GRaPE 2.5 models are here. There are 2 flavors for you to explore with this generation:

  • GRaPE 2.5 Quasar. The biggest, the baddest, the best GRaPE model to-date. Standing in at 27B Parameters.
  • GRaPE 2.5 Helios. Stronger, more reliable, and more capable than previous GRaPE models. Helios has 10B parameters.

GRaPE 2.5 models aim to dominate in expanding creativity. Previous GRaPE models have been more-assistant oriented, this is the first generation of GRaPE models where creativity has been a focus.

As a result, you may find situations where these models overachieve standard expectations.

Along with this, we are adding a few new features to the way GRaPE thinks.

  1. GRaPE thinks with thought summaries and todo lists. In testing, we found giving the model a steady TODO-list of items to check off as it works provides a more grounded process. It will revert to topics it thought about ages ago much less often than you would see in Qwen, or Gemma 4 models.

  2. GRaPE will generate 4-8 ideas before committing to thinking. In our attempt to improve the creativity of GRaPE 2.5 models, we found that having the model generate ideas before diving deep into thinking can help the model explore more options, more variety, and develop better angles-of-attack before proceeding.

  3. Thinking modes have been overhauled. When we trained GRaPE 2 models, we found that they were deeply attracted to the lower-side of thinking. While this can help for simple tasks, it degrades GRaPE's overall capability. We have adjusted the thinking modes, and the amount GRaPE was given during thinking, so it can tackle more complex problems.

  • Minimal: skips thinking
  • Low: below 1024 tokens
  • Medium: between 1024 and 4096 tokens.
  • High: between 4096 and 8192 tokens.
  • Xtra-hi: above 8192 tokens.

In SLAI apps, the thinking mode will automatically be appended to user prompts.

We currently have no plans to make more GRaPE 2.5 models. We are instead pursuing expanding our reach beyond fine-tuning models, and are working on GRaPE 3.

Thank you all for taking the time to read this. We are excited for you to try GRaPE Scribe for yourself. Go build something GRaPE!