Meta is one of those companies where, if there is anything remotely comparable, I'm happy to pay more to not use them. They've had a profoundly negative impact on society and Zuckerberg is not who I want controlling the future at the top of AI.
I feel the same about Grok w/ Elon. I will pay extra to use someone else.
I'm not an Amodei stan, but of all of these people he seems to have the most ethical focus. Again, not everything done perfectly and I have my gripes, but of the leaders of frontier labs, I'll vote with my money.
And, yeah, I wouldn't trust sama to watch my bag while I went to the bathroom.
If it was up to Dario we'd all be banned from using open-weight models, and we'd have to be investigated for PRC connections before sending our allotted five API queries a week.
"Avoid generic tangents" / "Please don't complain about tangential annoyances."
That's pretty much 90% of HN these days.
Apple releases a new iPhone? Here comes the flood of decade-old complaints about long-discontinued Mac butterfly keyboards and walled gardens.
Microsoft releases a new version of Windows? Here come the gripes about Azure.
Google changes something in GMail? Play Store!
It's like there's an army of bots out there determined to reduce the productivity of the Western tech bubble by diverting everyone into endless circular arguments about absolutely nothing of relevance to the topic at hand.
Is there a reason these pelicans always have roughly the same composition (side-view, 2d, biking right, flat ground beneath, etc)? I don't see any of that detailed in the prompt, yet they all seem to generate roughly the same image of differing quality.
The more generic your prompt, the more generic the response. It's a regression to the "mean" of the training data aka GIGO for AI.
It's like when you ask your average person off the street to draw a house - it'll almost always be square with a triangle roof, one door, and two windows.
In the pelican/bike example, it's probably a bit of a self-perpetuating snowball too. If the earliest examples were bike left-to-right, flat ground, etc. then they are also being scraped up in future LLMs.
I was going to ask the exact same question earlier but deleted it after thinking “I’m sure Simon has done some sort of discussion on this.” Since it does seem novel to you, too, it would be really interesting to read more about this phenomenon.
Search Google Images for "bicycle". Almost all bicycle product shots are staged the same way: side view, going left-to-right. It makes sense to me that given that skew in the training data, the model grounds itself in the bicycle.
and furthermore, this is because the drivetrain is ~always on the right side of the bike - if you want to inspect or admire a bicycle you look at the right side, as you might look under the hood of a car.
(Why the drivetrain is on the right, I don't know. But most bike parts follow open standards so it's quite entrenched.)
> and furthermore, this is because the drivetrain is ~always on the right side of the bike
While I'm sure this factors into things for advertisements for bike components, there is also just a general preference that westerners have for left-to-right motion. Not just in bike ads, but all ads with (or suggesting) movement. And also not just ads, but movies where directors believe left-to-right motion is associated with progression and right-to-left motion is regressive.
Yes, I do a thing where I ask the machine to generate responses in the form of a lizard talking to a cat. The lizard is always a green gecko and the cat is always orange, which I never specify.
I started using Spark 1.2 for development because if you're willing to let Meta train on your data it was dirt cheap and was actually really pleasantly surprised with it. It's not a frontier model by any means, but for work that didn't require a top of the line model, I really enjoyed using it.
I'm anthropomorphizing it a bit, but it felt like it knew its weaknesses and didn't try to impose it's opinions on me. What I mean by that is that it did what I told it and if there was something unexpected in the code that it put out it was often because I gave it ambiguous or conflicting instructions. It didn't try to go above and beyond and just acted like a tool, which is what I want from a coding agent 90%+ of the time. I also felt that it did a much better job of following established patterns in my code than many of the other current models do. I'm a huge fan of OpenAI's models and Spark 1.2 is what I expected 5.6 Luna to be.
I'm curious and a little excited to use 1.3, but honestly a little worried that as Meta pushes for better benchmarks that Spark will start to fall into the trap of trying to be "helpful" in ways I don't want it to be.
Tangential, but when I first started using Spark 1.2, it made me realize how much I miss 5.3 Codex. That model was the peak of coding models, IMO, in that it knew how to write good code, but didn't try to overstep or be "helpful" in unexpected ways. That got me thinking about how the major labs seem to be stepping away from coding focused models toward more general purpose ones and how I can't help but feel like that's a mistake.
A model that (at least in benchmarks) is getting closer to SOTA. A clear separation between what’s used to improve their products and what’s not (at least this is what they claim).
Good job Meta! Seriously. This is almost making me forget about the 18B$ lawsuit for children social media addiction.
DeepSWE scores 75.4 - that's the best score so far. And it's crazy cheap!
Google held the top a few hours today with Gemini 3.8 Flash, but now second to Spark 1.3. All this competition will drive prices down!
With the contributor pricing being more than 10x cheaper than the standard, that would make it best and cheapest on the DeepSWE leaderboard! It feels fast in my experience too. LLMs keep improving at an insane pace.
So one model is "Not used to improve our products" and is 10-20 times more expensive to the "Used to improve our products"-model.
Given this is Meta, my immediate assumptions that one is cheap because it lets me "be the product". I know I'm rushing to conclusions but there is zero trust here. The brain will do its thing. And the wording here is giving the brains a lot of wiggle room.
Given OpenAI and Anthropic's behavior, do you really expect them to be singled out for this practice? Zero trust has been in "LGTM" territory for years now. Meta's bet against people taking a principled stance arguably paid off great.
The previous version was, in my experience, the best free model available on OpenCode. It's been very good at simple/moderate tasks where I am precise in my ask and it doesn't need to make a ton of undefined assumptions. Hopefully this new version is also available on opencode for free.
I didn't like 1.2, It make some mistakes in a web app, so I quickly went back to Claude, Kimi K3 or Deepseek V4. Hope this one can clear agentic development, because Muse Spark models are fast and cheap.
Used Muse Spark 1.2 and was not impressed at all. Fast and cheap but even GPT 5.6 Terra felt much more capable. Also not really looking to support a company that was just forced to pay $18B for mental health damages.
I'm party using 1.2 to reverse engineer and re-implement an old game binary and it has been quite good and fast. The contributor pricing is very attractive, excited to try 1.3 and see if I feel a difference. 1.2 can get stuck outputting similar sounding thought summaries with no apparent progress when asked to solve bugs. Then I've switched to GLM-5.3-Flash which for this use case has been clearly better at finding suspected causes and following tracks.
I had no idea Meta has a coding agent harness. Does anyone have experience with it and can comment? The 1.3 contributor prices look very attractive. I'll probably start using their API if performance is good and the API is reliable with decent rate limits.
Meta is one of those companies where, if there is anything remotely comparable, I'm happy to pay more to not use them. They've had a profoundly negative impact on society and Zuckerberg is not who I want controlling the future at the top of AI.
I feel the same about Grok w/ Elon. I will pay extra to use someone else.
I'm not an Amodei stan, but of all of these people he seems to have the most ethical focus. Again, not everything done perfectly and I have my gripes, but of the leaders of frontier labs, I'll vote with my money.
And, yeah, I wouldn't trust sama to watch my bag while I went to the bathroom.
If it was up to Dario we'd all be banned from using open-weight models, and we'd have to be investigated for PRC connections before sending our allotted five API queries a week.
"Avoid generic tangents" / "Please don't complain about tangential annoyances."
"Avoid generic tangents" / "Please don't complain about tangential annoyances."
That's pretty much 90% of HN these days.
Apple releases a new iPhone? Here comes the flood of decade-old complaints about long-discontinued Mac butterfly keyboards and walled gardens.
Microsoft releases a new version of Windows? Here come the gripes about Azure.
Google changes something in GMail? Play Store!
It's like there's an army of bots out there determined to reduce the productivity of the Western tech bubble by diverting everyone into endless circular arguments about absolutely nothing of relevance to the topic at hand.
4.2266 cents, 38 seconds.
For comparison here's Muse Spark 1.2, which animated it without me asking it to: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
The 1.3 one is definitely better - better bicycle frame, better wing, better pelican hat.
UPDATE: Here's another one with five pelicans for each of the five Muse Spark 1.3 reasoning levels: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
The most expensive was reasoning level xhigh - 7.5 cents, 1m34s.
And I ran five pelicans at all reasoning levels for 1.2 as well, here: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
Is there a reason these pelicans always have roughly the same composition (side-view, 2d, biking right, flat ground beneath, etc)? I don't see any of that detailed in the prompt, yet they all seem to generate roughly the same image of differing quality.
The more generic your prompt, the more generic the response. It's a regression to the "mean" of the training data aka GIGO for AI.
It's like when you ask your average person off the street to draw a house - it'll almost always be square with a triangle roof, one door, and two windows.
In the pelican/bike example, it's probably a bit of a self-perpetuating snowball too. If the earliest examples were bike left-to-right, flat ground, etc. then they are also being scraped up in future LLMs.
as a kid I did them like this. nobody told me to do that. are we all so similar?
It's really interesting, isn't it? They almost always cycle from left to right - but I have had a few which cycle in the other direction.
The 2D / flat ground feels reasonable for a SVG, which implies a vector illustration.
I was going to ask the exact same question earlier but deleted it after thinking “I’m sure Simon has done some sort of discussion on this.” Since it does seem novel to you, too, it would be really interesting to read more about this phenomenon.
Search Google Images for "bicycle". Almost all bicycle product shots are staged the same way: side view, going left-to-right. It makes sense to me that given that skew in the training data, the model grounds itself in the bicycle.
and furthermore, this is because the drivetrain is ~always on the right side of the bike - if you want to inspect or admire a bicycle you look at the right side, as you might look under the hood of a car.
(Why the drivetrain is on the right, I don't know. But most bike parts follow open standards so it's quite entrenched.)
> and furthermore, this is because the drivetrain is ~always on the right side of the bike
While I'm sure this factors into things for advertisements for bike components, there is also just a general preference that westerners have for left-to-right motion. Not just in bike ads, but all ads with (or suggesting) movement. And also not just ads, but movies where directors believe left-to-right motion is associated with progression and right-to-left motion is regressive.
Sun is missing a few rays and not wearing sunglasses.
Yes, I do a thing where I ask the machine to generate responses in the form of a lizard talking to a cat. The lizard is always a green gecko and the cat is always orange, which I never specify.
Has any ab tried to game this yet and just made the most amazing pelican by hand and always reply with that?
If you have a grading rubric, huge points off for adding arms instead of using the wings as arms!
I think it's hilarious that this detail is enough for me to dismiss looking into the model, but here we are, and it is.
No one cares.
These benchmarks you guys invent for yourselves prove nothing.
A small model like Mistral 7b is just as useful to the end task as any larger model, if not more so because it’s faster.
Your big model may be able to draw pelicans or solve some esoteric nonsense but it cannot do real work in the real world.
These phoney benchmarks and experiments mean nothing.
None of the LLMs can replace a software engineer nor even a barista or car mechanic etc. not even close.
Instead of inventing fake benchmarks do something tangible and tell me how it performs.
Before laying off half the country and going full retard on AI
Someone did care enough to create a throwaway account to vent here, it seems.
What does the mean pelican look like at this point?
Also 3X token use vs. 1.2
lol
Definitely an upgrade over 1.2
I started using Spark 1.2 for development because if you're willing to let Meta train on your data it was dirt cheap and was actually really pleasantly surprised with it. It's not a frontier model by any means, but for work that didn't require a top of the line model, I really enjoyed using it.
I'm anthropomorphizing it a bit, but it felt like it knew its weaknesses and didn't try to impose it's opinions on me. What I mean by that is that it did what I told it and if there was something unexpected in the code that it put out it was often because I gave it ambiguous or conflicting instructions. It didn't try to go above and beyond and just acted like a tool, which is what I want from a coding agent 90%+ of the time. I also felt that it did a much better job of following established patterns in my code than many of the other current models do. I'm a huge fan of OpenAI's models and Spark 1.2 is what I expected 5.6 Luna to be.
I'm curious and a little excited to use 1.3, but honestly a little worried that as Meta pushes for better benchmarks that Spark will start to fall into the trap of trying to be "helpful" in ways I don't want it to be.
Tangential, but when I first started using Spark 1.2, it made me realize how much I miss 5.3 Codex. That model was the peak of coding models, IMO, in that it knew how to write good code, but didn't try to overstep or be "helpful" in unexpected ways. That got me thinking about how the major labs seem to be stepping away from coding focused models toward more general purpose ones and how I can't help but feel like that's a mistake.
A model that (at least in benchmarks) is getting closer to SOTA. A clear separation between what’s used to improve their products and what’s not (at least this is what they claim).
Good job Meta! Seriously. This is almost making me forget about the 18B$ lawsuit for children social media addiction.
DeepSWE scores 75.4 - that's the best score so far. And it's crazy cheap! Google held the top a few hours today with Gemini 3.8 Flash, but now second to Spark 1.3. All this competition will drive prices down!
With the contributor pricing being more than 10x cheaper than the standard, that would make it best and cheapest on the DeepSWE leaderboard! It feels fast in my experience too. LLMs keep improving at an insane pace.
But is the score really reflective of the quality or are both models benchmaxxing?
how much of it is from reallocation of staff to ai training and labeling
So one model is "Not used to improve our products" and is 10-20 times more expensive to the "Used to improve our products"-model.
Given this is Meta, my immediate assumptions that one is cheap because it lets me "be the product". I know I'm rushing to conclusions but there is zero trust here. The brain will do its thing. And the wording here is giving the brains a lot of wiggle room.
I'm confused what your surprise is here. It's plain and simple right to the point wording.
I don't see the wiggle room at all.
aren't they explicitly saying this with both their pricing and their wording? I'm not sure what you are alluding to?
the meaning is pretty obvious - they want to train on your chats & tasks and are willing to subsidize for the privilege of doing so.
Given OpenAI and Anthropic's behavior, do you really expect them to be singled out for this practice? Zero trust has been in "LGTM" territory for years now. Meta's bet against people taking a principled stance arguably paid off great.
The "contributor" pricing is the standout here at a ~20x discount, if you allow training on your data.
The model seems on par with Sol and Opus 5 on paper (admittedly on some older/saturated benchmarks, but very competitive for $).
Stats:
1M context, $0.10 input/$0.002 cached, $0.20 output (Mtok)
Why they didn't use LLM to create html table instead of https://lookaside.fbsbx.com/elementpath/media/?media_id=1048...?
The previous version was, in my experience, the best free model available on OpenCode. It's been very good at simple/moderate tasks where I am precise in my ask and it doesn't need to make a ton of undefined assumptions. Hopefully this new version is also available on opencode for free.
I didn't like 1.2, It make some mistakes in a web app, so I quickly went back to Claude, Kimi K3 or Deepseek V4. Hope this one can clear agentic development, because Muse Spark models are fast and cheap.
“contributor” pricing at $0.10/$0.20 is crazy cheap if it’s measuring up to Sol.
Definitely shows how important a user data flywheel is for RL and model improvement.
Used Muse Spark 1.2 and was not impressed at all. Fast and cheap but even GPT 5.6 Terra felt much more capable. Also not really looking to support a company that was just forced to pay $18B for mental health damages.
I'm party using 1.2 to reverse engineer and re-implement an old game binary and it has been quite good and fast. The contributor pricing is very attractive, excited to try 1.3 and see if I feel a difference. 1.2 can get stuck outputting similar sounding thought summaries with no apparent progress when asked to solve bugs. Then I've switched to GLM-5.3-Flash which for this use case has been clearly better at finding suspected causes and following tracks.
Practically free for "contributors" at 0.2 usd/mtok. That's going to be hard to say no to for hobbyists.
Could this be best intelligence / $ if you're willing to let zuck digest your data?
I declined the use of cookies and everything went black. No content at all. Dissapointed.
Blog post: https://research.meta.ai/blog/introducing-muse-spark-1-3 (https://news.ycombinator.com/item?id=49541149)
I had no idea Meta has a coding agent harness. Does anyone have experience with it and can comment? The 1.3 contributor prices look very attractive. I'll probably start using their API if performance is good and the API is reliable with decent rate limits.
This should probably be primary:
https://news.ycombinator.com/item?id=49541149