Training Preview

Own Your AI: Fireworks Training Preview

GLM 5.2 is live! Opus-level intelligence at open-source rates. Pay per token on serverless. Try it today.

Blog

Training Preview Own Your AI: Fireworks Training Preview

PUBLISHED 4/6/2026

Table of Contents What Teams Have Built

RL: Agentic Behavior That Prompt Engineering Can't Reach

SFT: The Foundation That Beats Closed Models

Classification: Fine-Tuning Wins Across Every Use Case

DPO: Alignment Without Writing a Reward Function Full-Parameter Training at Frontier Scale Numerical Parity Between Training and Inference Three Surfaces, One Platform

Training Agent

Managed Training

Training API Get Started

Table of Contents

Fireworks Training is now in preview: an end-to-end platform for training and deploying frontier models at scale. Three surfaces for three kinds of teams, from a conversational agent that handles everything, to managed infrastructure for ML engineers, to bring-your-own training loop on Fireworks-hosted clusters. All on the same infrastructure that already handles production inference for Cursor, Vercel, Genspark, and others. What's new today:

• Full-parameter training at frontier scale, from Qwen3 8B to Kimi K2.5 at 1T parameters • Custom loss functions via the Training API • Multi-LoRA serving for teams iterating across multiple adapters • Training Agent: describe your task, deploy your model

All three surfaces are in preview now.

What Teams Have Built

RL: Agentic Behavior That Prompt Engineering Can't Reach

Reinforcement learning is how teams push past the ceiling SFT hits on multi-step reasoning, reliable tool use, and mid-flight self-correction. Vercel used our RL infrastructure to build a custom "Auto Fix" model for v0. The model checks the output stream for errors and self-corrects without a second pass, reaching a 93% error-free generation rate , significantly outperforming closed frontier models, with a 40X improvement in end-to-end latency vs. the proprietary model it replaced and over 8,000 characters per second throughput. "Using a fine-tuned reinforcement learning model with Fireworks, we perform substantially better than SOTA. In our evaluation, Sonnet 3.5 compiled at 62%, and we got our error-free generation rate well into the 90s." — Malte Ubl, CTO at Vercel

Genspark applied frontier RL to Kimi K2 , a 1 trillion parameter open-source model, for deep research agents requiring multi-source investigation and chained tool calls. The RL-trained model unlocked a 33% increase in tool calls , surpassing a state-of-the-art closed model at 50% lower cost . "Fireworks enabled us to own our AI journey, and unlock better quality in just four weeks." — Kay Zhu, CTO at Genspark

Cursor ran RL rollouts for Composer 2 (now top-scoring on CursorBench) across 3 to 4 clusters worldwide, with training and production traffic sharing the same GPU pool via delta-compressed weight updates. Frontier RL doesn't require one mega-cluster. The assumption that you need co-located RDMA hardware rests on moving a full 1 TB checkpoint on every update. You don't. "Our RL inference scales elastically and globally because of it. When we have low prod traffic we scale up RL, when we have high prod traffic, we scale down RL." — Federico Cassano, Research at Cursor SFT: The Foundation That Beats Closed Models

RL gets the headlines, but the truth is most teams start their fine-tuning journey with SFT and the gains are immediate. The most common pushback we hear is that closed models are "good enough." On a customer support dataset, a fine-tuned Qwen3 8B Instruct model hits an F1 score of 76.38% vs 69.40% on leading closed source model. In fact Qwen3 0.6B, 4B, and 30B all outperform the closed model on this benchmark, at significantly lower cost. On a production customer operations dataset, fine-tuned Qwen3 30B reaches 91.71% versus the closed model's 82.48%. Classification: Fine-Tuning Wins Across Every Use Case

On a ticket routing task, fine-tuned Qwen3 30B reaches 80.91%, a 19-point gap over Claude Haiku (61.47%) and 9 points over Gemini Flash (71.93%). Fine-tuned Qwen3 models at every size beat or match Gemini Flash, produce zero invalid outputs (versus 15% for Claude Haiku), and run 2.5–20X faster at p50–p95 latency. DPO: Alignment Without Writing a Reward Function

For tasks where correctness is hard to label but preference is easy to express: structured outputs, compliance judgments, domain-specific alignment. DPO closes the gap between a capable model and a trustworthy one. Evaluated across 100 samples per task by an independent LLM judge. Full-Parameter Training at Frontier Scale

Most training services cap out at LoRA. LoRA is the right starting point: fast, cost-effective, and well-suited for rapid iteration. But LoRA and full-parameter training learn in meaningfully different ways . LoRA learns less and forgets less. Full-parameter produces behavioral changes that adapter-based methods can't reach. Fireworks Training now supports full-parameter training from Qwen3 8B on a single node to Kimi K2.5 (1T parameters) on 64x B200s. Getting there required solving distinct distributed systems problems at each scale: composable parallelism across FSDP, pipeline, context, and expert dimensions; Blackwell-native MXFP8 for expert computation; and streaming pipeline parallelism so RL rollout data trains as it arrives rather than waiting on batch accumulation. More detail in the training infrastructure writeup. LoRA and full-parameter run on the same platform. You don't have to choose your ceiling when you start. Numerical Parity Between Training and Inference

Fireworks runs production inference across DeepSeek, Kimi, Qwen, and others at scale. That's not background — it's the reason the training infrastructure is trustworthy. The numerical edge cases that surface in frontier MoE models aren't hypothetical for us. We've already debugged them, in production, on the same hardware. The hardest part wasn't performance. It was correctness. MoE models are numerically fragile in ways dense models aren't. A small hidden-state change can flip expert selection and cascade through subsequent layers. Kernel fusions that are mathematically equivalent produce numerically different results because floating-point addition is not associative. For RL workloads, KL divergence between training and inference logprobs anchors the reward signal. If those two stacks disagree...