RL research and pilots

There is published reinforcement learning work for GLM 5.3 and Qwen 3.8. Match each result to the model version and checkpoint before using it to design a job. Sources below were reviewed on 1 October 2026. A preprint, a vendor training report and an implementation guide provide different kinds of evidence.

GLM 5.3#

Z.ai’s GLM 5.3 release report↗, published 14 August 2026, attributes its gains over GLM 5.2 to post-training on the same base model. It describes long-horizon task environments, single-rollout asynchronous optimization (SAO) with compaction, verifiers and the slime framework↗. slime uses Megatron for training and SGLang for rollouts.

The report describes binary task rewards checked against oracle, no-op and unsolved states, and using solver trajectories to find reward shortcuts. This is evidence for GLM 5.3’s development process. It is not a published, validated Grid job for a customer dataset.

The SAO preprint↗, submitted 8 July 2026, studies single-rollout sampling, value-model training and token-level clipping for asynchronous agent RL. Its named deployment is GLM 5.2, so use it as algorithmic background for the continued approach described in the 5.3 report.

For a custom GLM 5.3 pilot, scope the official BF16 checkpoint↗, its licence, distributed trainer, rollout engine, environment and verifier. GLM 5.3 Flash uses a separate architecture. Its BF16 release↗ needs an independent compatibility review.

Qwen 3.8#

27B implementation guide. Unsloth’s Qwen 3.8 training guide↗ documents supervised adapter training and an RL entry point for Qwen 3.8 27B↗. Its RL entry uses the Qwen 3.5 Unsloth path with fast_inference=False, loading the published 4-bit checkpoint↗. This is a practical starting point for compatibility work. It does not establish a complete, validated Grid RL recipe.

2.4T research. The QwenGyre preprint↗, submitted 27 September 2026, studies elastic allocation of GPUs between training and rollouts for long-horizon agents. It reports experiments on Qwen 3.8 2.4T, with trajectories reaching 700K tokens. Those results do not demonstrate an RL result on the 27B checkpoint and do not determine a customer training sequence length or capacity quote.

Inkling and Nemotron#

Inkling Small’s release↗ describes distillation followed by agentic coding RL and makes weights and fine-tuning on Tinker available. Vendor training and vendor-platform support are useful evidence for scoping; Grid needs its own checkpoint and trainer validation.

The official Nemotron 3 Nano 30B A3B model card↗ documents SFT followed by multi-environment synchronous GRPO, using NeMo RL and NeMo Gym. It links the developer recipe and data. This evidence applies to that checkpoint, rather than establishing the same method for every Nemotron model.

Design a pilot#

Our suggested starting point is a bounded Qwen 27B pilot when the task has a reliable verifier and the smaller model meets the untuned baseline. This is an engineering recommendation, not an observed Grid training result. Use a larger model when the baseline and workload justify its distributed training requirements.

Before committing capacity, agree on:

1. The exact checkpoint, revision, licence and supported training and serving formats.
2. A held-out evaluation and an untuned baseline for quality, response time and cost per accepted result.
3. The rollout environment, tool access, sequence length and maximum task duration.
4. The reward or verifier, including tests for shortcuts and cases that should fail.
5. A bounded compute budget, stop conditions, acceptance criteria and the deployment plan.

Training and rollout inference have separate scopes and costs. RL rollouts provide sampling for a trainer you own. They do not perform weight updates themselves. Managed RFT is a separate invitation path with method eligibility confirmed per checkpoint.

Open a Qwen 27B pilot inquiry or a GLM 5.3 pilot inquiry from the RL section. The inquiry asks for the model, checkpoint, task, data, reward or verifier, environment, baseline, success criterion, timing and budget range.