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.
> 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. For the small model, 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 quarter to half a million dollars of spend on that run, with an average deal that would be available without any extra connections. 100k minimum, if they got a great deal.
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.
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.
> 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?
> 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.
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.
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.
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.
> 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.
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.
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.
Here is some more technical information on how this was trained, as well as a download link.
https://huggingface.co/thomsonreuters/Thomson-1.0-Small
(Full disclosure I’m a TR employee, although I had nothing to do with making this)
Full technical report PDF: https://huggingface.co/spaces/tri-fair-lab/publications/blob...
> 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.
Benchmark overall score comparisons:
> Large: Thomson 1.0-Large - 78.5, Opus 4.8 - 79.5, Gemini 3.1 Pro - 78.0, GPT 5.4 - 76.5, Snowdon 1.0-Large - 73.6, Qwen3.5 397B - 73.0,
> Small: Thomson 1.0-Small - 74.6, Gemma4 31B - 71.2, Snowdon 1.1-Small - 71.7, Qwen3.6 35B - 71.7
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. For the small model, 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 quarter to half a million dollars of spend on that run, with an average deal that would be available without any extra connections. 100k minimum, if they got a great deal.
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
> 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.
From the HF link posted above it's Qwen3.6-35B-A3B.
https://huggingface.co/thomsonreuters/Thomson-1.0-Small
Thanks! And from the Technical Report linked in there it looks like their Large unreleased model is a trained Qwen3.5-397B
Per Business Insider [1], it was built on Qwen.
[1] https://www.businessinsider.com/thomson-reuters-builds-ai-mo...
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.
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.
> 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?
It’s a cost-saving measure and a marketing story.
> 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.
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.
Has anyone found a link to the technical report? They don’t seem very keen to publicise their performance on evals…
This is it: https://us.aws.cdn.hf.co/xet-bridge-us/6a899b56f043f237307da...
It’s on hugging face, I suppose one could do those themselves?
https://huggingface.co/thomsonreuters/Thomson-1.0-Small
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.
Reuters, the news agency, is only roughly 10% of TR's revenue. They're still a pretty big player in Legal and Tax.
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.”
https://ir.thomsonreuters.com/news-releases/news-release-det...
I doubt the point is to "make the money back".
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.
> 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.
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.