Development Roadmap 2026 Q1

SpecForge aims to make speculative decoding more accessible. After being open-sourced for 5 months, SpecForge has seen better performance, reliability and extensibility. However, to better improve the training experience and speculative decoding ecosystem, we propose the following roadmap for 2026 Q1, feel free to join us if you are interested!

I will label each part with corresponding priority.

Model Release (high priority)

We will release SpecBundle phase 1 around Christmas in 2025. In 2026 Q1, we aim to push further to cover reasoning models, this will be called SpecBundle phase 2. The following models will be covered.

• Qwen3 thinking
• Qwen3 Next thinking
• Kimi K2 Thinking
• GPT-OSS (we have released before, but we will train a better one)
• Dpsk

VLM Support (high priority)

Currently, we only see a specialized eagle3 model for QwenVL, but we should try to make it generalized so that any VLM can be plugged in and get trained.

• General VLM training
• VLM performance tuning, some users report VLM has lower speedup #352

Model Card (high priority)

• SpecBundle Benchmark Website
• Add model card to SpecForge docs
• Add spec-bundle usage to SGLang cookbook

Long Context Support (medium priority)

• Sequence Parallel @uygnef

MTP Finetuning

• Support finetuning native MTP

Multi-layer MTP (low priority)

• Support multiple layer EAGLE3 @xiaoxi-s

Performance (high priority)

• Support flex attention backend on Blackwell cards
• Reduce the bubble time in training @FrankLeeeee
• Reduce memory usage in SGLang backend @FrankLeeeee

Ecosystem (medium priority)

• Support vLLM weights format
• Support TensorRT weights format

More spec algorithm support (medium priority)

• Support dllm draft @xiaomin-D @sleepcoo

Everyone interested in joining our development, welcome join our slack channel
https://sgl-fru7574.slack.com/archives/C09784E3EN6

No specific file, test, or entry point is named. Start by choosing one unchecked roadmap item and locating its existing training or model-support implementation; done would require a scoped change with validation for that item.