Rendered at 17:34:21 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
underlines 9 hours ago [-]
Please challenge me and explain why this is a break through worth the headline they use for their own work.
How I understand it, without really reading into it:
- Thomson Reuters did not "create" a frontier model, they gave some money, maybe a bit of their data to Imperial College London, and took Alibaba's Qwen 3.6 35B A3B Model for a basic fine tune.
- "They" (some undergrads at Imperial College London) used an existing ablation framework to undo some of the topic alignment of the original model.
- "They" fine tuned on some domain knowledge trying to preserve general knowledge - here maybe, just maybe some data came from Thomson Reuters.
As a result: one of 100's of Qwen 3.6 35B A3B sparse model fine tunes, just for publicity, to write overstating headlines like "X created their own frontier model"
walrus01 8 hours ago [-]
That's pretty wild that a Ctrl-F of "qwen" or "china" or "alibaba" on the thomson reuters announcement URL turns up no results. They want to portray it like they developed this thing from scratch, when that is by no means what actually happened.
dgellow 7 hours ago [-]
[dead]
dgellow 7 hours ago [-]
They didn’t say they created it no? I don’t think any of those points matter. It’s a business announcing a new product. Why does it matter if they built it internally or used external teams? I find strange to focus on those details
8 hours ago [-]
cootsnuck 14 hours ago [-]
This is going to increasingly happen over the years to come. Big organizations will become more sophisticated with operationalizing their data, training and running LLMs will continue to be demystified and accessible, and over time we'll get more and more specialized / industry-specific models.
It's going to become another way to monetize your informational assets if you're a big older enterprise with troves of data. All you need is time to figure out how to make it useful for yourself and then eventually sell access to it however you want.
Think of all the data that big orgs have that isn't accessible to all the AI labs to suck up.
make3 11 hours ago [-]
40M$ to get a marginally better model is surprising, why not just use the free weight models
jll29 9 hours ago [-]
Because "bigger = better" does not work for highly specialized expertise, where the knowledge is not available on the public Web. It is naive to believe it's all open out there merely because the Web is large and we have Wikipedia.
If you are a highly specialized professional intellectual property paralegal, a forensic tax investigator or a post-market pharmacovigilance analyst, you will need for-profit knowledge sources, and your answers will often require synthesizing and interpreting multiple sources.
simianwords 7 hours ago [-]
I’ll bet against this because of bitter lesson. Generalised models with context engineering will be more reliable and cheap than training your own.
adventured 11 hours ago [-]
They're trying to find a moat in the AI era, for their relatively gigantic news business (and they're among the few still standing giants in news). Most of these organizations are very scared of what AI might do to them.
bryanrasmussen 9 hours ago [-]
I'm pretty sure, given the reference to fiduciary grade standards that they are doing more to secure their moat in Legal, Accounting and related fields.
Reuters News is only slightly over 11% of their business. Legal and Accounting is closer to half.
on edit: just went and looked it up, adding in compliance offerings it is over 80% of their business.
make3 10 hours ago [-]
they should unite to block all news from getting to LLMs without LLM providers paying per access
dgellow 7 hours ago [-]
[dead]
johnnypangs 13 hours ago [-]
Here is some more technical information on how this was trained, as well as a download link.
> In this report, we argue that frontier performance can be achieved by a wide range of institutions through Continual Learning on readily available open-weight models.
> As opposed to existing limited approaches such as small-scale fine-tuning, prompt engineering, or tool-augmentation with a frozen model, our Continual Learning approach takes advantage of the effectiveness of a modern mid- & post-training stack while introducing safeguards preserving both plasticity and stability at each training stage and seeking to make the minimal number of high-impact interventions on the parameters.
For the large model, Thomson is utilizing the fine tuning stack they describe in the article, running it on Snowdon 1.0-Large, which in turn is a fine tune of Qwen3.5 397B. Same thing for the small model, but it's a fine tune of Snowdon 1.1-Small, which is a fine tune of Qwen3.6 35B.
As for the small version's run:
> The full pipeline consumed approximately 1.63 × 10²³ FLOP over 35,207 B200 GPU-hours, showing that these results are achievable with compute and personnel budgets substantially lower than commonly thought.
That would amount to around a quarter to half a million dollars of spend on that run. 100k minimum, if they got a great deal.
kennywinker 10 hours ago [-]
I wonder where the other $39.5 million went?
ustad 9 hours ago [-]
Men in the middle wages
12 hours ago [-]
dgellow 7 hours ago [-]
Thanks, that’s great, lots of details
hypfer 12 hours ago [-]
So it's a qwen fine-tune?
I mean that's a reasonable thing to do, but then the press release shouldn't be written the way it is written.
They're not as detached from the rest as the industry as the writing suggests.
__
> It is obtained by repurposing the open-weight Qwen3.6-35B-A3B model and substantially improving it on a wide range of performance domains.
nice wording on the HF page tho. "Repurposing". Lmao
make3 11 hours ago [-]
the press release says this explicitly
mkl 5 hours ago [-]
No it doesn't. There is no mention of Qwen at all.
hypfer 11 hours ago [-]
Yep. Explicitly enough to be legally safe for sure.
Which is the thing with press releases. Would've been nice to not do the bare legal minimum tho
dgellow 7 hours ago [-]
What isn’t clear in the press release? Reads fine to me
scirob 11 hours ago [-]
Maybe it's my dyslexic brain but "it's own frontier model" in my head converted to foundational model that was trained from scratch.
But this is qwen based.
But w/e I'm pro AI so more companies having more people with skills for more post training is cool
Arcuru 14 hours ago [-]
> starting from a strong open-source foundation and investing $40 million to train Thomson
Sounds like they spent $40 million finetuning an open weight model on their own data? I wonder what they built on.
postalcoder 13 hours ago [-]
From the HF link posted above it's Qwen3.6-35B-A3B.
Pretty cool someone is still doing this. Training in house LLMs was extremely popular in 2023-2024, back when domain-specific LLMs could easily top GPT in their field. In my field alone (tax/HR tech) I remember that Intuit, Workday, Indeed, LinkedIn were all training internal models.
It eventually stopped making sense because of inference costs. Running something internal with 30% GPU utilization is just too cost inefficient compared to using an API. Idk how Reuters will manage to solve this fundamental problem.
hypfer 11 hours ago [-]
> Running something internal with 30% GPU utilization is just too cost inefficient compared to using an API. Idk how Reuters will manage to solve this fundamental problem.
That's actually easy, because you can solve it through doing nothing and simply declaring that optimizing for lowest cost is not the main goal.
The fundamental-ness of that problem is entirely man-made and thus can easily be declared void as long as you have the cash to back that up.
Which might be a winning strategy in a world where everyone else is not doing that. Plus that your knowledge stays in-house, etc.
jll29 8 hours ago [-]
What makes sense also depends on one's business model: TRI charges premium dollars for access to their systems, so there is no need to optimize for cost; trust in the answers is the currency of knowledge workers in today's complex domains.
At the Thomson Reuters family of companies (technically then: Refinitiv Ltd. sold to LSEG), the first foundational model (in the sense of "trained entirely from scratch") was trained already in 2018 (i.e., pre-ChatGPT); it would even have been earlier, but the electricity wires and fuses in the rented 5 Canada Sq, Canary Wharf office had to be replaced first at the time to deal with the current needed to serve the GPUs.
dgellow 7 hours ago [-]
It’s based on qwen, not fully trained internally. I expect we will way more of this in the future, it’s pretty cheap to fine tune an open weight model for your specialized niche
porridgeraisin 12 hours ago [-]
Yep, this is the fundamental issue. It's a 35BA3B model and they probably finetuned it and evalled it in one bursty week on an 8xH100 rental just fine. But long term inference is always going to be easier in an API.
Unfortunately for reuters tho, they dont really have a choice. A lot of their data moat is not necessary live data as in linkedin, and the only way they can keep that moat is by doing this. I guess that justifies any cost.
jsrozner 10 hours ago [-]
I don't see why you couldn't see improvements in self-hosted or hosting-as-a-service model throughput? Basically API-style support for a company's internal LLM system. Why not? Or secure infra offered by AWS to self-host your own models that get served to the company just like any other company internal service can be hosted on AWS or similar?
It won't match Anthropic or OpenAI, but it could be economic?
hypfer 10 hours ago [-]
> Or secure infra offered by AWS to self-host your own models
There is no such thing as "secure infra" hosted by someone else.
This might still be fine, depending on your threat model, of course, but if your weights absolutely must never leave the confines of your org, you cannot use any shared hosting provider, because they just offer legal coverage of incidents. But if your moat is your knowledge, legal doesn't matter as much as the knowledge being suddenly unmoated.
porridgeraisin 9 hours ago [-]
It's possible, but it depends a lot on the nature of the data sovereignty guarantees. Today for many companies, the kind of guarantees they have with AWS amounts to basically legal coverage. And that is fine when the company is itself protecting data only due to legal or regulatory reasons. But if the company is protecting it for commercial reasons... then a different model is needed. It could happen, but it has to be worked out.
elpakal 14 hours ago [-]
> Our evaluation found Thomson’s citation quality generally competitive with leading frontier models, even when tested on Canadian employment-law questions without a Canada-specific setting.
That’s it? It was generally competitive with leading frontier models? Neat, but why would someone pay for frontier models and also a generally competitive additional product?
granzymes 13 hours ago [-]
It’s a cost-saving measure and a marketing story.
JSR_FDED 12 hours ago [-]
Cool that they did this on top of Qwen3.6-35B-A3B. If they have their own collection of valuable data this is the only way to make sure it doesn’t end up in general purpose models. That’s probably enough justification for the $40m spend - continued control of your destiny as an information provider.
peddling-brink 14 hours ago [-]
> Thomson Reuters is also making a “small” version of Thomson available as an open-weight model on Hugging Face for academic and non-commercial use to further aid in this validation.
Looking forward to the ERP fine-tune.
dash2 14 hours ago [-]
Has anyone found a link to the technical report? They don’t seem very keen to publicise their performance on evals…
I don't trust that they'll be able to make back that $40M.
This feels very much like a news agency getting into crypto or launching its own NFT line.
Or IBM selling Watson.
Or Mozilla chasing every which thing.
They're not stakeholders in the future of work. They're just wanting to stay relevant and pattern matching against what they see.
Reuters is too important for this.
If they were trying to use this as a narrative affront to OpenAI and Anthropic, maybe, but this is Reuters, not a deeply political organization seeking to land gotchas against big tech.
judge2020 14 hours ago [-]
Reuters, the news agency, is only roughly 10% of TR's revenue. They're still a pretty big player in Legal and Tax.
CircuitSeuss 13 hours ago [-]
Keep in mind line items can be deceptive. e.g. “Corporates” can mean “money we make from selling access to individual’s biometric data to the government.”
1. Marketing and expressing to their customers that they are not falling behind, and
2. Insulating themselves from frontier labs jacking up prices, nerfing the models they depend on, or otherwise unexpected changes in behavior.
I think the main goal is #2. Thomson Reuters might be a $40B company, but.... at this point it's not clear that that holds any weight in terms of not being fucked over by 2 companies aiming for $2t+ IPO valuations.
Edit: On second thought, there is probably a #3 too. They can serve inference for their own models significantly cheaper than frontier lab rates (assuming they're capturing continuous use of their hardware). I still think #2 is the primary goal.
blooalien 12 hours ago [-]
> I think the main goal is #2. Thomson Reuters might be a $40B company, but.... at this point it's not clear that that holds any weight in terms of not being fucked over by 2 companies aiming for $2t+ IPO valuations.
It's exactly the same sorta thinking re; Microsoft potentially fucking over the PC videogames industry that Valve used to justify the zillions of dollars and countless man-hours put into their big push for Linux gaming rather than tie themselves to a single proprietary company that could try to kick them out of the gaming industry. So far it's going pretty well for them. Depending on how they play their cards, this could also work out really well for Thomson Reuters as well.
hypfer 10 hours ago [-]
I doubt that 40 million are a lot of money for TR.
I also doubt that R&D spending having to "make back" directly is a winning strategy.
dgellow 7 hours ago [-]
Yeah, if R&D ROI was the main criteria to evaluate a company decision the AI labs wouldn’t be valued the way they are!
colechristensen 14 hours ago [-]
No it's just a value add to their existing data products and a moat against the big n LLM companies. The "news" part of the business is relatively small compared to everything else the do.
Like how Bloomberg does news but its far from their only or primary product. TR covers a different surface of data products than Bloomberg but it's a decent comparison.
maplethorpe 11 hours ago [-]
I know you're being facetious, but I genuinely think Thomson Reuters should be investing in NFTs as much as it is in AI. NFTs (while some of the shine has admittedly worn off) are an emerging infrastructure for digitally native ownership, and that's precisely the sort of institutional problem Thomson Reuters is positioned to solve (think tax records, medical records, etc).
People in tech suddenly tossing NFTs to the side because AI came along makes no sense to me. HN should be as bullish on NFTs now as it was in 2021. Jumping on to the AI train and acting like NFTs are bad now makes us seem flippant.
kennywinker 10 hours ago [-]
Tbf plenty of us were “acting” like NFTs were bad then too.
48484848n 11 hours ago [-]
[dead]
neontrashfence 11 hours ago [-]
Afaik... Lord Jacob Rothschild is 30% owner of Woodbridge, which holds Reuters. I believe the Thomson family owns the rest... The richest family in Canada. Woodbridge owns large shares of textbook companies, radio stations, news services, wires, scientific journals past and present.
How I understand it, without really reading into it:
- Thomson Reuters did not "create" a frontier model, they gave some money, maybe a bit of their data to Imperial College London, and took Alibaba's Qwen 3.6 35B A3B Model for a basic fine tune.
- "They" (some undergrads at Imperial College London) used an existing ablation framework to undo some of the topic alignment of the original model.
- "They" fine tuned on some domain knowledge trying to preserve general knowledge - here maybe, just maybe some data came from Thomson Reuters.
As a result: one of 100's of Qwen 3.6 35B A3B sparse model fine tunes, just for publicity, to write overstating headlines like "X created their own frontier model"
It's going to become another way to monetize your informational assets if you're a big older enterprise with troves of data. All you need is time to figure out how to make it useful for yourself and then eventually sell access to it however you want.
Think of all the data that big orgs have that isn't accessible to all the AI labs to suck up.
If you are a highly specialized professional intellectual property paralegal, a forensic tax investigator or a post-market pharmacovigilance analyst, you will need for-profit knowledge sources, and your answers will often require synthesizing and interpreting multiple sources.
Reuters News is only slightly over 11% of their business. Legal and Accounting is closer to half.
on edit: just went and looked it up, adding in compliance offerings it is over 80% of their business.
https://huggingface.co/thomsonreuters/Thomson-1.0-Small
(Full disclosure I’m a TR employee, although I had nothing to do with making this)
> In this report, we argue that frontier performance can be achieved by a wide range of institutions through Continual Learning on readily available open-weight models.
> As opposed to existing limited approaches such as small-scale fine-tuning, prompt engineering, or tool-augmentation with a frozen model, our Continual Learning approach takes advantage of the effectiveness of a modern mid- & post-training stack while introducing safeguards preserving both plasticity and stability at each training stage and seeking to make the minimal number of high-impact interventions on the parameters.
For the large model, Thomson is utilizing the fine tuning stack they describe in the article, running it on Snowdon 1.0-Large, which in turn is a fine tune of Qwen3.5 397B. Same thing for the small model, but it's a fine tune of Snowdon 1.1-Small, which is a fine tune of Qwen3.6 35B.
As for the small version's run:
> The full pipeline consumed approximately 1.63 × 10²³ FLOP over 35,207 B200 GPU-hours, showing that these results are achievable with compute and personnel budgets substantially lower than commonly thought.
That would amount to around a quarter to half a million dollars of spend on that run. 100k minimum, if they got a great deal.
I mean that's a reasonable thing to do, but then the press release shouldn't be written the way it is written.
They're not as detached from the rest as the industry as the writing suggests.
__
> It is obtained by repurposing the open-weight Qwen3.6-35B-A3B model and substantially improving it on a wide range of performance domains.
nice wording on the HF page tho. "Repurposing". Lmao
Which is the thing with press releases. Would've been nice to not do the bare legal minimum tho
But this is qwen based.
But w/e I'm pro AI so more companies having more people with skills for more post training is cool
Sounds like they spent $40 million finetuning an open weight model on their own data? I wonder what they built on.
https://huggingface.co/thomsonreuters/Thomson-1.0-Small
[1] https://www.businessinsider.com/thomson-reuters-builds-ai-mo...
It eventually stopped making sense because of inference costs. Running something internal with 30% GPU utilization is just too cost inefficient compared to using an API. Idk how Reuters will manage to solve this fundamental problem.
That's actually easy, because you can solve it through doing nothing and simply declaring that optimizing for lowest cost is not the main goal.
The fundamental-ness of that problem is entirely man-made and thus can easily be declared void as long as you have the cash to back that up.
Which might be a winning strategy in a world where everyone else is not doing that. Plus that your knowledge stays in-house, etc.
At the Thomson Reuters family of companies (technically then: Refinitiv Ltd. sold to LSEG), the first foundational model (in the sense of "trained entirely from scratch") was trained already in 2018 (i.e., pre-ChatGPT); it would even have been earlier, but the electricity wires and fuses in the rented 5 Canada Sq, Canary Wharf office had to be replaced first at the time to deal with the current needed to serve the GPUs.
Unfortunately for reuters tho, they dont really have a choice. A lot of their data moat is not necessary live data as in linkedin, and the only way they can keep that moat is by doing this. I guess that justifies any cost.
It won't match Anthropic or OpenAI, but it could be economic?
There is no such thing as "secure infra" hosted by someone else.
This might still be fine, depending on your threat model, of course, but if your weights absolutely must never leave the confines of your org, you cannot use any shared hosting provider, because they just offer legal coverage of incidents. But if your moat is your knowledge, legal doesn't matter as much as the knowledge being suddenly unmoated.
That’s it? It was generally competitive with leading frontier models? Neat, but why would someone pay for frontier models and also a generally competitive additional product?
Looking forward to the ERP fine-tune.
https://huggingface.co/thomsonreuters/Thomson-1.0-Small
This feels very much like a news agency getting into crypto or launching its own NFT line.
Or IBM selling Watson.
Or Mozilla chasing every which thing.
They're not stakeholders in the future of work. They're just wanting to stay relevant and pattern matching against what they see.
Reuters is too important for this.
If they were trying to use this as a narrative affront to OpenAI and Anthropic, maybe, but this is Reuters, not a deeply political organization seeking to land gotchas against big tech.
https://ir.thomsonreuters.com/news-releases/news-release-det...
It's likely split between two goals:
1. Marketing and expressing to their customers that they are not falling behind, and
2. Insulating themselves from frontier labs jacking up prices, nerfing the models they depend on, or otherwise unexpected changes in behavior.
I think the main goal is #2. Thomson Reuters might be a $40B company, but.... at this point it's not clear that that holds any weight in terms of not being fucked over by 2 companies aiming for $2t+ IPO valuations.
Edit: On second thought, there is probably a #3 too. They can serve inference for their own models significantly cheaper than frontier lab rates (assuming they're capturing continuous use of their hardware). I still think #2 is the primary goal.
It's exactly the same sorta thinking re; Microsoft potentially fucking over the PC videogames industry that Valve used to justify the zillions of dollars and countless man-hours put into their big push for Linux gaming rather than tie themselves to a single proprietary company that could try to kick them out of the gaming industry. So far it's going pretty well for them. Depending on how they play their cards, this could also work out really well for Thomson Reuters as well.
I also doubt that R&D spending having to "make back" directly is a winning strategy.
Like how Bloomberg does news but its far from their only or primary product. TR covers a different surface of data products than Bloomberg but it's a decent comparison.
People in tech suddenly tossing NFTs to the side because AI came along makes no sense to me. HN should be as bullish on NFTs now as it was in 2021. Jumping on to the AI train and acting like NFTs are bad now makes us seem flippant.