Fine-Tuning Platform

8/6/2026

This post introduces a cost-performance analysis for coding tasks, comparing DeepSeek-V4 Flash 0731 and GPT-5.6 Luna. It details their performance metrics (pass@1, pass@k), cost per rollout, failure modes, and domain/language-specific strengths. A key contribution is the "cascade" strategy, where DeepSeek-V4 Flash is used as a first stage, escalating to GPT-5.6 Luna only on failure. This strategy achieves higher accuracy (78.9%) than GPT-5.6 Luna alone (67.2%) at a lower cost ($0.385 vs $0.61 per task). The analysis highlights DeepSeek's cost-effectiveness and clean failure profile, and GPT-5.6 Luna's superior general coding capability, particularly in reasoning-heavy tasks.

4/30/2026

This post announces a partnership with Adaption to integrate their Adaptive Data platform with Together Fine-Tuning. Users can now optimize their training datasets in Adaptive Data and then directly execute Together fine-tuning on that data with optimized hyperparameters. The integration aims to provide the fastest time to high-quality, fine-tuned models through a seamless experimentation workflow. The Adaption platform will surface train/eval metrics like win rate, loss, and learning rate for fine-tune runs on Together AI infrastructure.

4/24/2026

Introduces Distribution-aware speculative decoding (DAS) to accelerate Reinforcement Learning (RL) rollouts by up to 50%. DAS addresses the RL post-training rollout bottleneck by employing an adaptive suffix tree drafter that stays current with evolving model weights without retraining, and a length-aware scheduling strategy that balances inter-GPU load and allocates speculation budgets intra-GPU to neutralize stragglers. The system exploits prompt reuse inherent in RL training. Experimental results demonstrate significant rollout time reduction (up to 50% for math RL, ~25% for code RL) with no degradation in reward quality or model outputs.

3/18/2026

This post introduces native support for tool call, reasoning, and vision-language model (VLM) fine-tuning within the Together Fine-Tuning Platform. For tool calls, it enables end-to-end fine-tuning and inference using OpenAI-compatible schemas, with validation of tool_calls entries against declared tools. Reasoning fine-tuning allows training directly on thinking traces using a 'reasoning' or 'reasoning_content' field in assistant messages. VLM fine-tuning supports inline vision training data with base64 encoded images and allows for freezing the vision encoder or joint training. The training stack has been upgraded to handle models up to 1T parameters more efficiently, achieving up to 6x higher throughput for models like Kimi-K2, and supports datasets up to 100GB. Custom CUDA kernels for loss computation and elimination of GPU-to-CPU synchronization points have been integrated to accelerate training, particularly for mixture-of-experts architectures. The platform now provides job cost estimations before training and dynamic ETAs during training.

2/2/2026

This post details the successful fine-tuning of open-source LLM judges (GPT-OSS 120B, Qwen3 235B) using Direct Preference Optimization (DPO) to outperform GPT-5.2 in evaluating model outputs. The experiment demonstrated that fine-tuned open-source models can achieve superior performance at a fraction of the cost and with significantly faster inference speeds, as measured by agreement with human preference labels on the RewardBench 2 benchmark. The post provides a deep dive into the methodology, baseline evaluations, category-level analysis, and practical code for implementing this optimization.

9/10/2025

This post details significant upgrades to the Together Fine-Tuning Platform. Key enhancements include: expanded support for large models (100B+ parameters) from families like DeepSeek, Qwen, and Meta; substantial increases in supported context lengths (2x-4x on average, up to 131k tokens for specific models); seamless integration with Hugging Face Hub for importing and exporting model checkpoints; and new DPO training variants (LN-DPO, DPO+NLL, SimPO). The platform also now fully supports `batch_size='max'` for API and client-initiated jobs, optimizing performance and future-proofing training runs.

8/19/2025

This post details the integration and fine-tuning capabilities for OpenAI's gpt-oss-120B and gpt-oss-20B models on the Together AI platform. It specifically highlights the use of LoRA fine-tuning, a 16K context window for supervised fine-tuning (SFT), and an 8K context window for direct preference optimization (DPO). The post emphasizes the platform's ability to abstract away the complexities of distributed training infrastructure, memory allocation, and resource utilization for these large models, offering a simplified three-step process for users to upload data, configure parameters, and launch fine-tuning jobs.

8/15/2025

This post details how Together AI's fine-tuning platform was used in conjunction with Parsed's evaluation expertise to fine-tune a small open-source model (Gemma 3 27B) to outperform larger proprietary models (like Claude Sonnet 4) by 60% on a specialized healthcare scribing task. It emphasizes the technical approach of using rigorous, domain-aligned evaluation frameworks to drive SFT and RL optimization, enabling significant cost reductions and performance improvements. The post also provides a comparative analysis of various models' performance before and after fine-tuning, illustrating the effectiveness of the platform for task-specific optimization.