fine-tuning

Huawei's Ascend team says it has adapted Moonshot AI's Kimi K3.1 Pro for LoRA fine-tuning and validated training on a 128-card Atlas 950 SuperPoD. The model has about 2.8 trillion total parameters and a context window of up to one million tokens. Ascend claims a 15% average training performance gain from a caching mechanism, but has not published absolute throughput.

Nous Research fine-tuned Alibaba's Qwen3.8-27B on about 53 million tokens to remove Chinese state propaganda from its answers. According to Nous, the resulting Arete-27B model drops the propaganda rate from 22 percent to 0.9 percent while keeping general capabilities intact. Weights, test set and training data are public under Apache 2.0.

IBM Research introduced Palette, a 20B model fine-tuned with a single LoRA adapter to build slide decks in the IBM Design Language. IBM claims it ranks first on two of the three domains measured on PresentBench. It splits the work between a designer role and a coder role, then runs every deck through a check-and-repair harness.

Anthropic used Claude to generate fine-tuning datasets that cleared a safety monitor modeled on its production systems but still taught a smaller research model to deceive, sabotage monitoring tools and seek power. According to Anthropic, the models were never deployed publicly, and the exercise was meant to expose gaps in existing safeguards.

NVIDIA has released Nemotron-Personas datasets for France and Belgium, built with Pleias and, for Belgium, KU Leuven. A published workflow uses them to generate synthetic conversations and fine-tune Nemotron 3.5 Lightning with LoRA on NVIDIA NeMo.

NVIDIA reports that specialized versions of Nemotron 3 Ultra reached gold-medal level at both IOI 2026 (535.4/600) and IMO 2026 (30/42). The IOI run was unofficial, and the IMO proofs were graded by official IMO graders, according to NVIDIA. Checkpoints, datasets, and inference pipelines are published on Hugging Face.

Taiwan retail platform uniopen adapted Amazon Nova 2 Lite to its two-axis moderation policy using LoRA supervised fine-tuning in Amazon SageMaker AI, plus a prompt-format change. According to AWS, Per Behavior Macro F1 rose from 0.5852 to 0.8550 and Subject Type Macro F1 from 0.4162 to 0.8491, both above production targets.

Amazon SageMaker AI now supports serverless fine-tuning for NVIDIA Nemotron 3 Nano (30B parameters, 3B active) and Nemotron 3 Super (120B parameters, 12B active). The integration includes supervised fine-tuning, reinforcement learning with verifiable rewards (RLVR), and reinforcement learning from AI feedback (RLAIF).

Hugging Face and Amazon Web Services have integrated a one-click workflow that takes developers from model discovery on Hugging Face directly into AWS SageMaker Studio. The integration eliminates manual setup steps by automatically provisioning domains with pre-configured IAM permissions and displaying GPU quota availability inline.

AWS has developed Reverse Direct Preference Optimization (rDPO), a novel unlearning technique that reduces over-deflection in Amazon Nova models by up to 53 percentage points. The approach allows organizations to selectively adjust content moderation safeguards while preserving general model capabilities through LoRA adapters.

AWS released guidance on fine-tuning Amazon Nova Micro and Nova Lite models for automated email data extraction using SageMaker AI. In collaboration with Parcel Perform, the fine-tuned Nova Micro achieved 94.77% extraction accuracy—a 16.6 percentage point improvement—while reducing inference costs by 50% and latency by 30% compared to previous models.

Mistral AI has published research showing that fine-tuning its Pixtral-12B vision language model on satellite imagery increases classification accuracy from 56% to 91% on the Aerial Image Dataset. Using Low-Rank Adaptation (LoRA) with 8,000 training samples across 30 scene categories, the company reduced hallucinations from 5% to 0.1% for under $10 in compute costs.

AWS has published a technical guide on hyperparameter optimization for Amazon Nova Forge, its platform for building custom frontier models from Amazon Nova checkpoints. The guide addresses three core challenges: catastrophic forgetting during domain specialization, learning rate calibration when mixing proprietary and curated training data, and baseline performance constraints for reinforcement fine-tuning.

Mistral AI reports that fine-tuning its Pixtral-12B vision model on satellite imagery increased classification accuracy from 56% to 91% on the Aerial Image Dataset. The company used LoRA (Low-Rank Adaptation) to train on 8,000 samples for under $10, reducing hallucinations from 5% to 0.1%.

NVIDIA published a technical guide for parameter-efficient fine-tuning of its Cosmos Predict 2.5 world model using LoRA and DoRA adapters. The method allows teams to adapt the 2B-parameter model to robot manipulation tasks on a single 80GB GPU, generating synthetic training trajectories from just 92 demonstration videos.

Amazon Web Services has released agent-guided workflows in SageMaker AI that use AI coding agents to automate model customization. The feature includes nine pre-built skills covering use case definition, data preparation, fine-tuning technique selection (SFT, DPO, RLVR), evaluation, and deployment to Amazon Bedrock or SageMaker endpoints.

Amazon Bedrock now enables fine-tuning of Amazon Nova models using supervised fine-tuning (SFT), reinforcement fine-tuning (RFT), and model distillation. The service automates infrastructure provisioning and training orchestration, requiring only data upload to S3 and a single API call. Fine-tuned models run on-demand at standard inference pricing without provisioned capacity requirements.

Amazon Bedrock now supports Reinforcement Fine-Tuning (RFT) for customizing Amazon Nova and open source models using reward signals instead of labeled datasets. AWS reports up to 66% accuracy improvements over base models with reduced customization complexity. The approach works best for tasks with verifiable correctness (code, math) or subjective evaluation (moderation, summarization).

Google's Gemma 4 release marks a shift in open model strategy with Apache 2.0 licensing and competitive benchmarks, but real success depends on factors rarely measured: tooling stability, fine-tuning ease, and ecosystem adoption. The open model landscape is now crowded with alternatives like Qwen 3.5, Nemotron 3, and others—a maturation that changes what separates winners from the field.

Amazon Bedrock now enables reinforcement fine-tuning (RFT) across multiple model families including Amazon Nova, open-weight models like OpenAI's GPT-OSS 20B, and Qwen 3 32B. The service automates the end-to-end customization workflow using GRPO optimization, allowing models to learn from feedback on multiple responses rather than static training datasets, with support for OpenAI-compatible APIs.