.svg)
.svg)
.svg)
.svg)
.svg)
.svg)
.svg)
Meta released Muse Spark 1.3 on September 2, 2026. It's the fourth Muse Spark release in five months from Meta Superintelligence Labs and, like the three before it, a closed-weight model you reach through Meta's own API or through Muse Code, the terminal coding agent Meta shipped in August.
OpenRouter listed it the same day under meta/muse-spark-1.3. Below is what changed since 1.2, where the benchmark numbers in the coverage came from, what the cheap "contributor" pricing means for your data and which parts of the Muse stack you can put on a server of your own. For the 1.3 model itself the self-hosting answer is no. There is a smaller Muse model from August where the answer is yes, and it gets its own section. Let's go :)
Muse Spark, Muse Glimmer and Muse Code
Muse is the brand Meta put on everything that came out of Meta Superintelligence Labs this year, so the names overlap and people mix them up. Muse Spark is the frontier model line: 1.0 arrived in April, 1.1 on July 9, 1.2 on August 5 and now 1.3.
It runs Meta AI inside Meta's own apps and it's the model behind the personal agents Meta keeps describing, the ones that "can work 24/7 on your behalf" in Meta's phrasing to Axios. Muse Code is the coding agent, a native binary for macOS and Linux that plans, edits and validates changes across a repository and fans work out to subagents in separate git worktrees when a job is big enough. Muse Glimmer is a 30 billion parameter open-weight model under Apache 2.0, distilled from Muse Spark, released on August 10 and built to run on one GPU.
There was a small Bounce Router post on Hacker News this week, a TUI that sits over Claude Code, Codex and Muse Code and switches to the next one when a subscription's usage runs out. Five weeks after launch, Muse Code is already being treated as a third option next to the two agents most people had installed.
What changed in Muse Spark 1.3
Meta's launch post says little about architecture... most of it describes behaviour. On Meta's internal coding evaluations, 1.3 needed about 20% fewer tool calls and about 25% fewer tokens than 1.2 to finish the same work. Those are Meta's numbers from Meta's harness, so I'd treat them as the vendor's claim until someone outside Meta reproduces them.
The rest of the list is what you'd expect from a model aimed at long agent runs. It holds onto instructions over longer sessions, it handles several tasks in one thread without losing track of which is which, it asks for clarification when inputs conflict and it's more careful before doing something irreversible, like a destructive git command or a deploy. Meta also says prompt injection resistance improved. For an agent with shell access (!!!) that's the line I'd read twice.
Context length is 1,048,576 tokens on the API, inputs are text and images + tool calling and structured outputs are supported. The knowledge cutoff isn't stated in the launch post and I couldn't find it on the model page either.
Muse Spark 1.3 xhigh vs max reasoning tiers
Muse Spark 1.3 ships in two reasoning configurations, and a lot of the launch coverage mixed them up. The one you get today through the API and Muse Code is called xhigh. The one that produced most of the headline scores is called max, and max is still in a limited preview while Meta finishes safety testing on it. VentureBeat pulled the two apart: on OSWorld 2.0, max scores 66.9 and xhigh 57.2; on JobBench 64.9 against 61.2; on GDPval-AA v2 an Elo of 1,754 against 1,709.
On Terminal-Bench 2.1 the gap flips, 88.8 for max and 89.2 for xhigh, so for terminal-heavy agent work the tier you can buy is already the better one. Artificial Analysis puts xhigh at 61 on its Intelligence Index and max at 62 - so just 1 point. The gap depends a lot on which benchmark you're looking at.
Mark Zuckerberg's line for the launch was that 1.3 rolls out "with frontier performance almost too cheap to meter". Alexandr Wang, Meta's chief AI officer, was more measured with Axios: "It's very competitive with frontier models." Both quotes are about xhigh, and the larger numbers on the scorecard belong to max.
Muse Spark 1.3 pricing and the contributor tier
Pricing didn't move from 1.2 - On OpenRouter, standard Muse Spark 1.3 is listed at $1.25 per million input tokens, $4.25 per million output tokens and $0.15 per million for cached input. The contributor endpoint, meta/muse-spark-1.3-contributor, is $0.10 in and $0.20 out, with cached input at $0.002. That's a 92% cut on input and 95% on output for the same model and the same 1M context.
The contributor model's description on OpenRouter includes this sentence: "prompts and outputs may be used to improve Meta's products." Meta gets to train on what you send. Wang told Axios that a "meaningful double digit" percentage of Muse Code users have taken that deal. For a hobby project or an open-source repo I'd take it too. For client code, anything with credentials in the context, customer data or a codebase under NDA, you'll be on the standard tier, and the price difference is what that costs. OpenRouter's description says the contributor endpoint is for experimentation, learning and early-stage work, which is fair.
Can you run Muse Spark 1.3 on a VPS?
No. The weights aren't published and there's no inference package, so there's nothing to install, on a CPU-only VPS or on a 96 GB GPU server. A VPS can host the software around the model.
Muse Code installs on Linux with one line, curl -fsSL https://dev.meta.ai/install.sh | bash, and it's a native binary that talks to Meta's API, so it runs on a server the way it runs on a laptop. Long refactors that outlive an SSH session and a shared box a small team points at one repository: those are the reasons people already run Claude Code on a VPS and the same reasons apply here. If you'd rather not sign into a Meta account from a server, an OpenRouter key gets you the model through any tool that speaks the OpenAI API, following the same pattern as routing Claude Code through OpenRouter with meta/muse-spark-1.3 as the model name. A self-hosted agent like OpenClaw can use it as its main model the same way.
Every file the agent reads still leaves the server and lands on Meta's inference, metered per token, under the terms of the tier you picked. Hosting the client yourself gets you persistence and an isolated workspace. It changes nothing about privacy.
Muse Glimmer on a GPU VPS
However... Muse Glimmer is the Muse model you can host.
Basically it's a dense 29.6 billion parameter model with a 1.8 billion parameter vision encoder, 52 layers, 128K context, Apache 2.0 with no user-count clause, and it was on Hugging Face the day Meta announced it, with Ollama, vLLM and LM Studio support the same week. Meta's stated targets were a 24 GB card for the smaller 4-bit quantisation (about 1% accuracy loss on Meta's evals), 32 GB for the better one (0.2%) and 55 GB and up for full BF16.
On Ollama it's one command, ollama run muse-glimmer, and the default tag is an 18 GB download. That download size is what decides the GPU. An 18 GB model plus KV cache for any useful context doesn't fit a 16 GB Tesla T4 or a 16 GB RTX PRO 2000 without spilling layers to the CPU, and once layers spill, an agent that makes twenty calls per task becomes painful to wait for. The 24 GB RTX PRO 4000 Blackwell on LumaDock's GPU VPS lineup is the first plan where the whole model sits in VRAM with room for context, and the 96 GB RTX PRO 6000 runs it in BF16 with the full 128K window if you'd rather skip quantisation.
By the way, all of those plans are full passthrough, so nvidia-smi inside the VM shows the same card and the same memory as the spec sheet.
Glimmer doesn't replace Spark 1.3. Meta's own scores put it at 51.2 on SWE-Bench Pro, 75.5 on MCP Atlas and 74.6 on DeepSearch QA, a strong 30B and well short of the frontier tier. What you get in exchange is a Muse-trained agent model whose prompts stay on your server, which is what you want for the code you wouldn't send to the contributor tier.
Muse Spark 1.2 open weights status
On August 10, alongside Glimmer, Zuckerberg said Meta would "also release the weights for Muse Spark 1.2, our latest foundation model", in the coming weeks. When I checked on September 10 they hadn't appeared, and the 1.3 launch post only refers to a "Muse Spark open weights release" with no version, size or date attached.
I have to give you some history here: in May, Wang had said Muse Spark "triggered checks for biological, chemical, cyber and loss-of-control risks" and was "not suitable for open sourcing", and Meta hasn't explained what changed between May and August. Until the files are on Hugging Face, the only Muse you can put on your own GPU is Glimmer.
I would plan for that.. Build the agent around an OpenAI-compatible endpoint, run Glimmer locally for the private work and route the hard tasks to Spark 1.3 through the API, and when the 1.2 weights land you change a model name. I don't know the parameter count of Spark 1.2 and Meta hasn't said, so which GPU that day calls for is an open question.
