Join the discussion
Write your take first — we'll ask for email only when you're ready to publish.
- Hacker News
- 40m sounds quite heavy for a qwen fine tune?by Havoc
- Has anyone found a link to the technical report? They don’t seem very keen to publicise their performance on evals…by dash2
- It’s on hugging face, I suppose one could do those themselves?by johnnypangs
- by jore
- > 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.by JSR_FDED
- > 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?
by elpakal - It’s a cost-saving measure and a marketing story.by granzymes
- 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.
by x313 - 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 nicheby dgellow
- 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.
- 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.
by jll29 - > 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.
by hypfer - 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
by scirob - > 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.
by Arcuru - Per Business Insider [1], it was built on Qwen.
[1] https://www.businessinsider.com/thomson-reuters-builds-ai-mo...
by EnderWT - From the HF link posted above it's Qwen3.6-35B-A3B.by postalcoder
- 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)
by johnnypangs - 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
by hypfer - Thanks, that’s great, lots of detailsby dgellow
- 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.
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.
by helloplanets - 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.
by cootsnuck - 40M$ to get a marginally better model is surprising, why not just use the free weight modelsby make3
- 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"
by underlines - 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 detailsby dgellow
- 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.by walrus01