SGLang Gives Meta Muse Spark Open Day-0 Boost [2026]

Meta’s newest open-weight model has barely been out for a day, and the inference ecosystem has already caught up to it. On October 15, 2026, the open-source inference engine SGLang and the post-training framework Miles announced day-0 support for Muse Spark Open, Meta’s freshly released any-to-text multimodal model. The timing matters: instead of the usual multi-week gap between a model drop and production-ready tooling, developers can run, fine-tune, and deploy Muse Spark Open the same day it shipped.

The model itself is a big one. Muse Spark Open carries 816 billion total parameters but activates only 42 billion parameters per token, a mixture-of-experts design meant to keep inference costs manageable despite the model’s scale. Intel called it “the most capable U.S. open-weights model released to date” in its own day-0 support announcement, a claim that puts Muse Spark Open squarely in competition with the leading Chinese open-weight releases of the past year. For readers tracking the pace of on-device and open-weight AI model launches, this is one of the more consequential ones of the fourth quarter.

Google · Preferred Sources

Don't miss new tech stories on Google

Add TrendinTech once in the Google app and our stories appear in your news suggestions.

Add Now

What Muse Spark Open Actually Is

Muse Spark Open is an open-weight model built by Meta that accepts text, image, audio, and video as input and produces text as output, according to Meta’s own developer documentation. That any-to-text design puts it in a different category from pure text-in, text-out language models: it is built to read a screenshot, listen to an audio clip, or watch a short video clip and reason about all of it in the same context window as a written prompt.

The headline number is the parameter count. At 816 billion total parameters with 42 billion active per token, Muse Spark Open sits in a similar design category to other large mixture-of-experts systems released over the past two years, where only a fraction of the total network fires for any given token. This is what makes a model of this size practical to run outside of a hyperscale data center at all. The official checkpoint, published for the NVFP4 quantization format, is listed under the name meta-models/Muse-Spark-Open-816B-A42B-NVFP4, which signals Meta’s intent for the model to be distributed and run in that lower-precision format by default rather than full 16-bit precision.

Hugging Face, which hosts the model weights, described the scope of the SGLang integration plainly in its own launch post: “The Muse Spark Open integration for SGLang supports text, image, audio, and video inputs,” confirming that the multimodal capability is not limited to Meta’s own serving stack but is baked into the third-party inference tooling as well.

Why SGLang and Miles Shipping on Day Zero Is a Big Deal

In most open-weight model launches, there is a lag between the model landing on Hugging Face and the broader inference ecosystem catching up with optimized kernels, quantization recipes, and serving configurations. That lag can run anywhere from a few days to several weeks, during which early adopters are stuck running unoptimized code or waiting for community patches.

SGLang, the open-source serving framework originally built out of UC Berkeley’s Sky Computing Lab research lineage and now maintained under the LMSYS banner, skipped that gap entirely. In its own announcement, the SGLang project described the model’s intended purpose directly: “Muse Spark Open is built to plan, act, and use computer across agentic tasks for life and work.” That framing, plan, act, and use a computer, points to agentic workloads rather than simple chat completions as the primary target use case.

SGLang’s announcement also laid out exactly what developers get on day one: “You can now build a powerful personal AI agent with an open stack, and SGLang to make it: Ultra-responsive with NVFP4 MegaMoE and sequence parallelism. Faster for agentic and coding tasks with our DFlash2 draft model. Customizable for specialized use cases via fine-tuning with Miles.” That last point is where Miles comes in. Miles is the post-training and fine-tuning framework that pairs with SGLang to let developers adapt the base model to narrower tasks without waiting for separate tooling to mature.

Meta’s own developer blog confirmed that this pairing was intentional and live at launch, not a later add-on: “Post Training: Fine-tuning with TorchTitan, SGLang + Miles, NVIDIA NeMo AutoModel, Hugging Face TRL and managed post-training APIs from Fireworks AI and Together AI are all live at release.” Five separate fine-tuning paths, available the same day the base weights dropped, is an unusually wide net for a model of this size.

One Early Benchmark: Miles and the Geometry3K Jump

The clearest evidence that Miles’ fine-tuning path works in practice comes from a benchmark result tied to Geometry3K, a geometry-reasoning dataset used to test multimodal models on diagram-based math problems. Using Miles, accuracy on Geometry3K moved from 88.2 percent to 94.9 percent after just 30 training updates. That is a meaningful jump for such a small number of optimization steps, and it suggests the post-training pipeline Meta and the SGLang team built around Muse Spark Open is tuned for fast iteration rather than requiring the long training runs typically associated with models in the hundreds-of-billions-of-parameters range.

For engineering teams evaluating whether to build on Muse Spark Open versus a smaller, cheaper model, that kind of rapid fine-tuning turnaround is often the deciding factor. A model that needs days of GPU time to specialize for a narrow task is a different cost proposition than one that can be nudged with 30 updates.

Hardware Reach: NVIDIA, AMD, Intel, and Google TPUs All at Once

What separates this launch from a typical single-vendor showcase is the breadth of hardware SGLang says it supports from day zero. The framework’s support spans NVIDIA, AMD, and Intel hardware, plus Google’s Tensor Processing Units through a dedicated backend called sglang-torchtpu. That is four major silicon vendors with working day-0 paths for the same 816-billion-parameter model, something that was rare even for much smaller open releases a year or two ago.

This breadth lines up with a broader shift the industry has been pushing toward standardized, vendor-agnostic AI infrastructure. It echoes the direction NVIDIA took when it open-sourced its Run:ai GPU scheduling tools earlier this year, and the push toward broader hardware access that NVIDIA has also pursued through its open-source AI tooling initiatives in Europe. A model that only runs well on one vendor’s chips limits its own addressable market; a model that runs across four vendors’ chips on launch day removes that friction immediately.

The NVIDIA Configurations: GB300, RTX Pro 6000, and H200

Meta’s reference configuration for the NVFP4 quantized checkpoint runs across four NVIDIA GB300 GPUs. That is the setup Meta itself validated and published alongside the model weights, making it the baseline configuration most production deployments are likely to copy first.

SGLang’s own prevalidated configurations go further, covering hardware that is more accessible outside of hyperscale data centers. One prevalidated SGLang setup runs on eight NVIDIA RTX Pro 6000 cards, each with 98 GB of memory, supporting a context length of 131,072 tokens. A second configuration swaps in eight NVIDIA H200 GPUs, each with 141 GB of memory, and extends the usable context window all the way to 1,048,576 tokens, commonly referred to as a one-million-token context.

That context-length tradeoff is worth sitting with. The RTX Pro 6000 configuration, while using less total GPU memory per card than the H200 setup, caps out at roughly one-eighth of the context length available on the H200 build. For workloads that need to process long documents, lengthy video transcripts, or multi-hour audio recordings in a single pass, the H200 configuration is the one that actually unlocks that capability. For more typical agentic or coding tasks that do not need a million tokens of context, the RTX Pro 6000 setup is the more cost-efficient path.

Configuration	GPUs Required	Memory per GPU	Precision	Max Context Length
Meta reference build	4x NVIDIA GB300	Not specified	NVFP4	Not specified
SGLang prevalidated config A	8x NVIDIA RTX Pro 6000	98 GB	NVFP4	131,072 tokens
SGLang prevalidated config B	8x NVIDIA H200	141 GB	NVFP4	1,048,576 tokens
Apple Silicon via oMLX / mlx-lm	Apple Silicon (quantized)	Not specified	3-bit or 4-bit oQ	Not specified
Intel day-0 path	Intel Xeon CPUs, Intel GPU D-series	Not specified	Not specified	Not specified

Running It on a Mac: Apple’s oMLX and mlx-lm Support

Perhaps the most unexpected branch of day-0 support comes from Apple’s side of the ecosystem. Both oMLX and mlx-lm, the tooling built around Apple’s MLX machine learning framework, support Muse Spark Open at launch. oMLX-optimized quantizations, referred to as oQ, are available in both 3-bit and 4-bit formats, which is what makes running a model with 816 billion total parameters on consumer-adjacent Apple Silicon hardware feasible at all.

Meta’s own account, published through its AI organization, confirmed the Apple Silicon angle directly: “To help make advanced models more accessible, Muse Spark Open is optimized to run on two Apple M5 Ultra Mac Studios with the latest MLX framework.” Pairing two Mac Studios to host a single model is an unusual deployment pattern for a model this size, but it tracks with a broader pattern of vendors finding creative ways to put large open-weight models within reach of smaller teams and individual developers, not unlike the idle-GPU pooling approach Apple has been exploring to put unused Mac GPU capacity to work for AI workloads more broadly.

Intel’s Day-0 Commitment Across Xeon and Data Center GPUs

Intel’s involvement is the other half of the hardware-diversity story. Intel reported day-0 support across Intel Xeon processors and its data center GPU platforms, including the Intel GPU D-series. Intel also said the model had been tested and was running on its upcoming data center GPU platform, the Intel GPU D-series, ahead of that hardware’s own general release, which is a notable vote of confidence from Intel in both the model and its own forthcoming silicon.

Intel framed the significance of the release in sweeping terms in its own announcement: “Muse Spark Open, an 816 billion-parameter any-to-text model, represents a major milestone in multimodal AI and the most capable U.S. open-weights model released to date.” That phrase, the most capable U.S. open-weights model released to date, is doing specific competitive work. It positions Muse Spark Open against the wave of large open-weight releases that have come out of Chinese AI labs over the past two years, several of which have topped leaderboards that U.S. labs had previously dominated.

Day-0 Support Across the Inference and Fine-Tuning Stack

Pulling together everything confirmed about the launch stack, the breadth of same-day tooling support is unusual even by the standards of 2026’s increasingly fast open-model release cadence. The table below summarizes what shipped on day zero across the inference and fine-tuning ecosystem.

Framework or Vendor	Hardware Target	Role	Status at Launch
SGLang	NVIDIA, AMD, Intel, Google TPU (via sglang-torchtpu)	Day-0 inference serving	Live
Miles	Cross-platform, paired with SGLang	Post-training and fine-tuning; Geometry3K accuracy moved from 88.2% to 94.9% in 30 updates	Live
oMLX / mlx-lm	Apple Silicon (two Mac Studios in Meta’s reference build)	Local inference with 3-bit and 4-bit oQ quantization	Live
Intel	Intel Xeon CPUs, Intel GPU D-series	Day-0 inference support, including on pre-release GPU hardware	Live

Competitive Field: Where Muse Spark Open Lands

The competitive framing around Muse Spark Open is being written by the people covering it, not just Meta itself. The New Stack described the release as Meta’s answer to China’s best open-weight models, a framing that reflects how much the open-weight leaderboard has shifted over the past two years toward labs outside the United States. VentureBeat similarly billed it as the new top U.S. open-weights model, a distinction that implicitly concedes Chinese labs had been setting the pace on pure open-weight capability.

That framing matters for enterprise buyers deciding which open-weight family to standardize on. A model’s raw capability score is only part of the decision; the surrounding tooling ecosystem, licensing terms, and hardware flexibility often matter just as much. Muse Spark Open’s day-0 coverage across SGLang, Miles, Intel, and Apple’s MLX stack gives it an operational head start that pure benchmark numbers do not capture. It also stands in contrast to the concerns some AI labs have raised about open-weight models carrying more risk than closed, API-gated ones, a debate that is unlikely to go away as more large open-weight releases land with this much day-0 infrastructure behind them.

On pure hardware terms, Muse Spark Open’s day-0 reach across NVIDIA, AMD, Intel, and Google TPU silicon also puts pressure on the usual pattern where new large models launch tuned almost exclusively for NVIDIA GPUs first, with AMD and Intel support arriving months later once community contributors or internal teams at those companies catch up. Readers who have followed the ongoing contest between NVIDIA’s RTX GPU lineup for AI workloads will recognize how rare simultaneous multi-vendor support still is for a model at this parameter count.

Historical Context: Meta’s Open-Weight Strategy

Meta has spent several years positioning itself as the leading U.S. advocate for open-weight AI models, a strategy that has drawn both praise for democratizing access to frontier-scale AI and criticism over the safety implications of releasing weights that can be fine-tuned or modified by anyone who downloads them. Muse Spark Open continues that strategy at a larger scale than most of Meta’s earlier releases, while also broadening the modalities the model can handle beyond text alone.

The pairing with SGLang and Miles on day zero also reflects how much the open-source inference ecosystem has matured since Meta’s earlier large model releases. Where past launches often required the community to build custom serving code from scratch, the infrastructure around Muse Spark Open was ready before the weights were even public, a sign that coordination between model labs and inference framework maintainers has become a standard part of the release process for major open-weight launches rather than an afterthought.

Market Impact: What This Means for Enterprise AI Buyers

For enterprise teams evaluating open-weight models, day-0 multi-vendor hardware support changes the calculus around vendor lock-in. A company that has already invested in AMD or Intel accelerators, rather than NVIDIA GPUs, does not need to wait for a separate optimization pass before testing Muse Spark Open on its own infrastructure. That lowers the switching cost for teams that might otherwise have defaulted to a smaller, NVIDIA-only model simply because the tooling was more mature.

The breadth of fine-tuning options, five separate paths live at release according to Meta, also lowers the barrier for companies that want a specialized version of the model rather than the general-purpose base weights. Combined with the Miles-driven Geometry3K result, where 30 training updates moved accuracy nearly seven percentage points, the signal to enterprise buyers is that customization does not require the massive compute budgets that model fine-tuning has historically demanded at this parameter scale.

There is also a cost angle buried in the hardware table above. Teams that do not need a million-token context window can deploy on the RTX Pro 6000 configuration rather than the H200 build, a choice that affects both upfront hardware costs and ongoing power draw. That kind of configuration flexibility, baked into SGLang’s prevalidated setups rather than left for each team to figure out independently, is itself a form of market impact: it compresses the time between a model’s release and its first production deployment.

What Comes Next: Five Predictions

Expect rival open-weight labs, both in the United States and China, to compress their own day-0 tooling timelines in response, since Muse Spark Open has reset expectations for what a launch-day ecosystem should include.
Watch for additional quantization formats beyond NVFP4 and the existing 3-bit and 4-bit oMLX options to appear within weeks, as the community optimizes further for memory-constrained hardware.
Expect more detailed throughput and latency benchmarks for the GB300, RTX Pro 6000, and H200 configurations to surface from independent testers over the coming weeks, since Meta’s own release materials did not publish comparative performance numbers across those three setups.
Expect Intel to use its day-0 support and testing on the unreleased Intel GPU D-series as a marketing point ahead of that hardware’s formal launch, given how directly Intel tied its announcement to Muse Spark Open.
Expect agentic coding tools and customer-service automation vendors to be among the first commercial adopters, given SGLang’s explicit framing of the model as built to plan, act, and use a computer across agentic tasks.

How Teams Can Get Started Today

For teams that want to test Muse Spark Open immediately, the fastest path depends on available hardware. Teams with access to NVIDIA H200 or RTX Pro 6000 clusters can follow SGLang’s prevalidated configurations directly, choosing the H200 build if long-context workloads such as document or video analysis are the priority, or the RTX Pro 6000 build for shorter-context agentic and coding workloads at a lower hardware footprint. Teams without large GPU clusters can test the model locally through oMLX or mlx-lm on Apple Silicon hardware, using the 3-bit or 4-bit oQ quantizations to fit the model into a more modest memory budget, following the same two-Mac-Studio pattern Meta itself validated.

Teams on Intel hardware, meanwhile, can rely on Intel’s own day-0 confirmation that the model runs on Xeon CPUs and the Intel GPU D-series, giving them a validated path without needing to wait for community-contributed optimization work. In every case, the underlying weights are the same 816-billion-parameter, 42-billion-active-parameter checkpoint, with the choice of framework determined mainly by what hardware a given team already has on hand.

Frequently Asked Questions

What is Muse Spark Open?
Muse Spark Open is an open-weight, any-to-text multimodal model released by Meta on October 15, 2026. It accepts text, image, audio, and video inputs and produces text output.

How large is Muse Spark Open?
It has 816 billion total parameters, with a mixture-of-experts design that activates approximately 42 billion parameters per token during inference.

What is SGLang?
SGLang is an open-source inference serving framework that announced day-0 support for Muse Spark Open, covering deployment across NVIDIA, AMD, Intel, and Google TPU hardware.

What is Miles, and how is it different from SGLang?
Miles is a post-training and fine-tuning framework that pairs with SGLang. While SGLang handles serving and inference, Miles is used to customize the base model for specialized tasks, as shown by its Geometry3K accuracy improvement from 88.2 percent to 94.9 percent in 30 training updates.

Can Muse Spark Open run on a Mac?
Yes. Meta’s own reference deployment runs the model across two Apple M5 Ultra Mac Studios using the MLX framework, with oMLX and mlx-lm providing 3-bit and 4-bit oQ quantized versions for local use.

What is the largest context window Muse Spark Open supports?
SGLang’s prevalidated configuration using eight NVIDIA H200 GPUs supports a context length of 1,048,576 tokens, commonly described as a one-million-token context window.

Does Muse Spark Open run on non-NVIDIA hardware?
Yes. Intel confirmed day-0 support across Intel Xeon processors and its data center GPU platforms, including the Intel GPU D-series, and SGLang’s broader support extends to AMD hardware and Google TPUs as well.

Why is day-0 support from SGLang and Miles significant?
Most large open-weight model releases require the broader inference ecosystem days or weeks to catch up with optimized serving and fine-tuning tooling. Muse Spark Open shipped with that tooling already live, across multiple hardware vendors and five separate fine-tuning paths, compressing the time between release and production deployment.

SGLang Gives Meta’s 816B Muse Spark Open Day-0 Boost [2026]