Fireworks AI Adds GLM 5.1 Training to Build Long Horizon Coding Agents

Fireworks AIFireworks AI

Fireworks AI added Z.ai's GLM 5.1 to its training platform, supporting supervised fine-tuning and direct preference optimization with a 200K context window. This allows developers to customize the flagship agentic model for multi-hour autonomous tasks without the numerical drift common in fragmented training and inference stacks.

Fireworks AI, an inference and training platform for fast model serving (running a trained model to generate outputs), added GLM 5.1 to its managed and API workflows. This update enables supervised fine-tuning (adapting a model to specific instructions) and direct preference optimization on the flagship model from Z.ai.
Context window
200K tokens
Training methods
SFT, DPO, RFT
Max model scale
1T parameters
Hardware
NVIDIA B200 clusters
Pricing
Per token or GPU-hour
Availability
Managed Training and Training API

The release follows Fireworks AI's Day-0 Kimi K2.6 support and the addition of DeepSeek V4 Pro, signaling a shift toward specialized foundries for agentic engineering. By using shared kernels, the platform eliminates numerical drift—ensuring model behavior during evaluation matches performance in production environments like Cursor or Vercel.

You can now fine-tune GLM 5.1 with a 200K context window using smart defaults or custom loss functions. This workflow is optimized for building agents that autonomously navigate codebases for up to eight hours. Managed training is priced per token, while the Training API uses per-GPU-hour pricing.

View the full update on fireworks.ai
Fireworks AI
Fireworks AI
@FireworksAI_HQ
X

GLM 5.1 from @Zai_org is now available on @FireworksAI_HQ Training Platform across the Managed and Training API workflows. Try SFT and DPO with smart defaults or your own custom loss function with a 200K context window, perfect for long-horizon agentic coding fine-tunes. RL coming soon. Get started: https://t.co/rqSamw3I3e

2retweets40likes
View on X

Still wondering? A few quick answers below.

GLM 5.1 is a flagship model from Z.ai designed for agentic engineering and long-horizon tasks. It is optimized for autonomous workflows that require planning and execution over extended periods, such as software development. The model is now available on the Fireworks AI platform for both high-speed inference and custom fine-tuning.

Fireworks AI supports several post-training techniques for GLM 5.1, including Supervised Fine-Tuning and Direct Preference Optimization. Developers can also use Reinforcement Fine-Tuning, which allows for defining reward functions instead of writing thousands of manual demonstrations. These methods are available across managed workflows and a flexible Training API for advanced research teams.

When training GLM 5.1 on the Fireworks AI platform, developers can utilize a 200K token context window. This large window is specifically intended for long-horizon agentic coding tasks, where the model needs to process and reason across massive codebases or long sequences of autonomous actions without losing critical information from earlier steps.

Fireworks AI uses the same infrastructure and optimized kernels for both training and production inference. This unified approach eliminates numerical drift, a common problem where a model performs differently in a serving environment than it did during evaluation. By sharing the same hardware and software stack, model behavior remains consistent from training to deployment.

Pricing for the Fireworks AI training platform depends on the chosen workflow. Managed training jobs are billed based on tokens processed or GPU-hours, depending on the specific job type. The Training API, which offers more control for custom loss functions and large-scale reinforcement learning, uses a predictable per-GPU-hour pricing model for rollout-heavy workflows.

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

Fireworks AI

See all AI news & updates from Fireworks AI →

Share this update