FlashNext-RL: A System Optimization Study for Long-Horizon Agentic RL of Qwen3.8-Flash-Next
Long-horizon agentic reinforcement learning (RL) requires both efficient rollout generation and training on long sequences.
ProofPaper ↗
Key points
- However, the memory requirements and model-specific operators of Qwen3.8-Flash-Next make this difficult with existing open-source implementations.
- To resolve these limitations, we introduce FlashNext-RL, a model-specific RL training optimization framework based on prime-rl, with vLLM for software engineering (SWE) rollout generation.
- Our main focus is BF16 training of the full 48-layer model on a single 8xB200 node, with the first two layers, the n-gram table, the sparse-attention indexer and the multi-token prediction head frozen.
- A controlled cumulative study on an 8-layer model further quantifies the contributions of individual optimizations.
Sources (1)
- [1]FlashNext-RL: A System Optimization Study for Long-Horizon Agentic RL of Qwen3.8-Flash-NextarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 11:53 PM
Long-horizon agentic reinforcement learning (RL) requires both efficient rollout generation and training on long sequences.
However, the memory requirements and model-specific operators of Qwen3.8-Flash-Next make this difficult with existing open-source implementations.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 20, 2026Fine-tuning GPT-OSS-20B for multilingual reasoning on Amazon SageMaker HyperPod
- Oct 20, 2026[AINews] DeepSeek DeepInfer: Open Source SOTA Inference running 155T toks/day/cluster making $22m/day at ~80% margin for a 81 day payback
- Oct 20, 2026MPS in PyTorch 2.15
- Oct 19, 2026Push-based KV transfer to improve TTFT and TPOT for disaggregated inference in vLLM
- Oct 19, 2026Improve TTFT for multi-turn agentic workloads with bi-directional KV transfer in vLLM
- Oct 19, 2026Qwen3.8-27B and Flash-Next on consumer hardware: engine numbers I could verify, and what they cost with 2026 RAM prices