Field Notes

synthesis

EU-finetuned Chinese open-weight models

StatusWorkingReviewdue UpdatedStands on9 sourcesSupportsnotes

Open-weight Chinese models reached near-frontier quality in mid-2026, the United States is moving to restrict them at the software layer, and no credible European lab has matched them on capability. That combination is real and it does describe a market gap. It is also no longer an unnoticed one: by July 2026, a funded competitor, Venice, has already proven the adjacent business model works at scale, and a second startup, Eustella, is already building close to the user’s specific proposal. The opportunity is narrower and more contested than “nobody is doing this.”

The user’s proposal

License or adopt a leading Chinese open-weight model, finetune it to remove Chinese Communist Party censorship and to better serve European languages and cultural context, and serve it from European data centres, optionally with privacy features comparable to Venice. The competitive logic: American models carry an American cultural bias, Chinese models carry CCP-aligned censorship, and Mistral is the only serious European alternative but is not yet competitive on capability.

The export-control premise is current and accelerating

The premise is well-grounded and moving fast as of July 2026. The Trump administration scrapped the Biden administration’s broad AI diffusion licensing framework in mid-2025, but reversed toward tighter controls through 2026: a January 2026 rule moved chip exports to China and Macau from presumptive denial to case-by-case review, conditioned on know-your-customer and remote-access safeguards, and in June 2026 the administration confirmed that chip export bans apply to Chinese firms even when operating outside China.1

The proximate trigger for restricting models themselves, not just chips, was Moonshot AI’s Kimi K3, a 2.8-trillion-parameter open-weight model released in mid-July 2026 that beat several current-generation American frontier models on coding and agentic benchmarks, becoming the largest open-weight model released to date.2 Its release directly catalyzed the policy discussion the user is responding to: reporting from that same week describes the administration “actively considering restricting access to cutting-edge Chinese AI models within US borders,” extending containment from the chip layer to the software layer.3 This is a live, unresolved policy question, not a settled fact to plan around; the note’s review date reflects that volatility.

Chinese models are competitive, and their licenses now permit this

The capability claim holds up. Beyond Kimi K3, Alibaba’s Qwen and DeepSeek’s V3 and R1 lines are described as having “overtaken” specific benchmark categories against US proprietary models within roughly a year, particularly in coding and agentic tasks.4

The licensing landscape changed enough since 2023 to make the user’s proposal legally straightforward on the model-license side specifically. Qwen has used Apache 2.0 since version 1.5, with no monthly-active-user cap, revenue threshold, or notification requirement. DeepSeek has used MIT since V2, though it layers use-based restrictions, such as a prohibition on military use, that a derivative must carry forward. Kimi K2 uses a modified MIT license whose attribution clause only activates above 100 million monthly active users or $20 million in monthly revenue, which is permissive for a venture at typical launch scale.5 None of the three impose geography restrictions on the weights themselves; downstream legal exposure comes from the deploying jurisdiction’s own law, not the license.

Whether censorship can be finetuned away is not fully settled

This is the technical premise the proposal most needs verified, and the evidence is suggestive rather than conclusive.

A Khoury College study of DeepSeek-R1 found political refusals concentrated on Chinese-specific topics, Tiananmen Square, party-leadership criticism, and Taiwan, distinct from the safety-oriented refusals in US models. Independent testing by Promptfoo found DeepSeek refused about 85 percent of a curated set of politically sensitive prompts, with Tiananmen Square blocked 100 percent of the time.6 Separately, testing by Enkrypt AI found that even when researchers successfully extracted an answer, 91.2 percent of DeepSeek R1’s responses on China-related controversies still leaned pro-government, suggesting the bias is not only refusal behavior but also a leaning in the substance of answers that are given.7

Two findings bear directly on removability. First, both Khoury’s team and independent researchers found the refusal behavior can be circumvented at inference time, without any retraining, by manipulating DeepSeek’s visible reasoning trace with confidence cues such as “I know that…” or by priming the first few words of an answer. That the model “already knows the answer” but is tuned to withhold it is evidence the censorship sits in RLHF-stage alignment rather than in missing training data, which is the more favorable case for finetuning removal. Second, and more cautionary, a separate analysis found the censorship behavior “persists across languages and transfers to models distilled from it,” suggesting it is trained deeply enough into the weights to survive at least some downstream modification, not merely a thin instruction-tuned veneer removable with a light pass.8 Community “abliteration” projects targeting GLM-4 and DeepSeek derivatives report success at removing refusal behavior specifically,9 but public evidence does not yet establish whether abliteration also removes the subtler leaning the Enkrypt AI study measured, as opposed to only the explicit refusals. A venture built on this premise should budget for evaluation research, not assume a single finetuning pass settles the question, and should treat “removed refusals” and “removed leaning” as two separate claims requiring separate verification.

The American-bias half of the argument is also supported

Independent of the Chinese-censorship question, the claim that US models carry a measurable cultural default is well-supported by a growing body of 2025 to 2026 research. Stanford-covered research found perceived partisan bias across ChatGPT, Claude, and Gemini, generally leaning left of the median US voter on polarized topics.10 Separately, cross-cultural alignment research found that, absent explicit cultural prompting, GPT-family models’ values align most closely with Anglosphere and Protestant European countries specifically, consistent with the broader finding that English-dominant training data produces outputs aligned with Western, educated, industrialized, rich, and democratic populations rather than a neutral default.11 This supports the user’s framing: American default alignment is a real, measurable, non-neutral choice, not merely a rhetorical foil to justify preferring Chinese base models.

Mistral is not the ceiling the proposal needs it to be, but it is not weak either

The claim that “Mistral isn’t really competitive” is directionally correct but needs calibration. Mistral AI raised at a roughly €20 billion valuation in 2026, crossed $400 million in annualized revenue in January 2026 on a trajectory toward $1 billion by year end, and remains the sixth-ranked model provider by EU traffic share in both 2025 and 2026, unchanged.12 On specific hard benchmarks, Mistral’s largest model is reported not competitive with Qwen’s dense 27B model or DeepSeek V4 Pro on agentic coding evaluation.13 Mistral is commercially strong and technically behind the Chinese frontier on the exact tasks, coding and agentic performance, where Kimi K3 and DeepSeek currently lead. That is a narrower and more specific gap than “Mistral isn’t competitive,” and it is the gap the user’s proposal should name explicitly rather than treat Mistral as broadly weak.

The competitive field has already moved

This is the finding most likely to change the user’s plan. Two entities are already executing close variants of this proposal.

Venice is proof that the underlying business model works, not merely that it is theoretically viable. As of its July 2026 Series A, Venice is profitable, past $70 million in annualized revenue, and valued at $1 billion, built on hosting “uncensored” open-weight models including Chinese ones, adjusting system prompts to reduce refusals, and layering privacy and encryption features on top.14 Venice is US-based and global rather than EU-hosted or EU-specific, so it validates the model without occupying the specific European-sovereignty position the user describes.

Eustella is the closer and more direct competitor. It is a European startup explicitly building on Qwen, DeepSeek, Moonshot’s Kimi, and GLM, deployed on EU-hosted infrastructure, targeting the same “alternative to ChatGPT and Claude for 100 million-plus European users” positioning the user’s proposal implies. Public reporting confirms the strategy but not the technical depth of its adaptation work: eustella states it addresses model selection through “finetuning, evaluation, and combining several models,” and says it examines “bias, censorship behavior, and gaps in training data,” without disclosing which models it has actually deployed or the specific technique used to reduce censorship.15 This is a live, funded, EU-hosted, Chinese-model-based sovereignty-positioned competitor operating in the exact niche the user proposed, as of the same month as this research.

A third, larger and slower-moving entrant changes the multi-year picture further. The European Commission selected the Domyn-led EUROPA consortium in June 2026 to build a 400-billion-parameter, fully open-source frontier model covering all 24 official EU languages from the outset, backed by up to 2.5 percent of EuroHPC’s total compute for a year, with a stated one-year shipping target.16 This is not a finetuned Chinese model and does not address censorship removal, since it trains from scratch on European infrastructure, but it directly targets the same underserved-EU-languages gap the user cites as part of the opportunity, with public funding a private venture cannot match.

Regulatory exposure specific to this venture

The EU AI Act’s general-purpose AI model obligations apply differently depending on how much compute the finetuning consumes relative to the base model’s training run. If the finetuning uses less than one-third of the original training compute, the modifier’s obligations are limited to documenting the modification itself, a copyright-compliance policy, and a training-content summary for the finetuning data, rather than the full obligations of an original model provider. If the finetuning crosses that one-third compute threshold, the modifier becomes the provider of a new model in its own right, carrying the full GPAI obligation set, and if the base model is later classified as carrying systemic risk, that classification would carry forward too.17 Given that finetuning compute costs are now genuinely low, frequently under $20,000 for a complete evaluated finetuning project on a mid-sized open-weight model, a venture pursuing light instruction-tuning or abliteration-style censorship removal will likely stay under that compute threshold and face the lighter obligation set, but a venture pursuing deeper retraining to fix European-language quality should verify its actual compute ratio against the base model before assuming light-touch treatment.

Assessment

Claim in the proposal Verdict
US is imposing export controls and considering restricting Chinese models True and accelerating; Kimi K3’s July 2026 release is the direct trigger for the current restriction debate
Chinese open-weight models are catching up to or beating US frontier models True on specific benchmarks, especially coding and agentic tasks, as of Kimi K3
Chinese models carry CCP-aligned censorship True and measured; the mechanism looks partly RLHF-stage and partly deeper, so removal is plausible but unproven at the “leaning,” not just “refusal,” level
Chinese models “normally provide the weights” True, and licensing has become genuinely permissive since 2023, removing what was the biggest non-technical obstacle
US models carry American cultural bias True and independently measured in 2025 to 2026 research
Mistral is not competitive Partially true; Mistral is commercially strong but specifically behind on coding and agentic benchmarks against Qwen and DeepSeek
This EU gap is open and unaddressed False as stated; Venice validates the business model, eustella occupies the specific EU-sovereignty-plus-Chinese-models position, and the publicly funded EUROPA effort addresses the EU-languages gap on a longer horizon

Decision

The underlying thesis is sound, but the venture cannot be positioned as first-mover; it must be positioned against eustella and Venice specifically. The defensible differentiation is not “we finetune Chinese models for Europe,” which eustella is already doing publicly, but a combination eustella has not yet demonstrated:

  1. Verify eustella’s actual technical approach before finalizing a plan, since public reporting describes their strategy without disclosing their finetuning method or deployed models. A direct product comparison, not just a positioning comparison, should come before further commitment.
  2. Treat censorship removal as an evaluation problem first. Budget for the Enkrypt-style leaning evaluation, not only the Promptfoo-style refusal-rate evaluation, since current evidence suggests these are separable properties that likely need separable treatment.
  3. Use Confidential AI computing and Private AI trust boundaries to differentiate on verifiable trust architecture rather than model selection alone, following the same lesson Private AI competitors draws about the broader private-AI market: TEE-backed inference is no longer a unique feature, so the product must prove which trust boundary a customer chose, not just claim a friendlier model.
  4. Track the EU AI Act compute-ratio threshold as a design constraint on how deep the finetuning goes, not only as a compliance afterthought.
  5. Revisit this note when Kimi K3-triggered US policy resolves, since a formal US restriction on Chinese model access would strengthen the EU-sovereignty pitch considerably for both this venture and its direct competitors.

  1. BIS rule change and reporting on extraterritorial application document the January and June 2026 export-control shifts. ↩

  2. Coverage of Kimi K3’s release and benchmarks describes its scale, performance, and market reception. ↩

  3. Analysis of the US restriction debate connects Kimi K3’s release directly to the administration’s consideration of restricting Chinese model access within US borders. ↩

  4. Coverage of Chinese open-weight model progress describes benchmark gains through 2025 and 2026. ↩

  5. Licensing terms for Qwen, DeepSeek, and Kimi K2 are summarized from current model-license comparison coverage; verify the exact license text for the specific model version selected before deployment. ↩

  6. Khoury College’s DeepSeek-R1 censorship research documents the refusal patterns and inference-time bypass methods. ↩

  7. Enkrypt AI’s testing of DeepSeek R1’s substantive leaning on China-related controversies, as reported alongside the Khoury findings, found persistence of pro-government framing even in answers extracted past refusal. ↩

  8. Findings that censorship-like behavior “persists across languages and transfers to distilled models” indicate the alignment is not purely a surface-level, easily stripped instruction layer. ↩

  9. Community abliteration efforts targeting GLM-4 and DeepSeek-distilled models report successful refusal removal; public evidence does not yet cover whether the same technique addresses substantive leaning. ↩

  10. Stanford-covered research on perceived partisan bias in ChatGPT, Claude, and Gemini. ↩

  11. Cross-cultural LLM alignment research finding default alignment toward Anglosphere and Protestant European values absent explicit cultural prompting. ↩

  12. Coverage of Mistral’s 2026 valuation and revenue and EU traffic-share ranking. ↩

  13. Benchmark comparisons showing Mistral’s largest model behind Qwen’s dense 27B model and DeepSeek V4 Pro on agentic coding evaluation specifically. ↩

  14. TechCrunch’s coverage of Venice AI’s Series A documents its revenue, valuation, and open-model strategy. ↩

  15. Trending Topics’ coverage of eustella describes its Chinese-model-based, EU-hosted sovereignty strategy. ↩

  16. The European Commission’s announcement of the EUROPA consortium describes the 400-billion-parameter, 24-language, EuroHPC-backed project. ↩

  17. EU guidelines on general-purpose AI model provider obligations describe the compute-ratio threshold for when a downstream modifier becomes a model provider in its own right. ↩

Built on 9 sources (9 external).

Working out connections…