quantises the experts only. Attention, the two 1.66B-parameter embedding matrices, the 1.8B vision encoder and the audio encoder stay in bf16, and since the experts are 97% of the parameters, the rest adds about 80 GiB. The V4-Flash checkpoint on the pair is quantised more completely, which is part of why 304B fits in 155 GiB. Our quantization explainer covers what NVFP4 does to a weight; the lesson here is that "NVFP4" on a model card does not say which weights.

What else would hold the 496 GiB? A 512 GB Mac Studio takes the weights with almost nothing left for context. Four GB10s give 416 GiB of allocation, still short, and tensor parallel wants a power of two, so the next step is eight boxes: 832 GiB, enough, on a fraction of a B200 node's bandwidth, and the case we made against four boxes applies to eight with more force. The other way round: 208 GiB at the pair's density holds about 400B parameters with room for context; at Meta's experts-only density, nearer 340B. A "Muse that fits two DGX Sparks" would be a model of well under half the parameters, and whether that costs quality is not something the tests below can show, because the 304B model already saturates them.

The same four tests, and a tie

These are the harnesses that scored the V4.1 decision in September, unchanged: ten coding tasks run against hidden tests, sixteen Serbian recall facts, five false-premise traps with nothing to recall, three deep explanations against a 30-fact checklist, six Serbian prose prompts, and seven broken repositories to repair with four tools and the tests re-run by the harness. Muse ran with reasoning off; the agent run used reasoning on low, because its tools do not fire with reasoning off.

* The speed is a property of the service, not of the model. Together serves Muse from Blackwell racks, and NVIDIA's smallest recipe for it is four GB200 or GB300 GPUs: hardware that costs an order of magnitude more than the two GB10 boxes on the other side of the table. The 3× says what a rack buys over a desk, not that an 816B model decodes faster than a 304B one.

A tie on everything that is pass or fail. The two scales lean to Muse by a little: it covered 10/10 on the Enigma explanation where the pair had 9/10, and both missed the same two H.265 facts, the quadtree split and the encoding cost. On length, the pair's argument essay stopped at 60% of the asked length and its business letter ran to 140%; Muse landed every prose task within 16% of the ask. Both are within what one rerun would move, since the deep set is three questions and the prose set is six.

Reading the answers rather than counting them: Muse's explanations are the more engineer-shaped. Its H.265 answer opens with block sizes and the 35 intra-prediction modes in numbered sections; the pair's opens with the standards bodies and the year. The pair's prose is warmer, and it ends a legal summary with advice where Muse ends with the summary. Both wrote clean Cyrillic and both drifted to Latin on the technical prompt, in the same place. Both refused every false premise in plain Serbian.

Every one of these tests is saturated by both models, so none of them can rank the two. What they establish is the absence of a gap: a free 816B model, run on at least four GB200-class GPUs, does not score above the 304B one in a box on a desk. Whether it would pull ahead on a harder set is open; this set is what we had, and it was built to find differences between models in this class.

What precision did we test?

Unknown, and that is a finding. Together's model page and launch post state the price, the 1M context and the four input modalities, and no precision. Together's general post on its serving says it quantises weights to "FP8, FP4 (nvfp4 or mxfp4), and hybrid precision" and calls the result effectively lossless. Two other launch partners, Baseten and Fireworks, say NVFP4 outright. So the quality numbers above are for Muse as Together serves it, most likely 4-bit or 8-bit experts on Blackwell, against the pair's NVFP4 V4-Flash: a like-for-like comparison of two quantised deployments, not of Muse at bf16. If you benchmark a hosted model, ask the host what precision you are paying for; three of the four partners here answered, in public, and the fourth did not.

What the endpoint is like

Measured from a desktop on 16 October; the asterisk is the same one as above, a rack against two boxes:

Four things that bit, and what to do about each:

1. Reasoning is on by default and not shy. Given room, a 350-word story prompt thinks for 2,400 to 2,900 tokens before writing the first word, 33 seconds; a two-sentence question thinks for 400, 5 seconds. reasoning_effort: low and medium change nothing on prose, minimal is rejected by the chat template, and Muse's own reasoning_strength dial, passed through chat_template_kwargs, makes it worse: 6,000 tokens and no story. Only reasoning_effort: "none" turns it off, and then the story arrives in 5.7 seconds and loses nothing on these tests. A "think briefly" system note cuts the thinking by a third, no more. If you keep reasoning on, add about 4,000 tokens to your max_tokens so the thinking does not eat the answer; a 600-token budget plus that allowance returned a complete 351-word story after 2,836 reasoning tokens.
2. Streaming is broken on this endpoint. With stream: true the answer arrives inside reasoning_content behind the raw template tokens to=user<|message|>, and content stays empty whatever the effort setting. Non-streaming replies are clean. Make one plain request and show it when it lands.
3. Send a User-Agent. Cloudflare in front of the API answers Python's default agent with 403 "error code: 1010". Anything with a named agent passes.
4. The free period ends without a name change. The -promo model name keeps working after 18 October at 20:00 Belgrade time and starts billing. Anything that calls it unattended should know the date.

Together's prices and Meta's numbers, for the record

The figures in this section are the vendors', not ours.

After the free window Together bills $1.25 per million input tokens, $4.25 per million output, $0.15 for cached input, for the same deployment under the same name without the promo suffix. For scale, DeepSeek's own API lists V4-Pro at $1.32 in and $3.96 out at peak hours. Muse is priced like a hosted frontier model with a 1M context and image input; nothing here tested V4-Pro, and nothing here says Muse is better than it.

Meta's model card reports JobBench 61.8, GPQA Diamond 89.0, DeepSWE 1.1 60 and Terminal-Bench 2.1 76.4, unverified by us, and recommends temperature 1.0, top-p 1.0 and reasoning_effort high or xhigh. Our harness runs were at reasoning off, the opposite end of that dial, so Meta's numbers and ours measure different settings of the same model. The point of ours was to see what a tie looks like at the speed people actually use a chat model, and with reasoning on the answers were not better, only later.

The verdict

• For an owner of two DGX Sparks, or one: there is nothing to upgrade to. Muse Spark Open does not fit at any published precision, and on every test we own it does not beat the 304B model that does.
• For three days it is free to compare, and worth an afternoon: ask the same question of both and read the answers side by side. Turn reasoning off for prose, leave it on for tools.
• "Open" here means weights you can download, a licence with a revenue clause, and a floor of 700 GB of VRAM. It is a service you rent, or a model for a lab, not a model for a desk.
• The number to keep is 2.4×: that is how far Meta's own 4-bit build overshoots the largest memory the GB10 class offers, and it is why the comparison is to a service, not to a box.

Related reading

• Meta Unveils Muse Spark: First Model From Superintelligence Labs — the closed model this one shares a name with
• Qwen3.8-27B vs Muse Glimmer 30B: Same 10/10, One Is 10× Faster — Meta's small open model on an RTX 5090
• Should You Upgrade to DeepSeek V4.1 Flash? We Tested It on Two DGX Sparks — the same harnesses, and why four boxes did not help
• DeepSeek-V4-Flash on Two GB10s: 304B Params, 1M Context, 83 Watts — how the pair runs
• Best Model for a Dual DGX Spark: DeepSeek-V4-Flash vs Qwen3.8-Flash-Next vs GLM-5.3-Flash — what else fits 256 GB
• 4-Bit Quantization Decoded — INT4 QAT, MXFP4, NVFP4
• LLM VRAM Requirements: Will It Fit an 8, 12, 16, 24 or 32 GB GPU? — the arithmetic behind the memory table