Train Past The Frontier Training Api Now Generally Available

Training API now generally available | Fireworks

Training API now generally available

Blog

Train Past The Frontier Training API Now Generally Available Train past the frontier: Training API now generally available

PUBLISHED 8/31/2026

Table of Contents Own your intelligence Training API, now generally available Why teams train on Fireworks

Correctness: Training and inference alignment

Performance efficiency: Maximized throughput across the RL loop

Development speed: Compressed iteration cycle Getting started: Fireworks Lab can help

Table of Contents

Own your intelligence

Explore training

Own your intelligence

The next frontier of competitive advantage is specialized intelligence. The most ambitious companies are training models with the capabilities that underpin differentiated products and operations, across coding, chip design, cybersecurity, and beyond: • Harvey post-trained Kimi K3 for long-horizon legal work using asynchronous RL. Harvey Tenet scored 19.7% all-pass on LAB against 11.5% for Claude Fable 5, while running at roughly a third of the cost per task. • Vercel used RFT and speculative decoding to fine-tune an open model for v0's auto-fixer, reaching a 93% error-free generation rate for v0 and 40x end-to-end latency improvement. • Heidi Health moved its clinical scribe onto open models it fine-tuned, going from POC to production in four weeks and achieving 3.5x lower latency. • Factory fine-tuned two small LoRA adapters (one to flag risky secrets, one to clear false alarms) on an open Qwen base. At a 5% false-alarm budget, the trained model caught about 70% of real secrets versus ~59% for GPT-5.5, while being faster and more cost-effective.

Across these workloads, training a base model for specific tasks improved quality, latency, cost, and/or behavioral consistency. Fireworks makes this level of specialization accessible to teams through three training surfaces, each designed for a different level of expertise, operational ownership, and control. Training API Managed Training Fireworks Lab Built for ML Researchers ML Engineers All Levels Workflow Bring your own data and loop for deepest control. From LoRA and per-token pricing to full-parameter and reserved capacity. Bring your data or evals, we run the loop. Pick your method (SFT, RFT, DPO) and run via UI/API. We provide embedded ML researchers and forward-deployed engineers. Engagements range from co-design to custom build.

No infrastructure to manage: Choose your method on Managed Training or write your own loop with the Training API. Today, we are announcing the general availability of our Training API and Fireworks Lab. Training API, now generally available

We’ve worked with ML teams around the world to understand where existing training workflows fall short: constrained model and method choice, limited control over parameters and training loops, rigid and underutilized compute, and fragmented training and rollout infrastructure that bloats costs, creates synchronization overhead, and introduces numerical drift. Those partnerships shaped our Training API. The Fireworks Training API connects your custom training loop to Fireworks-managed distributed training and rollout infrastructure. You orchestrate the loop in Python from wherever you choose, with full control over your loss or reward, data, and environment. Fireworks manages high-performance compute: the trainer computing gradients and updating the model, and the rollout deployment generating samples from the trained model. Fireworks also manages the interaction between these two, from weight synchronization to failed-swap recovery, and train-rollout alignment. All of this allows you to focus on refining your learning signal, not infrastructure challenges. python Copy

1 2 3 4 5 6 7 def reward_fn ( completion : str , row : dict ) - > float : """Return 1.0 if the model's numeric answer matches the ground truth.""" predicted = extract_answer ( completion ) truth = extract_answer ( str ( row . get ( "ground_truth" , "" ) ) ) if predicted is None or truth is None : return 0.0 return 1.0 if predicted == truth else 0.0

Example of a custom reward function defined in Python. This code can be dropped directly into a training loop using the Fireworks Training API. One API, two compute options.

Choose Serverless when speed to experiment matters most, or Dedicated when scale, control, and GPU economics take priority. Serverless training allows you to train LoRA adapters on top of shared infrastructure. You pay per token, and sampling runs in the same session. Use it to iterate on experiments, de-risk a larger run, or run RL loops that alternate between rollout and training. Deploy a promising checkpoint to production inference instantly through the UI (Managed Training) or API (Training API & Managed Training). With multi-LoRA deployment, each customer or use case gets its own tuned model without standing up separate infrastructure. Dedicated training is best when you need full-parameter training, models beyond the serverless pool, larger context lengths or LoRA ranks, or sustained throughput where per-GPU-hour pricing is more economical than per-token. Fireworks gives you dedicated, elastic capacity sized to each run, with the ability to scale GPU type and cluster size as needed. Serverless Dedicated Parameter mode LoRA LoRA and full-parameter Models A curated list of popular models Unlimited depth and breadth, up to the largest MoE models Provisioning Attach to an always-on shared pool, nothing to spin up Trainer and deployment provisioned for your run; support for disaggregation to speed up training across trainer or rollout Throughput Shared capacity and per-account rate limits No contention, scale up as you need Billing Per-token Per-GPU-hour

"The Fireworks self-serve API has boosted our experimentation velocity and overall productivity, freeing up Figma's AI team to spend more time on science and research, and less time debugging infrastructure. We're able to iterate faster and build features that make a real difference in design and product development workflows." — Sumithra Bhakthavatsalam, AI Research Manager, Figma

Start on Serverless in minutes, nothing to spin up: Start training → Why teams train on Fireworks

Fireworks is among the few organizations outside the frontier labs to have operated reinforcement learning (RL) across more than 10,000 GPUs. RL turns training and inference into one continuously coupled system;...