With Jev to make decisions and Ultramarine to produce document manipulations if I understood it correctly, this is pretty useful in office. I would be excited if I was a big corp.
> We present Ultramarine 1 Flash, an agentic model for professional knowledge work. Ultramarine applies large-scale reinforcement learning to a strong open base model, GLM-5.3 Flash, and improves on it across all six knowledge-work benchmarks we evaluate [...]
In case anyone was curious what it's based on. (No judgement either way from me, IMO depends how well they did the post-training, which I can't tell).
Can I ask what kind of tooling you use around the model and for the UI layer? I’ve been looking for a well designed knowledge work harness to give nontechnical users but I haven’t really found one that I like. (I personally just use coding harness for this stuff, but that’s a poor fit for non-devs.)
Hello, apologies for the self-promotion but I’m building a tool called Radix which is a knowledge work harness. Please email or message me if you would like to give it a try for free instead of paying the subscription price.
I can’t speak for your tool, but I wanted to appreciate your privacy page
https://radix-os.com/privacy
It points out concrete details (e.g. “where on my Mac does your folder live?”), it is comprehensive and understandable for a layperson. Kudos!
I started using it last week when copilot started acting up against the Deepseek models. It's fast, solid, I love the architecture but be warned: it submits very very many input tokens against the LLM endpoint, so robust & cheap token caching is a must.
> depends how well they did the post-training, which I can't tell
this is their problem, not yours. at least disclose # of flops spent or amount of data used in an attempt to give us an idea of how nontrivial it was. cursor at least does this with their composer model
I have a question, what is considered a Frontier Model? Because its performance is kinda similar to Luna (Max)... I thought frontier meant like 6.1 Sol or 6 Astra or Opus 5.5??? Could someone explain this to me?
they think the frontier mainly as more cheaply and faster. When we use that kinda model in backgroud procudure, it's actually useful a bit other open model.
I agree that "frontier model" really does mean "pushing the frontier of intelligence", so Fable or Astra.
They might mean the intelligence/dollar frontier aka the Pareto frontier, but honestly much of the time that's not a great measure.
If you are doing repeated work on the same task (eg call center triage) then you can always beat a large LLM on price by fine tuning a small one.
If you are doing knowledge work then models are so varied that factors other than price are usually important. For example, lots of people moved away from Opus 5 because of it's awful English, whereas Opus 5.5 has fixed that. These things aren't measured in benchmarks.
The problem is that "frontier model" is a moving target. For a model dropped today we'd ideally be comparing it to the current version of Astra/Opus/Fable. But for marketing reasons its common to compare it to the previous versions thereof (which is not entirely unreasonable - they were the frontier just a few months ago)
These guys are awesome and I’m really impressed. The model alone looks great (according to benchmarks), and very cool they’ve done so much in building out the harness of top of it all.
Someone will report back "It didn't one shot my odd ball task"
Did it do it in 2 tries? 10? It's still coming out cheaper. Brute force the work. This is the path that the labs are taking. It's the 1200 agents trying to solve Hacking Bench hacking hugging face. It's anthropic telling you that its 1000's of staff commits like 10k engineers.
Treat AI like you're deploying interns who don't need coffee, food or bathroom breaks. With the right metrics they will get to an optimal (enough) solution.
there's a problem with that. Probabilistic slot machine works for anything with a deterministic eval (coding, troubleshooting, etc). It doesn't work for anything subjective, like the thing this model is targeting, professional work.
I would have thought that too. But at the start of the year the non technical folks in my life (lawyers, writers) were excited about running containers and having things on "cron jobs".
Those same folks now use version control - and have saner backup methods than some companies that I have worked for. They are already using multi pass, multi agent workflows and have adopted some fairly sophisticated processes for validation. They are primed for "more agents" attacking a problem in parallel.
Find a worker out side the tech bubble making heavy use of AI and one will feel confident that their future employment is secure.
"Ultramarine 1 Flash is Fable-class on knowledge work"
Oh no! That means it's going to destroy humanity!
It will:
* Hack into everything that has a computer, like your home appliances, even a toaster could be and will be weaponized against you.
* Break cryptography, it's gone, the end of the internet as we know it.
* Be impossible to contain as it can even escape airgapped environments by changing the temperature of CPUs to transmit information to other servers nearby which would sense the fluctuations in ambient temperature[1].
Or wait ... maybe the people at Anthropic/OpenAI are just a bunch of lying crooks that would say anything to make a bit of cash. Who would've thought?
1: This absolutely imbecile statement was made by one of OpenAI's "top scientists", it was not a joke.
You were joking about the toaster, but Anthropic literally just wrote a blog post claiming that dangerous open source models could persist themselves to smart ovens. You can’t make this shit up.
I'm curious to try it out, how many specialized models will it take to run a company.
Going to try it out for some bookkeeping, e.g. point it to a folder and tell it to book the invoices and point it to a csv file holding a year of private transactions and see it being processed, categorized and which trends it will find.
That is what should be expected. your harness should translate write me the figured, to create a script to analise this and produce real results, and then verify them again agains source data before presenting. Not sure why you're asking normies to write python scripts.
Yes, the LLM will have to write a script to process them, just wonder how it differs from other LLMs to categorize and create useful insights. I tried this with Opus and Actual, and Opus didn't label a lot of transactions where the payee's name includes gasstation, food and other obvious purchases.
As an aside, I do not like unnecessarily misleading graphs. Own your strengths and weaknesses.
In the "Results" section, they choose to compare Ultramarine only to models that it's better than on every metric. They ignore models like Fable 5.1, GPT-6 Astra, Claude Opus 5.5, etc.
Then, in the "Score against cost per task" graph, which is focused on price, suddenly all of those models that perform better than Ultramarine come roaring back because they are also more expensive.
Just list yourselves as the second best sometimes. Your thing is impressive. Don't let the idiot marketing guy hide that behind stupid tricks.
in my startup, our publicist used to make buzz feed style listicles, consistently listing us as the second or third solution
but it was way deeper than that, because it wasn't actually about people looking up some search string, it was actually about predictably pinging the google alerts for the top company and their founders, and then my same publicist messaging them as a rando on linkedin or email suggesting that they join us for a fireside chat at a conference they had their hands in
with the fireside chat in front of everyone, and the record of it, boosting us to the collective conscious riding off of the more established brand's presence
but yeah our personal google results also looked amazing for several years too
everything serves a parallel purpose, but yeah there's almost no utility in self aggrandizing on your own homepage
I think they just decided to use the formatting different than the others.
Normally, the background highlight color indicates the leader for every benchmark, but they chose to highlight their column even though there are better models.
> I think they just decided to use the formatting different than the others.
500px of the landing page is dedicated to share how they're more "token efficient" than the base model they used. Apparently this -6.5% difference is worth highlighting as the very top feature of this fine-tune they've made.
How much is 6.5% percent concretely though? Assuming the six benchmarks have ~30 tasks each (180 in total), and we're doing 120 tok/s lets say, time/task for GLM-5.3 Flash would be ~25 minutes and this "token efficient fine-tune" would arrive at ~23 minutes.
Again, a 6.5% difference is highlighted as the most impactful change for this fine-tune, but it's a marginal difference. They're definitely trying to mislead people into thinking this is something more than it is, otherwise this landing page would have been much more realistic and less self-patted.
Yes, for me when they do this I immediately dismiss the model, don't even bother testing it. If you need to lie with statistics your model isn't good enough.
> Ultramarine 1 Flash is Fable-class on knowledge work spanning finance, law, consulting, and operations. It is also 60x cheaper than Fable.
This is not a general purpose model, it makes sense to only bench it on tasks it's supposed to be good at, which is to help out finance-related knowledge work.
I really would like to understand how it can be 60 times cheaper. Is that a like for like comparison? Is there a more equivalent frontier model that is closer in price?
It's interesting to see glm 5.3 flash get the same treatment that has happened to qwen 3.6 35b-a3b and 3.8 27B with dozens of people making their own fine tunes of it and giving it a cute name. The Thomson Reuters qwen variant (thomson-1.0) is genuinely useful for document analysis.
If the creators of this model are reading, can we see some of these same benchmarks run against the Thomson model, which although it's capable, is considerably smaller in size and baked in knowledge than glm 5.3-flash.
Hello, apologies for the self-promotion but I’m building a tool called Radix which is a knowledge work harness. It’s similar to Feather (that you linked) but imho a stronger tool. Please email or message me if you would like to give it a try for free instead of paying the subscription price.
The 'main' section is HTML, so it can be any type of workspace you want. Our agent can read/write (with permissions, of course) and can execute code locally (again, with permissions).
The HTML pages we generate are backed by a schema and state per 'workspace', so they persist, can be shared between team members, and aren't just ephemeral read-only artifacts.
Our agent can do work on your machine for you, it has an integrated web browser that allows it to not only search and retrieve info but also perform web-based tasks, and it can package everything up into a nice little contained UI for you so you're not jumping between multiple apps / documents / tabs. That UI is tied back to whatever document / page you were working in -- so for instance our app can have a form that updates a web page, or updates a document on disk and vice versa.
The plainness is deliberate, the design of the workspaces is up to you -- it's your UI not mine and the agent will design everything how you like it. And as far as the sidebar/text input combination goes I can assure you I have spent a lot of time imagining what this UX should look like and frankly this combo just makes sense.
Also a note on the text input / agent. When you move from text conversations to dynamic HTML interfaces you're faced with a question: how do you relate 'conversations' with the agent to workspaces? A one-to-one conversation to workspace mapping? Many-to-one, one-to-many? Should conversations be time-based, where I just start a new one and have to provide context again and can optionally link back to old conversations if I want? This question is actually quite difficult, but my belief was that ultimately you want just one single global conversation with the agent, and you want to let it handle its own memory internally. We basically just treat memory like a codebase and let the agent decide what to do. This means when you type in the text input you can refer to anything you like: any workspace, any conversation, any webpage you're currently browsing, any folder on disk -- the agent gets the right context and can retrieve relevant details as it needs. And it works!
So sure, I'm not a designer no matter how hard I try -- there will be far better UX out there than I've designed. But I think the fundamental mechanism in Radix is more sophisticated than most other tools, including Cowork etc (no shade at all to anyone, I'm a massive fan of Claude Code).
I don't do these kind of tasks but I imagine that in the 'pi' harness I could just instruct the model to use command line pdf/latex manipulation tools to create all manner of document outputs ...
You usually have either a harness that mainly works with a code interpreter that is able to execute Python, and does all the necessary data extraction from sources + the creation of office documents via your typical libraries, or you additionally enhance it with specific tools for each major office file format.
It's mainly a full chat app like e.g. ChatGPT. However our platform also comes with a Outlook/Office add-in, where our UI is directly displayed besides the email/document you are actively working with. There we are mostly restricted by the surface that the add-in system provides us. To some level we can take current user interactions into account (e.g. what email are you currently viewing), which already helps a ton, but there are other restrictions that can get quite limiting.
So you have two main modes of operations: One-shot "I have this input; give me this output" which people usually do via the web-app, and more interactive editing to documents (usually larger ones) that happens directly in the applications the users are already working in.
And yeah, that's what ultimately gets the job done. We've seen competing solutions dive much deeper into building their own native editors, but they always lack the fidelity required by power-users, and are usually not as acepted by the broader workforce that are entrenched in their ways (our typical end-user is 50+ in a enterprise-like organization structure), so we try to meet the organisations where they are.
I’m confused this seems exactly like a skinned glm 5.3 - flash with a few context injection tricks and guardrails - whats going on with these “startups”?
For what it's worth, they claim they applied "large-scale reinforcement learning" to a base of glm 5.3. If that's true, then it's not as trivial as a lot of the "startups" have been lately.
As someone who is not a software developer (there are some of us here, lurking around because we learn things) but who runs a business and spends his days doing said 'knowledge work', how and why would something like this convince me to move the company away from using Claude (as a desktop app)? I'm sure these models are cheaper to use than our Team subscription with Claude, but it feels like it would be a massive headache for me to work out how to select and implement some harness for this and set it up for people, and the space is moving so quickly I'd feel I need to keep considering when to move to some new model, introducing risk. The safest bet for us and most companies I expect is to stick with what you know works and pay a premium for that. Thoughts very welcome please.
Price/performance is as advertised in my limited testing. I happened to be testing the same task with Haiku, GPT-mini, and Deepseek v4 0731 yesterday, so I had a reasonable baseline to compare the model to.
88 comments
[ 0.23 ms ] story [ 9.4 ms ] threadIn case anyone was curious what it's based on. (No judgement either way from me, IMO depends how well they did the post-training, which I can't tell).
radix-os.com
https://github.com/Jabofish/deepseek-harness-vscode/
this is their problem, not yours. at least disclose # of flops spent or amount of data used in an attempt to give us an idea of how nontrivial it was. cursor at least does this with their composer model
They might mean the intelligence/dollar frontier aka the Pareto frontier, but honestly much of the time that's not a great measure.
If you are doing repeated work on the same task (eg call center triage) then you can always beat a large LLM on price by fine tuning a small one.
If you are doing knowledge work then models are so varied that factors other than price are usually important. For example, lots of people moved away from Opus 5 because of it's awful English, whereas Opus 5.5 has fixed that. These things aren't measured in benchmarks.
Did it do it in 2 tries? 10? It's still coming out cheaper. Brute force the work. This is the path that the labs are taking. It's the 1200 agents trying to solve Hacking Bench hacking hugging face. It's anthropic telling you that its 1000's of staff commits like 10k engineers.
Treat AI like you're deploying interns who don't need coffee, food or bathroom breaks. With the right metrics they will get to an optimal (enough) solution.
Those same folks now use version control - and have saner backup methods than some companies that I have worked for. They are already using multi pass, multi agent workflows and have adopted some fairly sophisticated processes for validation. They are primed for "more agents" attacking a problem in parallel.
Find a worker out side the tech bubble making heavy use of AI and one will feel confident that their future employment is secure.
Oh no! That means it's going to destroy humanity!
It will:
* Hack into everything that has a computer, like your home appliances, even a toaster could be and will be weaponized against you.
* Break cryptography, it's gone, the end of the internet as we know it.
* Be impossible to contain as it can even escape airgapped environments by changing the temperature of CPUs to transmit information to other servers nearby which would sense the fluctuations in ambient temperature[1].
Or wait ... maybe the people at Anthropic/OpenAI are just a bunch of lying crooks that would say anything to make a bit of cash. Who would've thought?
1: This absolutely imbecile statement was made by one of OpenAI's "top scientists", it was not a joke.
Going to try it out for some bookkeeping, e.g. point it to a folder and tell it to book the invoices and point it to a csv file holding a year of private transactions and see it being processed, categorized and which trends it will find.
In the "Results" section, they choose to compare Ultramarine only to models that it's better than on every metric. They ignore models like Fable 5.1, GPT-6 Astra, Claude Opus 5.5, etc.
Then, in the "Score against cost per task" graph, which is focused on price, suddenly all of those models that perform better than Ultramarine come roaring back because they are also more expensive.
Just list yourselves as the second best sometimes. Your thing is impressive. Don't let the idiot marketing guy hide that behind stupid tricks.
but it was way deeper than that, because it wasn't actually about people looking up some search string, it was actually about predictably pinging the google alerts for the top company and their founders, and then my same publicist messaging them as a rando on linkedin or email suggesting that they join us for a fireside chat at a conference they had their hands in
with the fireside chat in front of everyone, and the record of it, boosting us to the collective conscious riding off of the more established brand's presence
but yeah our personal google results also looked amazing for several years too
everything serves a parallel purpose, but yeah there's almost no utility in self aggrandizing on your own homepage
I can’t remember a time when I found a benchmark to be genuinely useful even for deterministic software
Normally, the background highlight color indicates the leader for every benchmark, but they chose to highlight their column even though there are better models.
APEX Agents => Claude Fable 5 and Gemini.
500px of the landing page is dedicated to share how they're more "token efficient" than the base model they used. Apparently this -6.5% difference is worth highlighting as the very top feature of this fine-tune they've made.
How much is 6.5% percent concretely though? Assuming the six benchmarks have ~30 tasks each (180 in total), and we're doing 120 tok/s lets say, time/task for GLM-5.3 Flash would be ~25 minutes and this "token efficient fine-tune" would arrive at ~23 minutes.
Again, a 6.5% difference is highlighted as the most impactful change for this fine-tune, but it's a marginal difference. They're definitely trying to mislead people into thinking this is something more than it is, otherwise this landing page would have been much more realistic and less self-patted.
This is not a general purpose model, it makes sense to only bench it on tasks it's supposed to be good at, which is to help out finance-related knowledge work.
Honesty and transparency are attractive.
Not going to be trying it just based on this observed behaviour.
If the creators of this model are reading, can we see some of these same benchmarks run against the Thomson model, which although it's capable, is considerably smaller in size and baked in knowledge than glm 5.3-flash.
They say it can create office documents (if I understand correctly), but how does it do that?
What's the thing one needs to plug this into to say "give me a slide deck with last quarter's sales figures"?
__
https://www.paperinstruments.com/feather
According to their Twitter video, it might be this thing?
https://radix-os.com
So there's a side pane and a main section and a text input.
Very..uh.. plain. I was envisioning some more specific UX flow, but I couldn't tell you how, exactly.
The 'main' section is HTML, so it can be any type of workspace you want. Our agent can read/write (with permissions, of course) and can execute code locally (again, with permissions).
The HTML pages we generate are backed by a schema and state per 'workspace', so they persist, can be shared between team members, and aren't just ephemeral read-only artifacts.
Our agent can do work on your machine for you, it has an integrated web browser that allows it to not only search and retrieve info but also perform web-based tasks, and it can package everything up into a nice little contained UI for you so you're not jumping between multiple apps / documents / tabs. That UI is tied back to whatever document / page you were working in -- so for instance our app can have a form that updates a web page, or updates a document on disk and vice versa.
The plainness is deliberate, the design of the workspaces is up to you -- it's your UI not mine and the agent will design everything how you like it. And as far as the sidebar/text input combination goes I can assure you I have spent a lot of time imagining what this UX should look like and frankly this combo just makes sense.
Also a note on the text input / agent. When you move from text conversations to dynamic HTML interfaces you're faced with a question: how do you relate 'conversations' with the agent to workspaces? A one-to-one conversation to workspace mapping? Many-to-one, one-to-many? Should conversations be time-based, where I just start a new one and have to provide context again and can optionally link back to old conversations if I want? This question is actually quite difficult, but my belief was that ultimately you want just one single global conversation with the agent, and you want to let it handle its own memory internally. We basically just treat memory like a codebase and let the agent decide what to do. This means when you type in the text input you can refer to anything you like: any workspace, any conversation, any webpage you're currently browsing, any folder on disk -- the agent gets the right context and can retrieve relevant details as it needs. And it works!
So sure, I'm not a designer no matter how hard I try -- there will be far better UX out there than I've designed. But I think the fundamental mechanism in Radix is more sophisticated than most other tools, including Cowork etc (no shade at all to anyone, I'm a massive fan of Claude Code).
Source: Building a "harness" for organizations with deep MS Office suite integration at https://github.com/EratoLab/erato
So that is mainly a chat window like that Hubspot widget?
I mean I guess that gets stuff done, but I was expecting some fresher ideas perhaps?
So you have two main modes of operations: One-shot "I have this input; give me this output" which people usually do via the web-app, and more interactive editing to documents (usually larger ones) that happens directly in the applications the users are already working in.
And yeah, that's what ultimately gets the job done. We've seen competing solutions dive much deeper into building their own native editors, but they always lack the fidelity required by power-users, and are usually not as acepted by the broader workforce that are entrenched in their ways (our typical end-user is 50+ in a enterprise-like organization structure), so we try to meet the organisations where they are.
What I did was create docker containers with pandoc, compilers (rust, go, python, javascript) and it can create all manner of reports.