Fireworks AI Adds Qwen 3.5 Training to Build Custom Reasoning Agents

Apr 30, 2026

Fireworks AI, an inference platform for fast model serving and compound AI systems, added Qwen 3.5 to its training platform via Managed and Training API workflows. The update supports supervised fine-tuning (adapting a model to specific instructions) and reinforcement learning (training via feedback loops) while maintaining a 256K context window.

Context window: 256K tokens
Training methods: SFT, DPO, RL
Fine-tuning types: LoRA, Full-parameter
Access: Managed UI, Training API
Customization: Custom loss functions, smart defaults

This follows a rapid expansion of the Fireworks AI training platform. Matching Alibaba's Qwen 3.5 release, Fireworks enables teams to build proprietary reasoning models that rival closed-source systems without the drift of fragmented training and inference stacks, while also building on Fireworks AI's safe tokenization to secure model boundaries.

You can now run SFT, DPO, or RL jobs using smart defaults or custom loss functions. The platform supports both LoRA (efficient parameter-efficient tuning) and full-parameter fine-tuning for advanced tasks. These workflows are available now through the Fireworks dashboard or Training API for the Qwen 3.5 model family.

View the full update on fireworks.ai

Fireworks AI

@FireworksAI_HQApr 29

Qwen 3.5 from @Alibaba_Qwen is now available on @FireworksAI_HQ Training Platform across the Managed and Training API workflows. Try SFT, DPO, RL with smart defaults or your own custom loss function with a 256K context window. We support Lora as well as full param fine tuning for your most advanced tasks! What would you like to see next? https://t.co/rqSamw3I3e

127

View on X

Still wondering? A few quick answers below.

Fireworks AI supports several training methods for the Qwen 3.5 model family, including supervised fine-tuning, direct preference optimization, and reinforcement learning. Users can choose between using smart defaults provided by the platform or implementing their own custom loss functions to optimize the model for specific reasoning, coding, or mathematical tasks.

Yes, the Fireworks AI training platform supports both full-parameter fine-tuning and LoRA, which is a more memory-efficient method called Low-Rank Adaptation. Full-parameter tuning allows for deep customization of the model weights for advanced tasks, while LoRA provides a faster and less resource-intensive way to adapt the model to new data.

When training Qwen 3.5 on the Fireworks AI platform, users can utilize a context window of up to 256K tokens. This large window allows the model to process and learn from extensive datasets, such as long documents or complex codebases, without losing the ability to understand long-range dependencies during the fine-tuning or reinforcement learning process.

Qwen 3.5 training is available through two primary workflows on the Fireworks AI platform. Developers can use the Managed Training interface for a guided experience or integrate the Training API directly into their own development pipelines. These workflows are designed to help teams move from training to production inference on a single unified stack.