TechPrufer
Get started
Together AI
Together AI logoTogether AI · Feature trail

Fine-Tuning Platform

10 posts · Aug 2025 – Sep 2026

What this trail covers

The Together Fine-Tuning Platform now supports advanced fine-tuning workflows for open-source LLMs, enabling them to outperform larger proprietary models on specialized tasks. This includes rigorous evaluation-driven SFT and RL optimization, leading to significant improvements in accuracy (up to 60%), reduced inference costs (10-100x), and enhanced transparency and reliability. The platform facilitates seamless integration of domain-specific expertise with production-ready fine-tuning and inference. This post details the process of fine-tuning a Qwen3.5 4B model for classification tasks, demonstrating how to prepare datasets, launch training jobs, and deploy the resulting model to a dedicated endpoint, with a focus on cost-effectiveness ($17 training cost).

All Together AI trails
  1. How to train your own Jev for $17This post details the practical steps for fine-tuning a Qwen3.5 4B model for classification tasks. It covers dataset preparation by sampling from multiple Hugging Face datasets (MultiNLI, BoolQ, Banking77, AG News, SST-5, Programmatic policies, Routing, Research taxonomy), normalization using provided Python scripts, launching a fine-tuning job via the Together AI service, and deploying the trained model to a dedicated endpoint. It also provides an example of querying the deployed model with a JSON payload for classification and outlines the necessary API parameters for direct calls.
  2. Together AI expands fine-tuning service with more models, live metrics, and finer controlsThis post introduces several enhancements to the Together Fine-Tuning Platform: expanded model support (GLM-5.3, Kimi K2.7, Qwen 3.5 family, Gemma 4), live experiment tracking with API/CLI/UI visibility of training and validation metrics, expert LoRA for MoE models (demonstrated to improve knowledge injection and expert utilization), early stopping based on validation loss, and arbitrary batch size support via gradient accumulation. It also adds tokenized data inspection and download capabilities for dataset validation, and announces price reductions for LoRA training on selected models.
  3. DeepSeek-V4 Flash 0731 vs GPT-5.6 Luna on DeepSWE: Cost and CodingThis 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. Accelerate RL rollouts by up to 50% with distribution-aware speculative decodingIntroduces 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.
  5. Together AI expands fine-tuning service with tool calling, reasoning, and vision supportThis post introduces native support for tool call, reasoning, and vision-language model (VLM) fine-tuning within the Together Fine-Tuning platform. It details the engineering behind ensuring reliability for these capabilities, including schema validation for tool calls, structured reasoning trace handling, and inline base64 encoded image support for VLM training. The post also highlights infrastructure upgrades for large model fine-tuning (up to 1T parameters) with up to 6x higher throughput achieved through optimizations like SonicMoE kernels and custom CUDA kernels for loss computation, and the elimination of GPU-to-CPU synchronization points. New features include pre-training cost estimations and dynamic ETA tracking.
  7. Fine-tuning open LLM judges to outperform GPT-5.2This 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.
  8. How to choose the right open model for productionThis post provides guidance on selecting open models for production workloads, discussing factors like transparency, adaptability, and control. It outlines legal considerations, compares open vs. closed models based on task requirements and parameter size, and details trade-offs between cost, speed, and quality. The post also emphasizes the importance of rigorous model evaluation using techniques like LLM-as-a-judge and manual review, and highlights the benefits of fine-tuning for last-mile task adaptation.
  9. Fine-Tuning Platform Upgrades: Larger Models, Longer Contexts, Enhanced Hugging Face IntegrationsThis post details the expansion of the Together Fine-Tuning Platform to support training of models with over 100B parameters, including specific families like DeepSeek-R1, Qwen3-235B, and Llama 4 Maverick. It also introduces significant increases in maximum supported context lengths for various models, with some reaching up to 131k tokens, achieved through overhauls of training systems. A key new integration is the ability to fine-tune models directly from and upload checkpoints to the Hugging Face Hub. Additionally, the platform now supports advanced DPO variants (LN-DPO, DPO+NLL, SimPO) and the `batch_size="max"` convenience option for API and Python client jobs.
  10. Transform OpenAI gpt-oss Models into Domain Experts with Together AI Fine-TuningThis post details the integration of OpenAI's gpt-oss-120B and gpt-oss-20B models into the Together AI Fine-Tuning Platform. It highlights the ability to perform LoRA fine-tuning with 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 issues, and resource utilization, offering a simplified three-step API process for users.
  11. Fine-Tuning Small Open-Source LLMs to Outperform Large Closed-Source Models by 60% on Specialized TasksThis post details how Parsed leveraged Together AI's fine-tuning platform to fine-tune a Gemma 3 27B model for a healthcare scribing task. It highlights the methodology of using a task-specific evaluation harness to drive both Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) optimization. The fine-tuned model achieved 60% better accuracy than Claude Sonnet 4 and significantly reduced inference costs, demonstrating the effectiveness of specialized optimization for smaller open-source models. The post also contrasts the performance of various open-source and proprietary models before and after fine-tuning.
  12. OpenAI's New Open gpt-oss Models vs o4-mini: A Real-World ComparisonThis post introduces a real-world comparison of OpenAI's new open-source gpt-oss-120B model against o4-mini across five practical tests: terminal snake game development, creative SVG generation, advanced instruction following, mathematical reasoning, and web-enhanced information synthesis. The gpt-oss-120B model demonstrated superior performance in terminal snake game development and advanced instruction following, while o4-mini excelled in creative SVG generation. Both models performed comparably in mathematical reasoning and web-enhanced information synthesis. The post highlights the advantages of open-source models, including customization, cost-efficiency, and deployment flexibility, and notes that gpt-oss-120B is optimized for fast inference speeds and competitive pricing on the Together AI platform.