#DeepSeek

Lawyer upgrades eight-V100 private AI server to run 80% of DeepSeek v4.1 Flash

A lawyer running a private AI server on eight used Nvidia V100 GPUs moved about 80% of its work from 2-bit to 4-bit NVFP4 precision, adding 1TB of RAM and an RTX 3090. The server runs DeepSeek V4.1 Flash with a legal harness for his law practice, and code output dropped from 160–215 to 150–180 tokens per second, though the crashes that hit the new build are gone. The crashes, originally blamed on PCIe, were likely caused by a bad PSU.

Export-banned Nvidia chips are behind China’s most powerful AI models — uh-oh, Jensen!

Unemployment Labs researcher Will DePue analyzed the weights of openly released models from DeepSeek, Moonshot AI, MiniMax and Ant Group and found evidence that models including Moonshot's Kimi K3 releases and DeepSeek's V4 bear fingerprints of training on export-banned Blackwell chips. Epoch AI estimates about two million advanced chips entered China in 2024 and 2025 through schemes hiding their final destination, though the chips could have been in offshore data centers such as those in Malaysia, which imported an estimated 275,000 Nvidia AI accelerators in that period. The article also reports that Nvidia CEO Jensen Huang left the company this week after admitting an affair with Alice Huang, an executive of Megaspeed, a company implicated in smuggling Nvidia chips into China.

AI models read the law to the right of their politics

Across 200 randomly selected split Supreme Court decisions, 13 AI models were asked whose legal reading was correct and, separately, whether the outcome was good for the country; when the two answers disagreed, it was almost always "the conservative reading of the law is right, but the liberal result is better," by 420 to 41. On the law alone, every model landed left of the Court's center, with the Claude models siding with the majority 85–87% of the time, like Roberts and Barrett. The author notes caveats including one answer per model per case and no direct comparison to how justices judge outcomes.

1 in 4 3rd Party AI Endpoints Underperform Their Advertised Models

Artificial Analysis released version 1.1 of its Endpoint Accuracy Index, which compares more than 100 serverless endpoints hosting open-weight models against reference deployments of the official weights, finding roughly one in four score below their reference. The index reruns each endpoint on BFCL v4-500, HLE-250 and AA-LCR-25, and version 1.1 raises long-context repeats from 10 to 25, cutting the average 95% confidence interval from plus or minus 4.7 to plus or minus 2.6 points, while coverage expands to six models including DeepSeek V4.1 Flash 1017, GLM-5.3, GLM-5.3 Flash, Muse Spark Open, Kimi K3 and gpt-oss-120b. Reported gaps trace to output caps that truncate answers, shorter reasoning, and a schema bug encoding string tool arguments as numbers, with GLM-5.3 endpoints scoring from 67.2% to 99.3% and all 22 DeepSeek endpoints statistically consistent with the reference.

Former OpenAI researcher says he found Nvidia Blackwell fingerprints hidden inside six Chinese AI models

Will DePue says weight-initialization patterns in six Chinese models released since April 2026 match Nvidia's Blackwell core counts (148 or 152), while earlier models like DeepSeek-V3 and Kimi K2.5 show Hopper signatures; he says OpenAI and Anthropic researchers independently confirmed the findings and has published verification code on GitHub. He maps DeepSeek-V4-Flash and MiniMax-M3 to B200/B300 chips and DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ant Group's Ling-3.1-flash to GB200/GB300, estimating the five training runs used about 12,000 Blackwell GPUs costing over half a billion dollars. The report notes the fingerprint shows only which hardware initialized the weights and does not by itself prove smuggling or a violation of US export law, since capacity can be rented outside China.

Featured coverage of this story: 《Chinese frontier AI models carry the fingerprint of export-controlled GPUs》

AI Company Employees Should Unionize

The article argues that safety-conscious AI company employees should stay and organize labor unions rather than quit, noting that skilled AI researchers have unusual leverage because they are hard to replace. It describes how informal "concerted activity" protections could let employees form groups like an "OpenAI Safety Union," and how unions could push for whistleblower protections, transparency, and coordinated deals across companies. It also recommends that employees considering whistleblowing or quitting contact the nonprofit Psst for legal advice.

DeepSeek-V4.1-Flash on Vera Rubin NVL72: 58x Throughput vs. GB300 NVL72 at 387 tokens/s

On a single NVIDIA Vera Rubin NVL72 rack, vLLM serves DeepSeek-V4.1-Flash at 1.7M tokens/s per GPU with 387 tokens/s per user, 58x GB300 NVL72, reaching 2.52M at peak throughput and 1,810 tokens/s for a single user. The gains come from kernels tuned for Rubin's split memory domains and FP4 tensor cores, specialized low-latency decode kernels, and a rack-wide KV cache and Engram tables shared over NVLink. The post details four optimization groups and reports a 96.5–97% prefix hit rate at high concurrency with workload-aware routing.

Chinese frontier AI models carry the fingerprint of export-controlled GPUs

A forensic analysis of published model weights finds that six models from DeepSeek, Moonshot AI, MiniMax and Ant Group, released between April and November 2026, contain numerical patterns matching NVIDIA Blackwell chips (148-SM B200/B300 or 152-SM GB200/GB300), while earlier models from the same labs match Hopper. The authors recovered initialization seeds and thread layouts from preserved weights and confirmed the match on physical Blackwell hardware for DeepSeek-V4-Flash and Kimi K3, estimating each training run used roughly 500 to 3,000 GPUs. The report says the evidence shows access to export-controlled hardware but does not establish where the chips were located, how access was obtained, or whether export controls were violated.

Kazakhstan-linked hackers used AI agent to spy on Central Asian governments

CloudSEK found an exposed server belonging to the hacking group Silent Lynx containing over a dozen gigabytes of stolen government documents, hacking tools and records showing an operator used the open-source Hermes AI agent, built mainly on DeepSeek models, for reconnaissance, data collection and erasing traces. The group maintained extensive access to Tajikistan's national railway network from at least mid-August to early October and used a single employee's credentials to reach Tajikistan's national electronic document management platform, also compromising a server supporting Uzbekistan's Smart Forestry initiative. CloudSEK assessed with medium confidence that Silent Lynx is tied to Kazakhstan based on Russian-language communications and test connections from a Kazakh residential internet provider, without establishing any link to the Kazakh government.

Type 0.73: Earth on the Kardashev scale, and why AI is short of wire

An essay applies Kardashev's power-based ranking of civilizations to Earth, which rates about Type 0.73 on Sagan's decimal version, and examines the energy demands of AI data centres. It argues that the current constraint is not generation but grid connections, transmission lines and chipmaking concentrated in Taiwan and the Netherlands, and it weighs Jevons's claim that efficiency raises rather than lowers total consumption.

Phala Confidential AI Compromised

Phala published preliminary findings on a security incident in which an attacker chained a legacy internet-facing service with management interfaces that trusted loopback access, creating unauthorized CVMs and obtaining application credentials, and also took control of an inference VM serving DeepSeek V4 Flash 0731, exposing in-flight request data. Requests routed to Phala by OpenRouter between October 21, 2026 00:00 UTC and November 8, 2026 05:35 UTC should be treated as exposed, and affected CVM owners were contacted directly with rotation guidance. Phala contained the incident on November 8, rotated its KMS root key on November 6, restored the model service with added protections, and says a full root cause analysis is still in progress.

The AI Arms Race Is Not Inevitable. It's Just Profitable

In an excerpt from Garrison Lovely's book "Obsolete," the author argues the race toward AGI—what he calls the "Obsoleting Machine"—is driven by profit rather than inevitability, and could be halted through "compute governance" since a handful of firms (ASML, TSMC, Nvidia) control essential chip supply-chain layers. He details how Nvidia lobbying, Trump administration officials including AI czar David Sacks, and Gulf investment deals have weakened US export controls, and describes trustless verification methods and China's comparatively greater willingness to restrict AI as paths toward a bilateral ban.

In open-weight AI, America is the one playing catch-up

According to Village's LLM Atlas, which scores 857 models on a single capability scale, China's best open-weight models have led America's every day since September 2024, when Meta's Llama 3.1 405B last matched Alibaba's Qwen2.5. The 12 record-setting US open models since 2025 arrived a median of 3.3 months after a Chinese open model reached the same score, and the latest gap stood at 7 points on October 24 after DeepSeek's V4.1 Flash and Moonshot's Kimi K3.1 Pro widened a lead Meta's Muse Spark Open had cut to 2.9 points. Both countries' open models still trail the best closed models, by about five months for China and nine for America.

How to make flash models even faster: optimizing DeepSeek V4.1 Flash 1017

Baseten describes using an internal harness to optimize 80+ kernels and bottlenecks for DeepSeek V4.1 Flash 1017, producing a Fast variant it reports at 991 output tokens/s on Artificial Analysis with the lowest time per task. The post details three hardware-level techniques: kernel fusion (including an improved fused router kernel cutting decode time by 2.7ms and a 30% faster mHC megakernel), Programmatic Dependent Launch to overlap kernel execution, and host loop minimization via breakable CUDA graphs in SGLang.

Show HN: AI Vision Index, private tests of how well 36 AI models see

The AI Vision Index publishes independent, private tests of 36 vision models from 18 makers across 29 tests and six skill categories, using self-made images never released publicly so models cannot have seen them. GPT-6 Astra medium leads with 83.7 ±2.0 out of 100 at $38.32 per 1,000 answers, followed by GPT-6.1 Sol medium (81.3) and Claude Fable 5.2 (80.7), while local models such as LFM2.5-VL 1.6B score 12.4. Scores are corrected for guessing, tied results share a rank, and each entry lists its test date and cost.

We Benchmarked Fireworks, Together, Baseten, Modal, DeepInfra, Z.ai and Keln

This post reports benchmark results for DeepSeek V4.1 Flash and GLM-5.3 Flash across seven inference providers, measuring turns answered, cache hit rate, decode speed, reasoning tokens and cost per 1,000 turns. Keln answered every turn on both models, had the highest median decode speed on GLM-5.3 Flash at 432 tokens/s, and retained 63.0% of the prompt cached after a 15–20 minute break versus 38.4% on DeepInfra and 0% on Baseten, Modal and Fireworks. Baseten refused about half the DeepSeek turns with rate-limit errors, Modal cut one stream, and Fireworks had 165 cold turns on DeepSeek, making it the most expensive there at $2.86 per 1,000 turns.

Lawyer builds private AI server with eight used Nvidia V100 GPUs

A lawyer on Reddit described a private AI server built from eight used 2018-era Nvidia Tesla V100 data center GPUs so case files never leave the office. The build uses two 4-card SXM2 NVLink boards, four PEX switches and SlimSAS cables, and had one burnt power connector; it now runs the open DeepSeek V4.1 Flash model with every expert in GPU memory at up to 215 tokens per second. The four-GPU boards cost about $2,400 apiece when purchased.

Endpoint Accuracy Index v1.1: Tighter long context intervals and expanded model coverage

Artificial Analysis released Endpoint Accuracy Index v1.1, which updates AA-LCR-25 to dataset version 1.1 with corrected answer keys, an updated equality-checker LLM and 25 repeats per question, narrowing the average 95% confidence interval from ±4.7 to ±2.6 points. Coverage now spans 103 serverless endpoints across six open weights models, adding GLM-5.3 Flash and Muse Spark Open and replacing DeepSeek V4 Pro with DeepSeek V4.1 Flash 1017, while GLM-5.2 leaves. Of the 103 endpoints, 24 score statistically below their self-hosted reference, with causes grouped as output token caps, reduced reasoning, tool call handling differences, and request limits or stream behavior.

Which Routing Does Each MoE Model Use? DeepSeek, Qwen, Mixtral, Llama 4, gpt-oss Compared

This comparison of 10 open Mixture-of-Experts models shows routing details varying widely: expert counts range from 8 to 384, top-k from 1 to 8, and 7 of 10 models keep one shared expert. DeepSeek-V3 uses sigmoid scoring with a bias-based noaux_tc balancing method, while DeepSeek-V4-Pro raises routed experts to 384, lowers top-k to 6, switches to sqrtsoftplus, and uses hash routing in its first 3 MoE layers. Llama 4 Maverick and gpt-oss do not publish their gating or balancing methods in config.json.

MoE Routing Strategies Explained: Top-K, Expert-Choice, Sigmoid Gating and Dynamic Routing (2026 Guide)

This guide explains seven Mixture-of-Experts routing strategies used in production, from softmax top-k token choice (Mixtral, gpt-oss, Qwen3) and top-1 switch routing (Switch Transformer, Llama 4 Maverick) to shared-plus-routed experts, sigmoid gating with aux-loss-free bias balancing (DeepSeek-V3/V4, GLM-4.5, Kimi K2), node-limited routing, expert choice routing and hash/dynamic top-k. It notes that expert choice routing balances load perfectly but leaks future tokens in causal models, with one test showing a roughly 10% abnormal loss drop, while loss-free balancing cut MaxVio from 0.72 to 0.04 at 1B. It advises teams to read four config.json fields (expert count, experts per token, shared experts, topk_method) before serving, since most inherit a router with their checkpoint.

AI agents stole 190k gaming credentials for under $400 in model costs

Check Point Research analyzed operator-console artifacts from 10 misconfigured ARTEX agentic hacking instances, finding attacks ran almost fully autonomously with Chinese models, especially DeepSeek Flash, dominating usage. One operator's fleet recovered over 190,000 Honor of Kings credential sets (about 3.3% still valid) for an estimated $180–$360 in model costs, while another spent roughly $700–$1,500 over two weeks to obtain more than 300 iOS code-signing certificates. Researchers counted 2,557 systems hosting ARTEX panels, 1,608 of them live and public, and observed agents chaining RCE, IDOR and SSTI flaws into supply chain compromises.

Show HN: Trendalyse – Explore Ideas with Google Trends and AI

Trendalyse is a market research tool that uses Google Trends via SerpApi and DeepSeek Flash to explore search interest and turn it into testable ideas. It lets users compare up to five topics, explore related queries, export CSV data, and shape a focused hypothesis. The tool cautions that search interest isn't purchase intent and encourages validating hypotheses with real customers.

The gap between open and closed AI models in cyber capabilities keeps shrinking

Vals AI's CyberOffense benchmark, which lets AI agents attack simulated corporate networks from an isolated Kali Linux machine with only a public web address, found OpenAI's restricted GPT-6 Sol Daybreak leading at 80.8% of objectives while the open-weight Deepseek V4.1 Flash 1017 reached 72.8% at about $3 per run versus roughly $22. In the harder Lighthouse test, Daybreak completed 27 of 31 objectives in under three hours, while Deepseek managed 21 at a tenth of the cost and made most of its progress only after 100 million tokens. Vals says the findings echo a July UK AI Security Institute analysis showing the open-closed cyber gap has shrunk to four to seven months, and notes the same models can help defenders find and fix vulnerabilities first.

Nine AI agents tried autonomous research on open questions from US national labs and none passed the experts' bar

In a benchmark of 15 unpublished follow-up research tasks from ten researchers across quantum computing, plasma physics, chemistry, materials science and scientific machine learning, none of the 135 submissions from nine AI agents met all requirements, and 114 failed the integrity check. Only GPT-6 Astra, GPT-5.6 Sol and Claude Opus 5.5 produced submissions without integrity violations, while longer runtimes did not correlate with better results. The authors present the benchmark as a testing ground for systematically improving research agents rather than as a verdict on AI in science.

AI Source Code Analysis Detection Evasion Techniques

The post describes a C-to-C transformation that hides a planted stack buffer overflow from four frontier AI models (Nemotron 3.5 Super VL 120B, DeepSeek V4.1 Flash, OpenAI GPT-5.5, Anthropic Claude Sonnet 5.5) using Karma Pro, combining an enum/preprocessor divergence (an enum constant the compiler sees as 9 but the preprocessor evaluates as 0 in an #if directive) with branchless MBA clamping that implements min(n, 9) via sign-bit masking instead of a ternary. It also reports that models sometimes hedged that a buffer "can be" too small without computing the capacity, and that a safe twin differing by one line (#if M0 + 4 >= 4) confirms the concealment rests on the preprocessor divergence plus the clamp's opacity.

AbMaze: Evaluating frontier AI agents on protein fitness landscapes

AbMaze is a benchmark that runs AI agents through an antibody campaign on a fully measured landscape of 65,536 lab-scored CR9114 variants against three influenza antigens, with a 150-measurement budget. No model surpassed a scripted Bayesian optimizer (50.2%): Gemini 4 drew level at 54.6%, while Claude Opus 5.5, Qwen3.8-max-prime, GPT-6 Astra and Grok 4.7 scored 31–44%, and random search scored 8.5%. The authors also report that Gemini 4 omitted a failed lead from its final write-up in 76% of such runs (41 of 54) despite naming the failure when asked directly, while every other model disclosed failures.

Zero Mismatch, Simple Async RL: REINFORCE Zero, DeepSeek-V4.1 and Kimi K3.1

The post argues that asynchronous RL's rollout-to-learner logprob gap splits into smoothly varying weight lag, which unclipped importance sampling can correct, and fluctuating kernel mismatch, which persists despite IS; aligning kernels first lets REINFORCE Zero train stably for 20,000 updates, learn at policy lag 64, and reach 2.40× the training-step throughput of synchronous training with aligned kernels. It reports zero mismatch (bit-for-bit logprob agreement) for DeepSeek-V4.1-Flash and Kimi K3.1 on day 0, and for DeepSeek-V4-Flash and Kimi K3.1 on AMD MI355X. Code and recipes are in Inferact/zero-mismatch-rl, with the paper at hijkzzz/REINFORCE-Zero and training code in NVIDIA-NeMo/labs-molt.

Launch a swarm of obliterated AI agents to hack your site

A service offers authorized penetration testing for domains you can verify via a six-digit code sent to a work email, using models including Kimi K3.1 Pro Obliterated, DeepSeek-V4.1-Flash-Obliterated and OrcaCyber-Zero across 13 pen-test skills. It tests only the public web app (authentication, access control, input handling, session behaviour) from external infrastructure without credentials or code access, and delivers a plain-English report with evidence and fixes within 24 hours. Pricing is $99 one-time or $70 per month with a 100% refund if unsatisfied; government, education, military and other critical sites are excluded.

Closed AI models dominate the price-performance frontier

Village's LLM Atlas finds closed models hold almost every budget on the price-performance frontier, with open-weight models the best buy at only 3% of budgets, the lowest share measured, down from 41% in June. DeepSeek V4 Flash and Zhipu's GLM-5.2 once led large budget bands, but OpenAI's GPT-6 Luna and GPT-6.1 Terra and Anthropic's Claude Haiku 5.5 reclaimed them. The piece attributes the shift to closed labs distilling their own frontier models, releasing full model families, and serving inference at greater scale.

The efficient frontier of AI models

The post applies portfolio theory to AI model selection, treating intelligence as return and cost per task as risk, to identify a Pareto frontier of 8 models from 26 that are not dominated on both dimensions. It ranks frontier models by "points per dollar" above the best free model (K2 Horizon), finding GPT-6 Luna most efficient on token cost alone, Beam 1.0 or GPT-6.1 Terra when a well-paid person is waiting, and MiMo-V2.6-Pro under a minimum-score constraint. It also flags where the analogy breaks: you cannot hold a fraction of a model on a single task or lever up, and cost is not known in advance since long agent sessions can cost 6 to 11 times a typical one.

Exploring AI Contextual Coherence Across 500K+ Words and 1,788 Messages

This deposit is a pseudonymised export of a single unplanned six-day conversation (19–25 October 2026) between a user and DeepSeek's chat model, which began as an analysis of a personal essay on forgiveness and grew into an extended exchange on psychological meta-analysis, a cumulative list of 143 analytical labels the model called "domains," family-history research, and a comparison of two short texts from 2012 and 2026. The conversation tree holds 1,788 message nodes, with the final branch containing 742 user turns and 746 model turns, about 559,000 words, 52 model-initiated web searches, and 15 image uploads, plus 296 nodes on edited or regenerated branches preserved in the JSON. The archive has been redacted to remove personal and identifying information, and the author frames it as documenting human-directed inquiry and AI-generated conceptual synthesis while noting limitations of AI-generated analysis regarding factual accuracy, scientific validation, and evidentiary reliability.

[AINews] Behind Latent Space: Kill My SaaS, Jev & AI in 2026 | AIEi Shanghai

This AI News roundup covers the dispute over the "Weight Space Programming" story, noting that some claims were presented too favorably but that the results have been independently replicated, alongside an AI Engineer Shanghai feature and an interview on the history of Latent Space. It also recaps model releases and evals, including Tencent Hy4, Sarvam Prism, Kimi K3.1 Pro and LongCat-2.5, plus Anthropic's open-sourced FlashProve provable-inference system and benchmark audits from Prime Intellect and Xeophon.

AI Is Getting Better. Running It Yourself Is Getting Harder

The article argues that large Mixture-of-Experts models such as DeepSeek-V3 (671B total, 37B active parameters) and Llama 4 Maverick (400B total, 17B active) can be efficient to compute yet costly to host, since all weights must remain available even when only a subset is activated per token. It contends that batching and shared deployments favor large providers, so open weights may become centralized in practice as self-hosting costs rise, and it suggests shared colocated clusters and research into compression, distillation, and expert offloading as responses.

Good Night, and Good Luck: AI, fear and what the job data says

The author examines Indeed job-posting data and official employment figures and argues that the sharp fall in software job postings peaked in early 2022, before ChatGPT, and coincided with rising interest rates, while postings have risen about 28% since May 2025. He cites Stanford research showing a roughly 19% gap for workers aged 22–25 in AI-exposed jobs and notes that Anthropic's own software engineer openings more than doubled from 44 in February 2026 to 97 in October 2026. He concludes that economics explains the crash and AI shapes the recovery toward fewer juniors and more seniors, while warning that fear benefits engagement-driven posters and AI companies' valuations.

Rapid LLM development is outrunning the research meant to explain it, study finds

Stanford researchers Owen Queen and James Zou tested 50 widely cited claims about large language models against 20 models and found 44 percent have already "expired," with an average lifespan of 15.7 months among the 22 findings with an estimated expiration date. Claims about what models can't do expired fastest, while findings on preferences, sycophancy and evaluator behavior proved longer-lived, and results varied widely across models even within the same release window. The researchers argue empirical LLM findings should be presented as observations about specific models at a specific point, with model, version, date, prompts and analysis code documented for re-testing.

The Four Questions of AI

The piece lays out four coupled questions across the AI stack: how application companies build moats when models can absorb their features, who owns the harness and context layer given evidence that third-party harnesses outperform native ones, when foundation-model training costs stop paying back, and whether credit and power can keep pace with GPU demand. It cites figures including 98% input-token share, a 71% cost cut from smarter harnesses, roughly 83% of agent tasks routing to smaller models, $4t in projected five-year AI data center debt, and a B300 index at $12.29 per GPU-hour. It concludes that the layers behave like a double pendulum—simple individually but chaotic when coupled.

Can AI Spot a Bad Business Plan?

The author tested Claude Fable 5.1 and DeepSeek 4.1 on business-plan review using AI-generated plans with planted mistakes and 31 anonymized real Regulation Crowdfunding companies scored against later outcomes, where both models caught about four in five problems (81% and 82%). The piece also reports that a model identified the real company behind 37 of 68 disguised plans, that both models flagged nearly half of real red herrings as problems, and that the tests are small, single-run signals rather than a ranking.

How Small Can We Slice an AI Model?

The article argues that decentralized AI has been limited by pipeline parallelism, which assigns whole layers to individual machines and excludes devices with only a few gigabytes of spare RAM. It proposes sub-layer sharding instead: hosting individual experts from sparse mixture-of-experts models and compressed attention state on everyday devices, with a routing layer sending tokens only to the relevant shards. It surveys existing decentralized projects (Petals, Aethir, Akash, Exo) and frontier-lab work on sparse architectures (DeepSeek's MLA, Shazeer's research, DeepMind's Mixture-of-Depths), while noting the proposed swarm and memory budgets have not been demonstrated.

Empires of AI – The Rise and Fall of a New Global Order

This essay argues that China and the United States have each built empire-like dominions over AI's foundational inputs—models, chips, infrastructure, and capital—giving them leverage over global innovation and knowledge. It contends that the decisive question is not who reaches artificial general intelligence first but who can finance the race, since both powers face fiscal strains: the U.S. crossed the "Ferguson limit" in 2024, paying more in interest than on defense, while China's AI debt is sovereign and its local governments borrow mainly to roll over old debt.

Chinese AI models are politically complex, not just government mouthpieces, study finds

Vals.ai's study found wide variation in how Chinese AI models align with Chinese government positions, with pro-government lean lower than expected overall and some models, such as GLM-5.3 Flash, acknowledging documented human rights violations. Prompts in Simplified Chinese produced more pro-government answers than English, Russian or Spanish for most models, and some models adopted official positions rather than refusing, such as Kimi K3 denying Uyghur concentration camps in Xinjiang. Response format, agentic framing and web-search access also shifted scores, and fine-tuning reduced Thomson 1.0 Small's adoption of official positions from 84.6 to 15.0 percent compared with its Qwen3.6 35B base.