Same type of company, three years apart, opposite results.

In 2023, Bloomberg trained a 50B model from scratch, fed with 363B tokens of proprietary financial data. The official conclusion: it wasn’t used in any products.

In 2026, Thomson Reuters spent $40 million, using less than 10% of its 175-year proprietary content, performing continued pre-training plus expert post-training on an open base. In self-reported domain benchmarks, it outperformed several leading closed-source models. The pure compute cost: $100,000 to $200,000.

Three things changed in the intervening three years:

1. Open base models became strong enough; the price of APIs with equivalent capability dropped by about 100 times.
2. Post-training toolchains matured: full-parameter asynchronous RL, evaluation systems used directly as optimization signals.
3. Frontier pre-training evolved into an arms race in the $10 billion to $50 billion range, with no lab maintaining a healthy, independent profit statement.

Scarce resources flipped from compute to evaluation plus data flywheels. Harvey performs full-parameter RL on an open base, outperforming GPT-5.5 and Opus 4.8 Max on legal benchmarks.

But the premise is narrowing: Moonshot AI’s Kimi K3 with its Modified MIT license, MiniMax’s $20 million revenue threshold, Alibaba’s revenue-sharing scheme – the window for pure open-weight models is closing.

The full story, with 12 company cases, the watershed of success/failure, and a decision framework:

The Third Path for Domain Models: A 175-Year-Old Company’s $40M Answer

Source: https://yage.ai/share/domain-model-third-path-20260828.html?utm_source=twitter&utm_medium=thread&utm_campaign=domain-model-third-path-20260828 ← Directory (https://yage.ai/share/)EN (https://yage.ai/share/domain-model-third-path-en-20260828.html)Deep NewsSuperlinear Academy (https://superlinear.academy/)Industry & CompetitionChina Tech Ecosystem

August 24, 2026, Thomson Reuters released its proprietary large model, Thomson 1.0 (https://huggingface.co/thomsonreuters/Thomson-1.0-Small). This 175-year-old professional information publisher did not train a general base model from scratch to directly compete with OpenAI. Their core assets are concentrated in proprietary datasets like the Westlaw case law database, Practical Law, and Checkpoint tax library. Thomson 1.0, as a layer in a multi-model stack, integrates with the existing legal product CoCounsel Legal and, in self-reported domain benchmarks, surpasses several leading closed-source models.

The accounting behind this investment reveals a clear structure. Thomson Reuters states a cumulative investment of $40 million over two years, including all expenses for talent acquisition and acquiring research teams. The base model used is Alibaba’s open-weight model Qwen3.6-35B-A3B. According to Thomson Reuters, the team injected proprietary content for continued pre-training, actually using less than 10% of their total content assets; subsequently, hundreds of legal and tax experts participated in fine-tuning, preference alignment, and evaluation. The entire training consumed 35,207 B200 GPU hours, translating to a pure compute cost of roughly $100,000 to $200,000.

Self-reported evaluation data shows Thomson 1.0 achieved a domain average score of 75.2 in legal tasks, higher than the base Qwen3.6-35B-A3B’s 72.7, Gemma 4-31B’s 70.5, and Anthropic Haiku 4.5’s 67.7. Thomson Reuters has made a smaller version open-weight for academic and non-commercial use, while the full version is entirely controlled internally. The technical report is titled “Continual Learning of Frontier Models for SovereignAI,” emphasizing that the team built a continuous learning pipeline, where training the model is just the first step.

Together, these two points bring a key industry question to the forefront. Can a traditional company with vast industry-specific proprietary data, using an open base plus private engineering, truly forge its own model path? Is this an unreplicable one-off, or does it represent a common solution for vertical domains? To understand the logic of this path, we need to go back to the opposite experiment from three years ago.

Three Years Ago, Same Type of Company, Opposite Results

In March 2023, financial information giant Bloomberg published the BloombergGPT paper. The team then used a 50B parameter scale, fed in 363B tokens of proprietary financial data from the Bloomberg Terminal, and trained a base model from scratch. At that stage, the common practice for vertical giants was to build their own base: proprietary data couldn’t be directly input into external commercial APIs, open-source base capabilities were still weak, so pre-training from scratch was the only way out.

This enormous project ultimately met an unexpected end. Bloomberg Terminal’s global product lead, Wayne Barlow, gave a candid response in an interview with Market & Alt Data Insight (https://a-teaminsight.com/blog/inside-bloombergs-askb-roadmap-the-data-moat-behind-the-ai):

“BloombergGPT was a research model and is actually not used at all in any of our products. Our products are built on a combination of models. We have some commercial models, some open-weight LLMs, some smaller, custom language models that we build ourselves.”

Bloomberg’s current AI product, ASKB, began testing in early 2026, covering about a third of its 375,000 terminal users. Its underlying architecture has evolved into a hybrid routing solution, jointly supported by general commercial APIs, open-weight models, and self-developed lightweight custom models. The 50B large model that carried the company’s hopes remained in the research stage and never entered actual production.

Two industry giants with deep data assets, at points three years apart, reached opposite business outcomes. Bloomberg invested a high budget, consuming 363B tokens of proprietary content to start from scratch, yet ultimately failed to deploy it in any actual product; according to Thomson Reuters, with a $40 million investment, using less than 10% of its proprietary content for post-training injection on an open base, it surpassed top general models in self-reported domain benchmarks.

The same type of data-rich company: training a 50B model from scratch in 2023 with no product use, vs. in 2026, spending $40 million on post-training atop an open base and winning domain benchmarks against frontiers. I initially thought this divergence stemmed from data cleaning difficulty or team execution. Later, understanding the division of labor between RAG and model weights, I realized the root cause lies in the architectural choice. In 2023, the industry generally viewed general models with RAG versus proprietary model training as two competing, mutually exclusive paths. Facing this choice, Bloomberg bet everything on the weight layer, attempting to forcibly train massive financial facts into the 50B model. In 2023, this approach had its rationale, as general models’ financial understanding was insufficient and RAG technology was still in its early exploratory stages.

However, betting the entire moat on the weight layer quickly faces the冲击 of open models. Once open base models match capabilities, the first-mover advantage of training domain data into weights evaporates rapidly. Models lacking a proprietary data flywheel, evaluation loop, and industry distribution naturally cannot gain a foothold in production. Thomson Reuters’ successful path in 2026 offers a new answer: open bases handle general reasoning, RAG carries dynamic facts, and the proprietary data flywheel is dedicated to evaluation and behavioral alignment. The same problem, three years later, has a completely different solution, with the weight layer receding into a replaceable basic component.

Why Now

From 2023 to 2026, three core dimensions of the entire technology stack have changed significantly. Shifts in base capability, post-training toolchain maturity, and capital barriers have collectively reshaped the industry landscape.

The first change is that open base models have crossed the usability threshold. In 2023, the market lacked high-quality open-weight bases; vertical companies building custom intelligence had to pre-train from scratch. By 2026, open models like Qwen 3.8 Flash Next and GLM 5.3 Flash have tightly followed frontier capabilities in domain understanding. According to Artificial Analysis tests, GLM-5.2 scored 51 points, while Kimi K3 with 2.8T parameters reached 57 points, only 3 points behind top closed-source bases.

Meanwhile, the price of compute rental and API calls has plummeted. In March 2023, at the launch of GPT-4, calling a million input tokens cost about $30; by the end of 2025, the price for models of equivalent capability dropped to around $0.30, an overall reduction of nearly 100 times. The marginal cost of renting general capabilities has fallen sharply, making it financially unjustifiable to spend huge sums pre-training a general base from scratch.

The second change is the maturation and standardization of the post-training toolchain. Asynchronous reinforcement learning for full parameters, preference alignment algorithms based on corporate value lists, and the method of converting domain expert evaluation systems into continuous optimization signals were still in a lab-tinkering stage in 2023. By 2026, this methodology has evolved into a highly complete infrastructure, allowing vertical teams to rapidly layer high-density professional judgment onto open bases.

The third change comes from a fundamental economic shift in frontier general bases. Pre-training general large models has evolved into an arms race in heavy industry that ordinary product companies cannot enter. According to confidential financial data obtained by The Wall Street Journal in April 2026 and reported in a SaaS industry observation (https://www.saastr.com/anthropic-just-passed-openai-in-revenue-while-spending-4x-less-to-train-their-models), the combined model training and operational expenditure of OpenAI and Anthropic alone approached $65 billion in 2026.

OpenAI completed a $122 billion funding round in April 2026, with a post-money valuation of $852 billion and monthly revenue of $2 billion. Even so, its estimated single-year compute budget for 2028 is $121 billion, with losses of $85 billion, pushing the breakeven point beyond 2030. Training GPT-4 in 2024 cost about $78-100 million; by 2026, the pre-training budget for a single frontier base model has ballooned to $10 billion to $50 billion. Currently, no frontier lab can maintain a healthy, independent profit statement.

The competitive entities in the frontier arms race are also transforming into trillion-dollar behemoths. In February 2026, SpaceX acquired xAI, developer of Grok, in an all-stock deal; the combined entity was valued at $1.25 trillion. SpaceX subsequently completed its IPO on NASDAQ (https://www.cnbc.com/2026/06/12/spacex-stock-jumps-2-trillion.html) in June, offering shares at $135 each to raise approximately $75 billion, with a first-day market capitalization exceeding $2 trillion. NVIDIA is simultaneously the largest shareholder in the funding rounds of OpenAI, Anthropic, and xAI. OpenAI’s latest funding terms even stipulate that $35 billion of the funds are contingent on the company achieving artificial general intelligence or completing an IPO.

In comparison, the total budget for targeted post-training in vertical domains only needs tens of millions of dollars, with pure compute costs compressed to hundreds of thousands. The capital scale and risk profile are no longer in the same order of magnitude.

Harvey’s Three Phases

If Thomson Reuters provides a static enterprise blueprint, then the legal AI unicorn Harvey demonstrates a highly representative dynamic evolution trajectory. This company, focused on lawyer professional workflows, has undergone two key technical restructurings over the past three years.

The first phase was from 2023 to 2024. Early on, Harvey was deeply tied to OpenAI, fine-tuning closed-source models on U.S. case law. Third-party case studies showed that participating practicing lawyers preferred this custom model 97% of the time in tested scenarios. As subsequent frontier general bases were upgraded, general model capabilities quickly caught up with fine-tuning results. Harvey decisively abandoned its custom fine-tuned model, shifting to a multi-model routing solution centered on general APIs.

The breakthrough of the second phase came in June 2026. Harvey shifted to collaborating with Applied Compute, implementing full-parameter, fully asynchronous reinforcement learning on the open-weight GLM 5.3 Flash model on the AC2 compute platform. Harvey reviewed this experiment in its official blog (https://www.harvey.ai/blog/training-a-legal-agent-with-applied-compute), citing the core reason for choosing GLM 5.3 Flash as having the most solid initial baseline among candidate open models.

The experiment produced self-reported breakthrough evaluation data: on a professional legal standards benchmark, the pass rate increased from 0.853 to 0.913, and the full-item compliance rate rose from 0.059 to 0.126. Its professional score set a new industry record, directly surpassing OpenAI’s GPT-5.5 xhigh and Anthropic’s flagship model, Opus 4.8 Max. This marked the first time a vertical post-training approach defeated the top closed-source models of the same period on a professional benchmark. In its official technical review, Harvey used two passages to reveal the core principles of post-training. The first states:

“post-training, harness optimization, and grader design cannot be treated as separate problems”

They then provided a further summary:

“the model only learns as well as the environment allows”

These two statements refocus optimization on the evaluation environment itself. The quality of the grader and the test set directly defines the capability ceiling the model can ultimately reach. If the evaluation environment itself cannot provide high-discrimination scoring signals, no amount of compute investment can create a gap.

Broadening the view to 12 representative enterprises in legal, financial, and medical fields, the industry exploration shows a clear stratified landscape. 2 to 3 companies have clearly successfully traversed this path, including Thomson Reuters, Harvey’s second-generation solution, and the medical AI company Abridge, whose valuation surged 12 times in 12 months; 3 companies are in a hybrid transitional state, like Bloomberg shifting to small models and commercial API collaboration, C3 AI’s Narwhal, and Salesforce’s xLAM framework; 2 companies persist in purely renting external APIs, including the financial intelligence platform AlphaSense and Thomson Reuters’ old rival LexisNexis; 3 have encountered setbacks or scaled back strategies, including BloombergGPT (which abandoned production deployment entirely), Harvey’s first-generation fine-tuned model, and the retired Databricks DBRX.

Analyzing the gains and losses of these 12 companies reveals a clear watershed. Companies that stumbled often viewed training models as the technical identity and ultimate goal, either sinking into the quagmire of pre-training from scratch or relying on public corpora for superficial fine-tuning. Successful teams positioned self-developed models as an engineering tool to optimize product costs, invocation latency, and control, always embedding them within a composite system underpinned by frontier general models.

What Is Truly Scarce

Of the $40 million total budget self-reported by Thomson Reuters, pure compute spending accounted for only $100,000 to $200,000. This stark ratio clearly marks the true focus of vertical post-training: basic compute has become a highly standardized public good purchasable with capital at any time. What truly costs tens of millions of dollars and cannot be bought directly on the market is the 175-year accumulation of proprietary content assets (covering case law, practice guides, and tax data), the fine annotation and deep involvement of hundreds of senior domain experts, and a professional evaluation system with high discrimination.

Viewing RAG frameworks and post-training within the same system, the division of labor between them becomes very clear. “General model plus RAG” and “proprietary model post-training” are actually two complementary planes within the same hybrid architecture, responsible for different layers. The fact layer is dynamic, citable, deletable, and rollable, best kept outside the model and carried by RAG and external knowledge bases. The behavior and professional judgment layer is stable, high-frequency, and scorable, needing to be written into weights through mid-to-post training.

This division of labor directly points to where scarce resources actually lie. The weight layer itself has been commoditized by open bases; simply stuffing facts into weights is expensive and hard to maintain. Truly scarce assets are, on one hand, the dynamic facts and knowledge indexes that the weight layer cannot carry, and on the other, the proprietary data flywheel and evaluation system capable of continuously scoring behavior. What determines the system’s upper limit is precisely the ability to manage facts externally and refine behavior internally.

This change is also confirmed in the evolution of the application layer. In 2023, mainstream teams were accustomed to preparing hundreds of Q&A pairs to fine-tune a small model for specific tasks. Today, ordinary developers have fully shifted to in-context learning and agent workflows; fine-tuning is no longer a routine development option. Conversely, on the side of platform enterprises with deep data assets, continuous mid-to-post training on open bases is generating enormous momentum, driving Thomson Reuters and Harvey to achieve results surpassing general closed-source models in their self-reported professional evaluations.

The transmission chain of the post-training stack: open bases are purchasable, private data and evaluation systems are not, ultimately solidifying as domain product cost and control. While seeing the potential of this path, the negative signals disclosed in the technical reports also need careful consideration. Thomson Reuters’ self-reported data shows that after intensive legal domain injection, the model’s general capabilities experienced measurable regression. On the mathematical reasoning benchmark AIME, the score dropped from the base’s 93.3 to 90.0; the command-line environment benchmark Terminal-Bench fell from 45.2 to 40.5; programming test scores also decreased from 39.8 to 37.4. The notion of completely avoiding forgetting during domain injection does not hold up in real tests.

This capability trade-off precisely explains why Thomson Reuters must maintain a multi-model coexistence architecture design in the CoCounsel product. The vertically developed model tackles high-value professional logic, while the general frontier model provides fallback assurance in general reasoning and open tasks. Professional post-training comes with a clear general performance cost, making a multi-layered architectural hedge an essential engineering safeguard.

The Narrowing Premise

The ability of vertical companies to forge a third path rests on a critical assumption: that the open-source community can continuously provide high-quality, commercially unrestricted top base models. However, in 2026, this premise is undergoing substantive narrowing.

In July 2026, Moonshot AI released Kimi K3 (https://www.bleap.finance/en-us/blog/kimi-k3-review), adopting a Modified MIT license that explicitly requires commercial products with over 100 million monthly active users or over $20 million in annual revenue to sign separate licensing agreements. Shortly after, as recorded by industry media in the report on Alibaba shutting down the free tier of Qwen Code (https://decrypt.co/364501/alibaba-shuts-down-free-tier-qwen-code), MiniMax also adjusted its licensing model to a community commercial license with a $20 million annual revenue limit. Meanwhile, Alibaba has also begun implementing revenue-sharing mechanisms for large-scale commercial customers of its Qwen series.

The business strategies of domestic open-source vendors are collectively shifting. The phase of using permissive open source to exchange for a developer ecosystem is gradually ending. Achieving commercialization through controlling open weights has become the mainstream demand, and the window for pure, unreserved permissive licenses is closing rapidly. DeepSeek released the V4-Flash model weights under the standard MIT license at the end of July, becoming one of the few remaining pure open-source channels, while the final distribution method for its flagship V4-Pro weights still has numerous conflicting industry reports.

Reflecting on Thomson Reuters’ choice of Qwen and Harvey’s bet on GLM, third-party teams introducing open bases must place open-source license review at the forefront of project initiation. If future top open models generally come with commercial thresholds or revenue-sharing requirements, the entire financial model for the vertical post-training path will need recalculation.

Three Paths, One Decision Rule

Synthesizing the technological and commercial realities of 2026, enterprises face three clear evolutionary paths in model strategy. Companies of different scales and data reserves need to match entirely different combinations of elements.

The first path is developing a frontier base from scratch. This path requires massive capital of $10 billion to $50 billion, bearing the pressure of losses without an independent profit statement. The threshold includes hundreds of billions in external financing commitments, years-long high-intensity compute procurement contracts, and gigawatt-scale energy supplies. Players have shrunk to frontier labs deeply intertwined with capital giants.

The second path is purely renting general commercial interfaces. Under this logic, models degrade into plug-and-play runtime components; product value completely resides in application orchestration, interaction design, and workflow integration. This is the rational choice for the vast majority of ordinary software companies.

The third path involves investing tens of millions of dollars to build privatized mid-to-post training capabilities on high-quality open bases, serving exclusively industry giants with high-value data assets.

Hierarchy of the three paths: self-training frontiers are in the 10B-50B range with no independent profit statements; the third path operates in the ~$40M range serving data-rich domain companies; pure rental leaves value in the orchestration layer. Facing these three paths, the decision framework for enterprises is actually quite clear. For vertical enterprises that possess vast exclusive proprietary data, possess top-tier industry evaluation capabilities, and are deeply engaged in high-value, low-tolerance scenarios, the third path provides a powerful strategic defense lever. However, before starting, the team needs to self-examine against three rigid prerequisites: First, whether they have an automatically iterative continuous learning pipeline to ensure rapid retraining with each upstream base update; second, whether internal evaluation standards are rigorous enough to directly serve as optimization signals for reinforcement learning; third, whether a multi-model hybrid route has been established, with general frontier models providing fallback for complex long-tail tasks. For conventional software teams lacking proprietary data accumulation, blindly pursuing post-training will only repeat the mistakes of early fine-tuning.

The future pace of evolution depends on two key indicators. One is the pace at which open-source vendors tighten base weight licensing, which will determine whether the foundation of the third path can remain secure long-term. The other is the penetration speed of general frontier models into professional vertical scenarios. Once general bases achieve another generational leap in reasoning capabilities within legal or financial fields, the relative advantage window brought by vertical post-training will face new squeezes.