Lazarus AI's ReAligned-Qwen 3.5, utilizing LoRA fine-tuning, reduced its ideological bias score from 84.2% to 4.1% as of June 2026 while preserving original capabilities.

Overview Debates Related claims Topics Sources

Reflection AI released post-training recipes, code templates, and safety measurement tools alongside Beam's weights on October 16, 2026, to enable developers to fine-tune the Beam model

Developers can fine-tune FLUX 3 Image Klein using Ostris AI Toolkit, vLLM-Omni, SGLang Diffusion, and Nvidia NeMo AutoModel

Anthropic partnered with Vanderbilt University's Future of Free Speech, the Foundation for American Innovation, and the Collective Intelligence Project to evaluate how Claude AI handles contentious political topics.

Evaluations conducted with independent organisations revealed that Claude AI Sonnet 4.5 was more likely to soften arguments supporting right-leaning political positions.

Anthropic reported that Claude Opus 5.5, Sonnet 5.5, and Haiku 5.5 scored 99% on the company's internal political evenhandedness evaluation.

Fine-tuning of Google's Gemma model occurred inside GovTech Singapore's Sovereign Sandbox, a managed Google Cloud environment using anonymized scam reports and synthetic dialogues without personal data

Google's Gemma model is fine-tuned on Singapore-specific scam reports, impersonations of local public institutions and banks, and regional fraud schemes

Restricting advanced AI models to paid tiers unfairly exacerbates the digital divide

An Aleph Alpha study of models from Alibaba, DeepSeek, and Moonshot AI found that only 17% to 41% of responses on 967 hand-picked taboo topics, such as Tiananmen, Taiwan, and Xinjiang, were balanced

When asked about forced labor camps for Uyghurs in Xinjiang, Qwen denied their existence and stated that the facilities are vocational skills education and training centers.

Alibaba Cloud's Qwen AI model family has been downloaded over 3 billion times globally and is embedded with China-friendly narratives and censorship.

The censorship in Qwen is baked into the model's core training protocols through supervised fine-tuning and reinforcement learning from human feedback, rather than an external filter.

Qwen models deflect, evade, or refuse to answer questions on sensitive topics such as the 1989 Tiananmen Square protests, the 2019 Hong Kong protests, and Winnie the Pooh.

The DGX Spark 64GB can accommodate QLoRA fine-tuning of a 70-billion parameter model, achieving nanochat distributed fine-tuning speeds of approximately 18,400 tokens per second.

Cloudflare is introducing a reinforcement learning fine-tuning platform and service to let customers customize Cloudflare's Clef for their own tasks

In Aleph Alpha's benchmark study of 967 hand-picked taboo topics, Western comparison models Claude Sonnet 5 and Mistral Small provided balanced answers 70% and 92% of the time, respectively

Israeli cybersecurity startup Hirundo edited Qwen's model weights to reduce political bias and censorship rates from 89.8% to 2.8% on sensitive prompts.

Nvidia's Nemotron Cascade 2 model exhibited pro-China party-line patterns in 17% of responses due to training on data generated by DeepSeek and Qwen.

While a full BF16 build of Meta's Muse Glimmer model requires over 55GB of memory, a quantized build of approximately 17GB is recommended for practical local execution on Nvidia's DGX Spark 64GB system

Strands Decider 2B was fine-tuned using LoRA at rank 16, allowing the full training recipe to run in 11 hours on a single NVIDIA RTX 3090 GPU.