Muse Code and Muse Spark 1.2
research.meta.aiHey guys I'm just wondering. Usually when someone announces a new model, they'll show you some fancy viz/video/images: "These are what my model can produce." I'm wondering if anyone is keeping track of these? Like in a gallery form, "Use this prompt to produce this output".
By itself is useful ("I want something like this, I'll just reuse the prompt and tweak"), but it can also be used as a "draw me a pelican on a bicyle" alternative. Basically feeding those prompts over model releases.
Meta is offering a 10x discount on input ($0.10 vs. $1.25/Mtok) and 20x discount on output ($0.20 vs. $4.25/Mtok) if you opt in to let them train on your data.
I actually really like that pricing strategy. It's very transparent
I love the idea of this pricing strategy but there is no way meta is not training on your data regardless of your monthly invoice
This makes it a very interesting alternative to Deepseek for personal work where I don't care about the training - judging by the AA benchmarks it seems like overall cost per task is similar to the new Deepseek Flash but with better benchmarks (and inbuilt vision capabilities).
I think that's a fair offering tbh
Probably to compete with DeepSeek, which AFAIK also retains data (or at least OpenRouter says they do)
FYI: there are providers of deepseek that offer the same or lower pricing and zero retention policies.
Unfortunately, none with the same caching performance as DeepSeek proper.
DeepSeek is really crazy cheap, though, and they don't have a giant pool of other invasive personal data to correlate it with.
I hate to say this and this is because I fucking despise meta. But between DeepSeek and Meta, and trust they handle the training data correctly, I trust meta.
What do you mean by "correctly"?
Meta, please offer this on OpenRouter too (ZDR + Non-ZDR, official Meta Provider).
Yikes that’s compelling pricing.
I think it's limited to US or at least EU is excluded.
They chose to compare against Open AI’s mid tier model Terra instead of Sol and still lost some benchmark against it.
They left Opus in and got beat in all but one benchmark.
Nothing wrong with trying to improve, but why the marketing games?
Instead of trying to say in the post you’re “closer” to frontier, first set a clear goal to beat the Chinese labs on price or performance and demonstrate it convincingly.
Then when your ready, come back and talk frontier without playing hide the model.
Given the current throughput figures on OpenRouter (~180 tk/s), its likely a much smaller param count on the order of something like Luna. I think the better, more timely comparison (re: your point on Chinese labs) would be to DeepSeek-V4-Flash-0731.
It's definitely confusing from a presentation perspective, but they are somewhat coherent comparisons if you account for the inference heuristics involved.
(They could in theory be gaming the decode speeds with much larger than normal batch sizes given the TTFT is pretty high at around 8s)
> They chose to compare against Open AI’s mid tier model Terra instead of Sol and still lost some benchmark against it.
If you scroll very slightly farther there is a benchmark that includes Sol, showing it outperforming Terra (as expected) and Spark 1.2
While I won't take their limited benchmarks with much salt, if it actually is this close to opus, but at a third the cost, that's pretty solid. Now, Terra is pretty damn affordable too and you're right that it's suspicious that they don't put Sol in there at all.
We can throw benchmarks in the bin by now. Each one I've seen is heavily biased and skewed. It holds very little reliable data points (unfortunately)
If you look at papers on benchmarks, they're usually created to expose gaps in how models are trained. It should be no surprise that models get better on them over time, because you can't get better at what you don't measure.
Cherry picking the benchmarks you present is where the falsehoods lie.
My conclusion is the opposite. If benchmarks were meaningless, surely Meta would be able to find some benchmark that shows they are better than Sol and Fable. The fact that they can't do that tells me that benchmarks still do mean something.
Or they spent time optimizing their model to real world problems they're facing and didn't waste time trying to game a benchmark.
Or they did try to game the benchmarks and just didn’t do it well enough.
Benchmarks are one data point, not the only one, but the easiest one to compare.
Right, but the point is that you can't conclude that a model is necessarily bad because it's not hitting the same scores on benchmarks. I just don't agree with lacker's conclusion, because their logic doesn't seem to consider that. Scoring lower on a benchmark doesn't strictly mean they have a bad model, but it may be the case. Like you said it's one data point, but being the easiest, and obviously most gamed, means you should probably weigh them less heavily.
If you got the $20 in free credits from Meta for signing up when muse-spark-1.1 was release, please note that there's now small print stating "While using free credits your content may be used for product improvement" which was not present at muse-spark-1.1 launch when the credits were given out.
If you don't mind Meta retaining your data, the "Contributor" pricing is deepseek-v4-flash-level of low, roughly 1/10th normal muse-spark API pricing currently. Attractive if you're OK with them retaining and using your data.
The API costs for the version of their model that feeds things back to meta is also drastically lower.
This is a nice release and a solid improvement over Spark 1.1. It compares favorably with Grok 4.5. Not SOTA, but solid releases. I think they need to really get this more competitive with Deepseek V4 Flash / Luna pricing to move the needle.
If you are happy to share data for training, the contributor mode offers amazing price $0.10 / $0.20
Yes, that is the really compelling thing here IMO. Its a viable deepseek competitor for many people, and I missed that on the first pass.
https://pbs.twimg.com/media/HO-59jQaoAA_JZ1?format=jpg
Very interesting they have a way cheaper "contributor" version "used to improve our products", how much of that is price discrimination vs the data being that valuable?
Roughly DeepSeek V4 Flash pricing, though you can get V4 from providers that don't train on your data
Last I heard, everyone at Meta was using Claude Code.
Any insiders know how Muse Code is doing internally?
Everyone is still using claude or codex if they aren’t forced off of it. Nobody is going to use a worse tool in this culture.
it's actually interesting that they're not being forbidden to use claude/codex, is Meta paying for it or is it personal accounts?
If there were, do you believe it would be in their interest to answer this publicly?
> > Any insiders know how Muse Code is doing internally?
> If there were, do you believe it would be in their interest to answer this publicly?
If it were being adopted like gangbusters in their organization, sure!
So... the fact that nobody is volunteering the information is probably a valid signal of how things are actually going...
You should play Blood on the clocktower
I wonder why they didn't compare with GPT-5.6-sol, only Terra?
Clearly they're positioning it as a mid model.
Which is in itself a bit weird as mid models nowadays are a golden mean fallacy. Terra is much less popular than both Luna (cost-sensitive) and Sol (performance-sensitive).
Claude Sonnet is a weird exception to the mid models because Anthropic doesn't do much with Haiku and Opus is too big.
I don’t know where you get your statistics, but I love Terra and use it all of the time. It is the default fastest model in ChatGPT/Codex today. I saw today a notice saying that the model had hit capacity briefly
But why include Opus then?
Haven't you seen the kernel optimization case study at the bottom of the page? They compare against GPT-5.6 Sol and their model is worse.
Presumably because it's worse than Sol, same reason they compared it to Opus 5 not Fable.
Their bigger model is not ready - watermelon code name was still being prepared for release as of a month ago
Here's the Muse Spark 1.2 pelican: https://tools.simonwillison.net/markdown-svg-renderer#url=ht...
I think it's a bit of an improvement on the Spark 1.1 pelican: https://simonwillison.net/2026/Jul/9/muse-spark-1-1/
Lol is that a helmet? Clearly they're taking the Anthropic path and are safety pilling their models
I wonder if at this point the labs ha.. please Simon, stop this bs everytime. Its time.
This one is actually a pretty good case for the continuing value of the benchmark, because it lets me visually compare Spark (8th April), Spark 1.1 (9th July), and Spark 1.2 (5th August): https://bsky.app/profile/simonwillison.net/post/3mseqv5z4qk2...
Pricing: https://dev.meta.ai/docs/pricing-rate-limits
Interesting that they have separate API pricing for "we can train on your data" (whereas iirc most of the big players either make that distinction only between subscriptions and API usage, or train on everything). Wonder how it compares to Deepseek V4 Flash given that they're similar on pricing and data policy.
Ive been poking with the muse code binary - seems to be written in rust, looks similar to codex but either its a very hard fork (i also see dissimilar things like config format is different, no acp, etc) or is just heavily inspired by it (more likely).
Does this muse code have any muse spark 1.2 usage included? Can't understand from the docs.
Why should I leave Claude or GPT and switch to Meta's aMUSEment model?
> Muse Spark 1.2 is available today in Muse Code and in Meta Model API with expanded global access
Wasn't the previous one us only? This is probably the biggest part of the post
Anyone know if muse code is open source?
I do think some of features in their harness seem interesting (workers in separate worktrees at once), recovery from crashes seem interesting.
Is this becoming a race where we have a usual flow of a company .. AI models, Coding agents, image generation tools, and more AI models ?
Why does every AI lab feel the need to build their own coding agent…? Don’t we have more than enough already?
Outputs are a little bit more deterministic if you control the harness.
It is easy to benchmark across one harness, one system prompt and extract the most performance when you control the harness.
Telemetry, marketing
Open the weights.
the soak tub in the kitchen was nice
Fun, this is currently on the front page at the same time this https://news.ycombinator.com/item?id=49187977 is
Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1, with improvements in code generation, complex debugging, codebase understanding, and end-to-end developer workflows. In Muse Spark 1.2, we significantly scaled up training compute on coding tasks while expanding training environment diversity. The model also maintains its strength in other key areas like general agents.
Interesting, it seems like their muse code is built upon Codex CLI?
I wish they would add a ZDR endpoint on OpenRouter
Will someone at Meta for the love of God make it so none of this stuff goes through Facebook.com? You want customers but most corporate firewalls block social media. Also, a lot of devs do not want their work stuff tied up to their facebook account. For the love of all things show the IG / FB logins as optional and do email as primary.
I am not a fan of Meta but I do cheer for any competitors against OpenAI and Anthropic, the duopoly is getting tiresome.
I could sign in with my Meta account that is independent and not linked to IG or Facebook. Just click "Login with Email" on dev.meta.ai.
I honestly think they're kinda banking on piggybacking off of Facebook account integrity systems to avoid the problems that other LLM providers are facing in trying to prevent mass free trial signups for token relays and so forth.
It's not a good system obviously. Google did this as well for Gemini-CLI, but forced it to be linked to personal Google accounts (which caused a great deal of onboarding friction).
China says hi.
Not in my case, I don't see any of my employers (past or current) trusting a country like China with their data.
The decades of US brainwashing children into thinking China is the big bad guy has worked unfortunately.
+10000 to that
The only company less trustworthy than OpenAI and Anthropic is meta.
Theres no actual evidence they didn't just distill Kimi K3
At this point, it doesn't matter who is distilling from who.
Is there any actual evidence that they did?
It seems like one day, Google or Meta might produce a coding model worth discussing. That day is not today.
If anyone from Meta is reading, please can you publish the cost and latency for each of your benchmarks, like OpenAI does? Show us how the reasoning effort level affects them in 2D charts. This needs to become standard practice.
Lol worse than DeepSeek
First of all, you have login to use it. Why?
After everything that you have seen with Meta, would you really trust them with a coding agent? You don't even know if your prompts are being analyzed by them on the side or if your code base is being uploaded to them. This goes for the rest of them that have closed harnesses and closed models gated by a login.
Think twice before falling for this announcement and ask yourself what they are not telling you.
Muse Spark 1.1 was released July 16th, less than a month ago. A new version release this soon (particularly after Kimi K3's release drastically overshadowed it) is a bit sus and it appears that Meta is trying a first launch do-over.
Doesnt seem suspect to me, training runs have checkpoints and there is no reason you cant release a checkpoint even if you are still training the model
Frequent minor version bumps are pretty common these days. Opus 4.7 -> 4.8 was 42 days.
Which was in itself a do-over because Opus 4.7 received a lot of bad press on suspicion of being a regression from 4.6.
Somewhat surprised that Meta with all their resources couldn’t make a model that matches Composer on any frontier. All the Sparks are dominated by some other model everywhere along the frontier. Nothing fancy here since Llama defined the open model.
The use traces must be crucial to functionality which is why they’re keeping prices so low.
They rebooted less than one year ago so this is decent progress. Obviously users don't care about progress though.
Yeah, progress is useful as an internal metric, but I'm going to measure against the present frontier unfortunately. Eager to see what they come up with in the future.