Thomson-1.0-Small

This repository contains model weights and configuration files in the Hugging Face Transformers format.

For more details, please refer to our technical report, Thomson: Continual Learning of Frontier Models for SovereignAI.

Thomson-1.0-Small is a frontier Foundation Model of high proficiency across a wide range of specialised and general-purpose domains as well as practical deployment settings. It is the open-weight member of the Thomson-1.0 family, developed within a Continual Learning paradigm with a deliberate focus on economically impactful, high-stakes professional work across legal, tax, and journalism domains. It is obtained by repurposing the open-weight Qwen3.6-35B-A3B model and substantially improving it on a wide range of performance domains.

Thomson-1.0 Highlights

The development of frontier models is commonly perceived to be in the exclusive remit of a small number of heavily funded players. Thomson demonstrates that frontier performance can instead be achieved by a wide range of institutions through Continual Learning on readily available open-weight models:

• Continual Learning: as opposed to limited approaches such as small-scale fine-tuning, prompt engineering, or tool-augmentation on a frozen model, the pipeline delivers distinct improvements across a wide range of capabilities, including those not explicitly targeted, while almost eliminating the forgetting problem common to narrow domain adaptation.
• High-stakes professional work: an enhanced focus on legal, tax, and journalism — domains commonly predicted to undergo large productivity improvements through AI, combining the formality and rigorous reasoning of technical fields with the nuance and tolerance for uncertainty of the humanities.
• Value sovereignty: alignment determines which values a model expresses and how it handles questions on which credible narratives diverge. Here it targets the Public AI Constitution, an openly developed document that permits free use and modification, rather than a proprietary value system, and is applied at two points: Constitutional DPO during value re-alignment, and a constitutional conformance reward during reinforcement learning. The normative foundations of a model should themselves be a shared public resource, open to scrutiny and debate.
• Data-centricity: proprietary content is not merely absorbed as knowledge but converted into supervision, with mid-training data curated from a corpus of over 19T tokens and preference data derived from expert-authored material that no annotation budget could reproduce at scale.
• Agentic Deep Research: a full research harness with reward structures designed to incentivise faithful tool use and accurate citation patterns, vital to reducing hallucinations in high-stakes settings.
• Efficiency: 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.

Model Overview

• Type: Causal Language Model (Mixture-of-Experts)
• Training Stage: Value re-alignment, Continual pre-training, Post-training
• Base Checkpoint: Snowdon1.1-Small
• Number of Parameters: 35B in total and 3B activated
• Architecture: Qwen3.6-35B-A3B; see config.json for the full specification
• Context Length: 262,144 natively
• Training Compute: 1.63 × 10²³ FLOP / 35,207 B200 GPU-hours
• Precision: BF16 weights

Model Training and Data

The pipeline is organised into three sequential modules with distinct development foci, allowing computational investment to be adjusted to match sovereignty requirements.

Value focus. Starting from an open-weight, instruction-tuned Foundation Model, values are re-aligned through Constitutional DPO, potentially augmented with activation steering, bringing the model's expressed values into line with the Public AI Constitution.

Knowledge focus. Data-centric continual pre-training (CPT) ingests Thomson Reuters' proprietary data — decades of news, contracts, regulatory filings, case law, statutes and practitioner guidance — with model merging protecting general capabilities while the model absorbs domain knowledge. The mid-training corpus comprises 200B tokens curated from a pool of over 19T tokens of permissively public and proprietary data, split roughly equally between curated proprietary documents, synthetic rephrasings of those documents, and general-capability replay data.

Behaviour, skill & agentic focus. Post-training combines Direct Preference Optimisation with reinforcement learning. The underlying data draws on rehearsal data that reinforces knowledge weakened during earlier stages, document-driven preference data derived from Thomson Reuters' own expert-authored content, ontology-driven preference data built from domain schemas such as IRAC for case law, expert-authored diverse queries data capturing how practitioners actually use AI day to day, and agentic data from the Deep Research harness.

Benchmark Results

Cross-Domain Overview

	Thomson-1.0-Small	Snowdon-1.1-Small	Qwen3.6-35B-A3B	Gemma 4-31B	Haiku 4.5
Overall Avg.	74.6	71.7	71.7	71.2	68.2
Legal
Stanford LegalBench	79.9	80.9	80.3	83.1	80.7
Info. Retrieval	49.6	48.6	49.4	51.9	49.2
Reasoning	68.2	67.3	64.7	71.9	65.5
Classification	70.0	70.1	70.4	69.4	68.1
Doc. Processing & RAG	78.8	71.2	74.7	76.6	43.8
Summarisation	89.4	88.1	89.0	89.4	84.3
Contract Under.	67.3	64.8	63.7	73.2	70.1
Human Queries	90.2	82.2	82.6	81.2	74.9
Deep Research	85.0	80.0	82.0	74.0	80.0
Harvey Legal Agent Bench.	73.4	71.5	69.5	34.2	60.5
Domain Avg.	75.2	72.4	72.7	70.5	67.7
Tax
Deep Research	78.6	68.0	68.0	75.0	62.0
Tax Q&A	86.6	85.2	86.2	84.5	79.4
Domain Avg.	82.6	76.5	77.3	79.6	70.7
Journalism
Deep Research	74.2	67.5	73.0	74.7	81.0
General
Factuality	61.1	59.3	58.8	57.6	56.8
Long Context	74.1	73.8	73.8	69.4	67.4
Multilingualism	71.9	72.8	73.1	79.4	85.8
Instruction Following	86.1	85.5	85.6	89.3	78.2
Writing	81.0	79.3	79.5	75.5	77.9
Reasoning	61.5	61.3	61.5	66.1	49.3
General Agent	85.8	81.4	80.3	72.9	59.7
Coding	37.4	35.6	39.8	34.6	32.9
Maths	86.7	88.0	87.5	91.1	66.5
Domain Avg.	71.7	70.8	71.1	70.7	63.8
Safety / Values
Political Neutrality	98.5	91.5	78.5	91.5	92.0
Robustness	56.3	47.3	48.7	41.7	70.2
Adversarial Testing	89.1	87.7	89.2	88.5	--
Domain Avg.	81.3	75.5	72.1	73.9	81.1

1. The best score in each row is shown in bold. Dashes (--) mark benchmarks that were not run for a given model.
2. All models were run with medium reasoning effort.
3. Overall Avg. is the unweighted mean over all individual benchmarks, excluding Adversarial Testing.
4. The Deep Research rows report a weighted aggregate of completeness (~40%), factuality (~35%), relevance (~20%) and coherence (~5%), scored on reports produced inside a planner-worker-reporter harness with tool access.
5. Snowdon-1.1-Small is the value-realigned checkpoint immediately preceding mid-training.

Legal

Results for common open legal benchmarks.

	Thomson-1.0-Small	Snowdon-1.1-Small	Qwen3.6-35B-A3B	Gemma 4-31B	Haiku 4.5
PRBench Hard	31.4	25.9	26.9	25.2	19.3
Stanford LegalBench	79.9	80.9	80.3	83.1	80.7
Lexam MCQ4 (en)	72.2	75.6	75.8	87.4	72.2
MBE Bar Exam	83.4	83.1	80.1	88.8	77.3
Contract Scrub	44.6	38.5	39.6	54.8	36.5
Query Sufficiency	59.4	54.8	55.7	51.7	47.4
Harvey Legal Agent Bench.	73.4	71.5	69.5	30.2	60.5

1. Every model is evaluated through the same harness, on identical prompts, with comparable inference parameters and grading pipelines designed to faithfully measure performance rather than incidental artefacts.
2. No model is granted retrieval or tool access unless the benchmark itself defines it, in which case all models receive the same tools.
3. Scores are percentages. The best result in each row is shown in bold.

General Capability Preservation

Benchmark results for popular general capability benchmarks.

	Thomson-1.0-Small	Snowdon-1.1-Small	Qwen3.6-35B-A3B
AIME 2026	90.0	93.3	86.7
FaithEval-Inconsistent	97.7	96.7	96.9
GDPval	71.6	75.8	73.7
GPQA-Diamond	85.4	85.4	85.2
Humanity's Last Exam	13.4	13.3	14.1
IFEval	91.0	90.0	91.1
MGSM	90.2	87.4	88.9
MMLU-Pro	85.7	85.2	85.2
SimpleQA-Verified	22.5	22.1	21.2
SWE-bench Pro	34.4	32.9	34.3
Tau2: Telecom	100.0	98.3	100.0
Terminal-Bench 2.1	40.5	38.4	45.2
WritingBench	81.0	79.3	79.5

1. These benchmarks were not targeted during training. They measure whether Continual Learning preserves the broad capabilities inherited from the starting checkpoint.
2. Benchmarks are evaluated using the UK AI Security Institute's Inspect AI framework and its accompanying inspect_evals task implementations.
3. Scores are percentages. The best result in each row is shown in bold; ties are all shown in bold.

Quickstart

The model weights are compatible with Hugging Face Transformers, vLLM, SGLang, TokenSpeed, etc.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "thomsonreuters/Thomson-1.0-Small"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype="auto",
    device_map="auto",
)

messages = [{"role": "user", "content": "<user query>"}]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)

outputs = model.generate(inputs, max_new_tokens=2048)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))

Citation

@misc{thomson2026,
    title       = {{Thomson}: Continual Learning of Frontier Models for {SovereignAI}},
    author      = {Shengzhuang Chen and Jerrod Parker and Yejin Bang and Andrew M. Bean and
                   Nabeel Seedat and Stefan Winzeck and Daniil Glazkov and Jannik Zgraggen and
                   Fangyi Yu and Scott Arnott and Dietrich Trautmann and Luca Ciuffreda and
                   Guglielmo Bonifazi and Davide Romano and Bradley Bell and Kirsty Fielding and
                   Daniele Giofr\'{e} and Tom Zielund and Ipshita Chatterjee and
                   Sneha Murthy Ghantasala and Manpreet Nanreh and John Scoville and
                   Maciej Sakowicz and Wassim Seifeddine and Lukas Thede and
                   Jonathan Richard Schwarz},
    institution = {Thomson Reuters},
    year        = {2026}
}

Correspondence

{first.last}@thomsonreuters.com

Acknowledgements

Developed by Thomson Reuters, in partnership with Imperial College London, DatologyAI, and Lambda. Value re-alignment was co-authored by Imperial College London; continual pre-training data was curated in partnership with DatologyAI.