On October 15, 2026 01, Meta let anyone download Muse Spark Open, a coding model with 816 billion 25 parameters. Meta says only about 42 billion 04 of them work on any single token. It reads up to a million tokens at once. Meta says it runs on two Apple M5 Ultra Mac Studios.
Three days earlier, Meta shipped Muse Wonder. It kept that one. According to Meta's launch post, the higher-capability version, Muse Wonder Full, goes only to subscribers on the Power Usage and Maximum Usage Muse Code plans.
The weights you can own are the second-best weights Meta has. For anyone choosing a foundation for a coding agent, the choice is plain: good enough and yours, or best and rented. We will admit the gap up front. Nobody appears to have published head-to-head task completion rates for Spark Open against Wonder, so the deciding math has to come from your own repo.
Meta opens the tier below its best
Opening the second-best model is one part of a three-tier machine, and each tier has one job. One report covers the open release, and Meta's own pages describe the split. Sorted by job, the October releases look like this.
What it takes to run Spark Open yourself, and what one bad patch costs
The open tier is Muse Spark Open, and its job is adoption. Meta's documentation says builders can serve it with SGLang or vLLM. TensorRT-LLM and llama.cpp also work. Both can load BF16 or NVFP4 checkpoints. A checkpoint is a saved copy of the trained weights. NVFP4 is a compact 4-bit number format that shrinks them.
The hosted tier is Muse Wonder. Builders reach it through the Meta Model API, and it also runs in Muse Code. Its job is usage revenue. The premium tier is Muse Wonder Full, which Meta limits to Muse Code subscribers. That one exists for recurring revenue and to give people a reason to pick Meta's own coding product.
What holds the tiers together is one line in Meta's developer documentation. It says the same request patterns work across Muse Spark and Muse Wonder when you change the model identifier. So a team that builds on the Spark family has already written code shaped for Wonder. Moving up a tier costs one string.
The architecture keeps the giveaway cheap to run. Spark Open is a mixture-of-experts model: it holds many specialist sub-networks, and a router picks a few for each token. Divide 42 by 816 and about 5% 26 of the weights do work on any token. You get a huge model's knowledge and pay a much smaller model's compute bill per token.
Cost per accepted patch decides it
A strategist judges a business on cash. A coding agent deserves the same test. The useful unit is the cost per accepted patch. The token bill and the retries come first. Human review time goes on top. So do hardware and security work, and the total gets divided by the patches you would actually merge.
Owning a model sets a memory floor you can check on a napkin. At BF16, each parameter takes 2 bytes, so 816 billion 27 parameters need about 1,632 GB. At NVFP4, each takes about 4.5 12 bits including scaling overhead, so roughly 459 GB before other overhead. A million-token context adds working memory on top of that.
Two Mac Studios is Meta's floor for running the model. Serving it to a team of engineers at the same time is harder. TechCrunch reports that a Disrupt 2026 panel will explore how open models give flexibility while builders take on the deployment and infrastructure decisions. In cash terms, owned hardware is a fixed cost, and it pays back only when it stays busy. A rented API is a variable cost that drops toward zero in a slow month.
Meta positions Wonder for agents that hand work to sub-agents and push through long multi-file tasks, which is exactly where coding agents break. If Spark Open fails more often on long tasks, review hours will eat the token savings fast. Engineer time is the expensive input here.
Nobody knows yet whether that gap is wide or narrow on real repository work. Artificial Analysis's Wonder post charts Muse Wonder and Wonder Full on the Artificial Analysis Intelligence Index. Muse Spark 1.3 appears on the same chart. But an index score says little about whether a 40-file refactor passes your test suite. Until someone runs both models on the same harness with the same pass rules, every claim about the gap is marketing.
Privacy is the strongest case for owning. Source code and credentials can stay inside your walls. The catch is that everything around the model has to stay local too. Hosted embeddings and cloud telemetry can leak repository data even when the checkpoint sits on your desk.
Safety becomes your problem as well. Technology Review reported that open models can be retrained to behave differently, and that many experts favor staged release. A coding agent holds shell access and secrets. Once you host it, you are the one logging it and patching it.
Meta has run the open playbook before. Wired reports that Meta's Llama models were the first frontier open-weight offerings from a big lab. They were built to run on your own hardware. What's different now is the tier above. Llama was the best Meta had. Spark Open is the floor beneath a product Meta sells.
Here is what the evidence supports deciding now. Teams whose code legally cannot leave the building should start provisioning for Spark Open today. Small teams with uneven usage should rent Wonder and skip the hardware. We think most teams in between should rent first and treat Spark Open as an exit they rehearse every quarter.
Three things would change that call. One is published task completion results for both models on a shared harness. Another is a cheap hosted price for Spark Open from an outside provider. The last is clear license terms from Meta on commercial use and fine-tuning.
Cheap to run, unproven, and climbing fast
Only a sliver of the model works per token.
Spark Open is a mixture-of-experts model, so about 42 of its 816 billion parameters fire on any token. You get a huge model's knowledge at a much smaller model's compute bill per token.
Index scores say little about a 40-file refactor.
Artificial Analysis charts Wonder and Wonder Full on its Intelligence Index, but no one has published task completion rates for Spark Open against Wonder. If Spark Open fails more on long multi-file tasks, review hours will eat the token savings.
Meta's flagship model became a free download in 37 days 28.
Muse launched on Muse Spark on September 8, 2026 10, and Spark Open shipped October 15 once Wonder sat above it. If the cadence holds, Wonder-class weights could open around early 2028.
Wonder-class weights go local by 2028
The dated facts suggest a cadence. Meta's newsroom says the company began overhauling its AI work fourteen months before October 12, 2026 09, which puts the start around August 2025. Muse, Meta's consumer agent, launched in the United States on September 8, 2026 10, running on Muse Spark. Spark Open shipped October 15, 2026 01.
Count the days. September 8 to October 15 is 37 days 28. In just over five weeks, the model powering Meta's flagship consumer product became a free download, and what triggered it was a better model shipping above it. Meta's research page already lists a newer Muse Spark 1.3 as well.
Project that forward with the one clock we have. Fourteen months took Meta from rebuild to frontier. If the next jump takes as long, a successor to Wonder lands around the end of 2027, and if the same pattern holds, Wonder-class weights could open in early 2028. That projection rests on a single release cycle, so treat it as a guess with a date attached.
Technology Review reported that open models trailed the leading closed models by only months. Put the two facts together and the builder's position gets clearer each year. Whatever counts as "best and rented" today is likely to be "good enough and local" within one or two cycles. The open tier keeps climbing, on Meta's schedule.
That points to an asymmetric bet. Writing your agent against Meta's request patterns and your own test harness costs little and keeps both doors open. The harness is the asset that compounds: twenty real tasks with real pass rules can grade every model Meta ships through 2031. The risk worth avoiding is a pile of prompts and tooling that only works inside one vendor's subscription product.
Price one repo task both ways
You can produce the one number that settles this for your team in a weekend. Here is the build.
Step one: pull 20 closed issues from your own repository that have a merged fix and tests covering it. Record the commit just before each fix. This set is your eval set, meaning a fixed group of tasks you grade every model against.
Step two: write a harness. A harness is a script that runs each task the same way every time. It checks out the pre-fix commit, hands the agent the issue text, lets it edit files and then runs the tests. Log pass or fail for each task.
Step three: run Muse Wonder through the Meta Model API. For every task, log input tokens, output tokens, wall-clock time and pass or fail. When a patch passes but you would still edit it before merging, log your cleanup minutes.
Step four: point the same harness at Spark Open. Meta does not serve Spark Open through its own API, so use an outside provider such as OpenRouter. If you have the hardware, run it locally with vLLM or llama.cpp. Keep the prompts, tools and pass rules identical across both runs.
Step five: compute cost per accepted patch for each model. Add the token bill, plus cleanup minutes times your hourly rate, plus a share of any hardware. Divide by patches you would merge. At $100 30 an hour, thirty minutes of cleanup costs $50, and that figure sits right next to your logged token bill.
Expect the first run to break. Tests will flake, environments will drift and the agent will loop on a missing dependency. Fix the harness before you blame either model. Then rerun the full set whenever Meta ships a new Spark or Wonder version, because that cadence is now measured in weeks.
Price one repo task on both models
- Build a 20-task eval set. Pull 20 closed issues from your repo that have a merged fix and covering tests, and record the commit just before each fix.
- Run one harness against both models. Send Muse Wonder through the Meta Model API and Spark Open through OpenRouter or a local vLLM or llama.cpp setup, keeping prompts, tools and pass rules identical. Log tokens, wall-clock time, pass or fail and your cleanup minutes.
- Compute cost per accepted patch. Add the token bill, cleanup minutes times your hourly rate and a share of any hardware, then divide by patches you would merge. At $100 30 an hour, thirty minutes of cleanup is $50.
The asset that compounds is your harness.
Meta built Spark Open as the floor beneath a product it sells, and moving up a tier costs one model string. Teams whose code cannot leave the building should provision for Spark Open now, and small teams with uneven usage should rent Wonder. We think most teams in between should rent first and rehearse the exit every quarter. Twenty real tasks with real pass rules can grade every model Meta ships, so build those first.