Muse Spark 1.1
ai.meta.com
ai.meta.com
From Terminal-bench-2.1 details,
> We use a bash-tool-only agent harness to evaluate 89 Terminal-Bench 2.1 tasks from the official repository, where resources are capped at 6 CPU cores and 8GB RAM.
This disqualifies the results. Each terminal bench task has a cpu upper limit and RAM upper limit. Overriding either is disqualification.
For reference, in tbench-2.1,
1. 0 out of 89 task allow 6 cpu cores (highest is 4, and i think only 1 task)
2. 8 out of 89 tasks allow 8GB RAM
This kind of shady benchmarking (I was talking about it just yesterday in a different context https://news.ycombinator.com/item?id=48838212) takes all joy out of building a harness to improve benchmark performance of a model because no matter what you do, you won't beat the headline (cheating) number. This is presumably why this model is not in the official benchmark leaderboard https://www.tbench.ai/leaderboard/terminal-bench/2.1
As an ex Meta employee, this is a little sad but not massively surprising. 'Number go up' is the core performance evaluation metric until PSC is done and you move on.
I thought the benchmark was supposed to be about terminal use and specifically chaining together lots of bash tool calls. Which test cases does this matter for?
A few examples from memory:
1. This task [1] asks the agent to train a CNN under 1 CPU, 2GB RAM, 10GB storage. If you allow high resources, weaker models often succeed (the most clock time actually goes in waiting for the network to train).
2. This task [2] asks agents to implement a complete MIPS interpreter in JavaScript in 1 cpu and 2GB RAM. A common failure mode is OOM, at least in the earlier buggy versions that models run to get feedback. When OOM hits, the task is killed, no do-overs.
3. A lot of tasks involve building projects with a single core supplied. If you use -j12 type options, it will actually be _slower_ to build and the task will more likely miss the timeout. Having more threads squeezes the end to end time. This is a big one actually since the most common failure mode (from what I have seen) is the task timeout hitting before the agent finishes
[1] https://github.com/harbor-framework/terminal-bench-2-1/blob/...
[2] https://github.com/harbor-framework/terminal-bench-2-1/tree/...
The test cases of "don't melt my computer" and "be a good (computational) neighbor"
https://www.anthropic.com/engineering/infrastructure-noise
Is anthropic benchmark maxxing and cheating on terminal bench too? They don't follow the strict resource "limits" either
Sure for old tasks you could argue that now its not required to boost because infra errors are alleviated with better default limits. My point more so is that its a strange thing to index on because if you wanted to cheat on the benchmark, it does not particularly seem like something that shifts results? Once the API is out maybe I'll eat my words, but I don't really believe that if you manually tried to reproduce the results with lower limits you'd see significantly different results
Did harbor / tb2.1 cap the swap available to docker runs?
There used to be a bug that would allow dockerized instance runs to use more memory than the specs allowed. Some of the original tasks weren't really possible to complete without exploiting swap. Even the oracle solutions didn't pass if you stopped docker from having access to swap.
I think crack-7z-hash and filter-js-from-html had that problem off the top of my head, but i haven't looked at this in months, so i'm not sure
uv tool install llm
llm install llm-meta-ai
llm keys set meta-ai
# paste API key here
llm -m meta-ai/muse-spark-1.1 "Generate an SVG of a pelican riding a bicycle"
For comparison, here's the pelican I got from Muse Spark 1: https://simonwillison.net/2026/Apr/8/muse-spark/
I'm also increasingly worried that I'm not providing enough value in my model reviews. It's really hard to get a useful and credible idea for the strengths of the new models.
He doesn't have to match Anthropic or OpenAI model revenue if he can deflate theirs by 99%.
All he has to do is keep spending a few billion dollars developing frontier models, release them as open weights, and turn coding models into a commodity. He also needs a good OSS reference harness to match. Very few people are in a position to do this and for it to make business sense.
That's quite likely where things are headed regardless, and he could speed it up significantly.
We should all hope models move from proprietary products to commodities the way compilers did.
This may be one of the best things Zuck could do for the world.
If meta releases an open-weight LLM that is not Chinese made, cheaper to run than the SOTA premiums, etc, it would lower the number of people paying for frontier labs models. We saw with with early LLAMA models, but they didn’t keep up in the race with v4.
Meta would benefit from this, not from increased revenue at the hands of open LLMs, but from reduced competition. Meta competes with Google for ad spend, and lowering the Google revenue (or increasing costs) from AI reduces the competitive advantage. OpenAI wants to build an ad engine, so same thing will apply there too - make it less-revenue-generating to compete. Meanwhile G, OpenAI, and Anthropic are huge talent sinks that they have to compete with, especially for ML talent which is core to Metas business goals (ads). Finally, Meta needs lots of GPUs to train their ad engine models. By reducing the revenue-per-GPU of these labs, they’re reducing demand on a core revenue generating supply they have to compete for.
I guess we'll see how Meta did this time.
Software is easy to define as “working”: just run it. But - useful software requires an absolute truck worth of code - 100k loc before you’re talking about a real product, or else dozens of iterations of a toy you make for yourself before it’s useful enough to quit toying with and just use for what you wanted it for.
Sure, the success of software is hard to anticipate and what “good” UX is is hard to pin down - that’s not what I’m talking about. I’m talking just making the code and having no lint errors. That shit is a slog but it’s a slog with a clear goal amenable to hill climbing.
Through that lens software is mostly pattern matching. It’s very rare that an activity in software construction is out of distribution because even if the core of the thing is novel it needs a massive blanket of UI and a tech stack and a production environment to run in and observability and and and and. You get it I hope.
Meanwhile most work out there is a mess of undocumented, un-codifiable detail with no objective criteria for success, only a very wide gradient of “job well done” to “what is this garbage go and fix it”.
We are solving the easy parts of software and soon all that’s left will be the parts that are just like other work. And then we engineers will also be doing mostly squishy subjective judgment stuff.
the muse family is no longer open-source. it is still priced very "cheaply", but a significant shift in approach.
https://dev.meta.ai/docs/getting-started/pricing-rate-limits
If they have a really good model, it makes sense to subsidise it, to gain users, before they align prices with competitors.
I really dont see how anyone's willing spend more than $1.50 per mm output. Let alone $15-50. Does anyone actually pay for usage based billing as a consumer?
https://platform.claude.com/docs/en/about-claude/pricing
Model Base Input Tokens 5m Cache Writes 1h Cache Writes Cache Hits & Refreshes Output Tokens
Claude Fable 5 $10 / MTok $12.50 / MTok $20 / MTok $1 / MTok $50 / MTok
Claude Opus 4.8 $5 / MTok $6.25 / MTok $10 / MTok $0.50 / MTok $25 / MTok
Note Fable costs $50 MTok and Opus 4.8 costs $25 / MTok.
Even with usage based billing I'm below $1 writing code all day.
same, pi as a harness works best, better than claude code and open code.
Grok 4.5 has a relatively high $0.50 per 1M cached input token rate, compared to $0.15 on this model.
Grok 4.5 cached input costs the same as Opus 4.8 cached input, which is going to make it a lot more expensive to use for multi-turn coding than many would assume from the $2/$6 headline numbers they led with.
There's a further sting in the tail, Grok 4.5 is only $2/$6 for the first 200k of context. Go above that, and the pricing is $6 / $12 - and you're still capped at only 500k context anyway.
Here's the xAI pricing on OpenRouter:
https://openrouter.ai/x-ai/grok-4.5?endpoint=0e927811-b1a8-4...
Compare with Grok 4.5 which came out at $2/$6 but then quietly charges $0.50 per 1M cached input tokens. That's as high as Opus 4.8!
Not the way you're implying?
The GLM 5.2 hype was blowing way before this. Neither xAI nor Meta have really made a difference in a different way - similar results / similar pricing (to GLM 5.2).
Not sure I agree. Angular fell behind in popularity but was (is? unsure atm) still eminently usable. I gave gemini a test drive recently and it was horrendous, as in "picking dirt cheap Chinese model over gemini any day" bad, and with overzealous guardrails to boot. 3.1 pro feels a year behind and is extremely lazy. 3.5 flash feels like a model you’d run on your 128gb macbook, not something that was released a month ago and which costs a fair bit when used through api.
In any case: as of right now I think that we went from a three horse race to anthropic / openai as premium choices vs whatever is the Chinese fotm for a fraction of the cost. 3.5 pro better be a miracle if google wants to hang out with the big boys, otherwise their only strategy is hoping that both US labs go broke and they remain the last man standing.
Likewise, Gemma4 models are unbeatable at their size.
So I use Gemini for coding? Hell no, but that’s not the same as Google failing writ large.
Googles only real goal is to retain the ownership and dominance of the online world they have now, and Gemini is doing exactly what it needs to do
Hell, difference between Opus and Fable isn’t that big. I would still prefer Opus for most task because it’s cheaper. Break things down into smaller tasks and create deterministic parts of the process and you need the more expensive models less and less.
Anthropic needs to get their token efficiency under control or they will lose long term.
- Chinese models
- Grok
- Meta
- OpenAI
- Anthropic
I think this is a win. I'm building like crazy to take advantage of all these subsidized tokens while I can.
On the one hand, because it is easy to build products, more and more people will build. And more and more products and features will be built. However, a lot of people who are non-technical will also try to build, but they get stuck, and then they will need engineers. The sheer volume of product built by both experienced technical companies and non-technical novice startups and founders and wannabe founders is going to be massive. That is the bull case for having more software engineers needed in the near future.
On the other hand, in a year or so, people will build all these products, and most of them won't be able to market them, sell them and make money. Eventually, there won't really be a need for that many software engineers.
I think overall the bull case is probably going to win net net.
And those people won't need to be software engineers.
> but they get stuck, and then they will need engineers
You've implicitly assumed here that the AI systems will always be worse than the average engineer. That is IMO myopic. I'm not sure that it's even true now let alone in the nebulous future.
Most of what we do as engineers is precisely describe or analyze the behavior we want or the behavior we don't want. All other engineering skills that are useful are ultimately downstream from understanding the behavior of software enough to know which parts to keep, improve, or jettison. Chatbots can take care, somewhat, of analysis or expansion of instructions.... but they can't read minds. I don't see that changing any time soon.
I don't know who needs to hear this, but neither can humans.
You've implicitly assumed here that AI systems will always be worse at contextualizing and framing questions than the average engineer. I'm not sure that it's even true now let alone in the nebulous future.
You haven't narrowed the fundamental myopia of the assumption here, just dressed it in slightly different clothing.
How would they know what to ask or contextualize if they don't know what the user wants?
> How would they know
How would you? The answer is the same.
If you're implying chatbots can ask their "client" what to build, good luck with that—contractors are at least liable for what they produce and have extreme incentives to ensure that their clients are happy. To the extent of refusing to build anything if they don't know what they want....
If you had psychic mindreading powers you would understand what I mean.
Plus all the other things that software engineers generally have not learned to a professional level even if they picked up the basics on the job by osmosis, because figuring out the customer's needs (and what they'll pay you for which may be different) is the job of a business analyst, a PM, or a UX researcher, and those are different skills and two of them may come with a Business Informatics degree rather than a CompSci one.
LLMs can be "eh, better than nothing" at many things, not just code.
Plus, this is software "Engineering" we're talking about, which famously gets scare quotes in comparison to all the other forms of engineering because unlike them we don't have as standard things like professional liability insurance to cover serious professional errors of judgment the way someone who signs off on a bridge that collapses would have.
This question is completely disconnected from reality. If you try to sue a human for proposing something more complex than what you need you will waste a lot of money and then lose the lawsuit.
Also the annual cost of too much small company infrastructure is less than the cost of even a single good human engineer.
I’m “seeing” more devs stepping into the SendCutSend stage where they’re cleaning up/fixing/productizing vibe coded projects so maybe there will be some new demand in that space?
Another example is when the WWW first became available, and suddenly everyone COULD be a publisher (browsers even included built-in HTML editors), and for a while MySpace pages proliferated until the excitement died down and people went back to being media consumers.
I expect we'll see the same thing with consumer use of generative AI. Suddendly everyone is generating 3-D worlds/games with Fable because they can, but I expect that just as with the web the novelty will wear off and they'll leave it up to the pros.
Professional use of GenAI, and coding in particular, is certainly here to stay, but it seems we're still in the early experimental/hype phase. At least tokenmaxxing has passed, and it seems most companies are now paying attention to, and limiting, how much they are spending, but it doesn't seem we've yet progressed to the stage where companies are paying attention to what they are actually getting out of it - is the money spent showing up on the bottom line in the form of increased revenues.
I'm hoping vibe-coding plays out the same way.
I think for a lot of type of software we have now reached peak employment.
Someone payed a few k just for a normal website.
Do you think that Chinese software industry is that relevant to the kind of software market talked about on HN? I.e. lots of enterprise b2b and infra companies.
Chinese companies have always had a very low willingness to pay for software which kinda breaks the flywheel of B2B SaaS companies and companies to service those companies all the way down.
They have had real issues with deflation rather than the inflation most Western countries have seen over the past five years.
Are we still left with this mindset? Maybe once upon a time but it has definitely been changing.
There's plenty of B2B and enterprise SaaS companies in China serving the Chinese market. Maybe not as many, but no longer the very low of the past.
I also would not say enterprise were not willing to pay, even many years ago. It's the SME that refused to pay. Large CRM, ERPs etc have always existed.
Happy to be wrong though if I'm missing something.
Exactly it's mindset. I used pulled up 1st result when I used Google [0][1] and it is ~20%. It also forecasts huge growth potential.
[0]: https://www.grandviewresearch.com/horizon/outlook/software-a...
[1]: https://www.grandviewresearch.com/horizon/outlook/software-a...
> reflecting that SaaS in China is much more services and implementation heavy
I wouldn't read into it too much. There's creative accounting and other things involved. Also what is services and not e.g. "premium support"?
> So it feels like the direct to business SaaS's like Salesforce aren't really there and then then the selling to SaaS titans like Datadog have a much smaller flywheel to work in.
US companies struggle for many other reasons that I rather not go into but definitely not a SaaS nor China issue. I mean yes it is a China issue as in businesses need to adapt to different cultures rather than 1 size fit all but that applies everywhere.
What I really want is Claude as a deep part of the operating system.
If that happens then a whole lot of the abstraction of software vanishes along with what we think of today as software jobs. I think many new forms of knowledge work would emerge from this though.
I would think that needs massive local compute but I can't imagine that is not the future down the line.
Luckily, China is on the verge of a true breakout, I’m not sure what exactly it will be - but I’d make a very large wager the “next iPhone” is Chinese, and will constitute a full blown “Sputnik moment” for the US and SV.
If Americans weren’t forbidden to own Chinese EVs they’d know this. But tariffs mean the breakthrough will be even more unexpected.
Since Chinese actually “sell stuff” I’m guessing their unbeatable lead in AI efficiency, manufacturing, and distribution will produce a step change breakthrough within a decade.
1. You have a task and it requires an expert to perform it (software engineering, where we were) 2. You automate the task, it still requires the expert to babysit it (where we are now) 3. Management works out that the expert is just monitoring what the automation is doing and has actually lost expertise because they are just watching, not doing. Pay collapses. (where we are going)
- DeepSeek
- GLM (Z.ai)
- Minimax
- Kimi (Moonshot)
- Hy3 (Tencent)
- Qwen (Alibaba)
(Each one of these with weights available to download and run locally)
No wonder we still can’t get climate change under control
This is was historically a money issue, being green used to be wildly more expensive.
Now being green is cheaper, the limiting factor is how fast PV and batteries can be made or imported.
Recent reports of the sum of all US data centres currently in planning, has a power demand exceeding the (capacity-factor-adjusted!) global annual supply of new PV.
This would be less of a problem, but still a problem, if Trump wasn't trying to get in the way of anything green, or if the companies building data centres decided to also support factories to make more PV.
* Planned new demand: 300 GW; PV factory capacity ~ 600 GW nameplate, but the capacity factor is 14% so that's really 84 GW on average.
I do not know if competition is good, we will see in a few years.
Looking forward having a physical job for a change :D
I do not mean Suckerberg or Eric Schmidt.
To be fair, seems more correct to compare against similar strength models if your main edge is pricing.
Second, compare to older versions of competitor s models.
Still does not look good? Compare to own previous models.
“Good” benchmarks to gauge development skills at the moment seem to be:
DeepSWE [0] by Datacurve
FrontierCode [1] by Cognition
And then there’s TerminalBench, which I’m certain has been saturated in post-training to no end, so I wouldn’t think of it as a gold standard anymore.
But yeah, in general, it’s not going to get easier knowing which benchmarks are actually measuring “frontier” capability, and which are just getting results inflated by way of time/token budget [2], ever again.
[0] https://deepswe.datacurve.ai
Paid prompts are not used for training. Data retention is not entirely clear, they will keep it if they think you're breaking the rules, but it's not clear how long they keep it if they don't think you're breaking the rules.
What kind of use case would be best for that shape?
Bug diagnostics is about being okay at coding but better at tooling.
Given a good diagnostic report, it can be handed to opus for the fix.
Opus is okay at writing reports, but it still regularly gets table widths wrong in typst documents, leaving the last column full of text but only a handful of characters wide.
I see models fail on tool calls that involve API requests to a specific API, internal or cloned Makefile calls, npm run commands, etc.
It's some kind of parsing or integration error due to what I think is codex not anticipating server-side tool calling and how meta treats those ids... first couple times running codex with muse, it would fail on its first non-web search call.
Got it fixed, not personally sold on the bespoke server-side tool calling and indefinite file storage yet, but also a very cool model that I'm enjoying using so far!
https://github.com/accretional/awesome-muse-spark/blob/main/...
I don't know where I need to sign up to try it out. What is pricing? Is it API or subscription, what?
I had the exact same experience with Grok 4.5 as well.
:(
Well, Vietnam is not in the list of restricted territories.
Anyway, what is "your region" ?
Is this where I am now, or is it where I activated my Oculus 2 five years ago ?
The reason: Its writing style feels "unique", and I find it pleasant to read for science-based topics.
I never ask _ONLY_ Meta AI, but the answer it gives is almost always in a distinctly different style than other frontier LLM's.
I think this is because of the unique JEPA architecture they have, but that's a layman's hunch.
> The extra resources enable the agent to try approaches that only work with generous allocations, such as pulling in large dependencies, spawning expensive subprocesses, and running memory-intensive test suites.
> An agent that writes lean, efficient code very fast will do well under tight constraints. An agent that brute-forces solutions with heavyweight tools will do well under generous ones. Both are legitimate things to test, but collapsing them into a single score without specifying the resource configuration makes the differences—and real-world generalizability—hard to interpret.
So changing the resource limits changes the benchmark. Yet their score table claims their score to be for Terminal-Bench 2.1, not Terminal-Bench 2.1 with raised limits.
I for one am really glad to get competitive models that will push the major labs to bring prices down. While Chinese open source labs are also great, unfortunately when it comes to US/Western political pressure it won't often have as much of a bearing on labs bringing prices down, especially for enterprises.
Also if these numbers are true, this is truly breaking ground finally for Meta.
eg. Model X is weaker than Fable, but competes well with Opus/Sonnet and costs 1/5th as much etc - something similar playing out with Grok 4.5.
Meta's AI org when from a total mismanaged dumpster fire for multiple years to delivering a competitive model in less than a year on essentially their first try?
so yeah, this is essentially their first try with a completely new org.
Even given that, this is their second try (they had Spark 1.0). Spark 1.0 was uninteresting, this is potentially interesting, but we can't really try it yet it seems (at least not in Openrouter).
Ultimately, competition is now fierce in this broad level of intelligence/cost: Spark 1.1, Grok 4.5, GPT 5.6 Luna, GLM 5.2
Sonnet not in the same ballpark of pricing (more expensive than Opus in many cases). Haiku has been basically abandoned.
I have questions regarding if I should even care but I don't so Meta please keep enjoying the irrelevance. lmao
I'm going to assume the only "region" that's permitted is the USA.