Hikaru-1, Update 1: reassessing the base model.

We are pausing forward progress on Hikaru-1 to take stock. This is not a setback so much as a checkpoint: a point in development where scope and integrity matter more than momentum, and where the responsible move is to confirm the foundation before building further on top of it.

Why we reopened the base model question

Hikaru-1 originally began as a finetune of IBM Granite 4.1 (8B). Granite earned that position on strength alone: solid tool-calling and reasoning performance for its size. But extensive testing surfaced a problem that benchmarks do not capture. Granite is, by design, a corporate model, and it carries a corporate personality with it. That personality resisted the shift toward the Hikaru persona regardless of how we adjusted training. We concluded that no amount of finetuning would fully separate the model from its origin, so we ruled it out and reopened the search.

An evaluation across 40+ models

We tested candidates ranging from 3B to 230B parameters. Seven stood out as the strongest, though not the only ones considered:

• Mistral-NeMo
• Llama 3.1
• Qwen 3
• Qwen 3.6
• Gemma 4
• LFM2
• Ministral-3

Of these, Ministral-3 (8B) placed at the top, just ahead of Qwen 3.6.

Why Ministral-3 (8B)

Size was the first filter. 8B parameters is a testbed scale: large enough to be meaningful, small enough to iterate on quickly. That alone put several of the larger candidates out of consideration for this stage.

The second filter was scope. Ministral-3 does not carry capabilities Hikaru-1 has no use for, such as audio input or built-in reasoning modes, which add complexity without adding value to what we are trying to test. Ministral-3 does support vision input, but that capability can be dropped from the model entirely, leaving a pure text-and-tool model. That is the surface area Hikaru-1 actually needs, and nothing more.

What this means going forward

Hikaru-1, and the following point releases from 1.1 through 1.9, will be built on the Ministral-3 (8B) base. This gives us a stable, consistent foundation to run the Explanatory Reinforcement Learning (ERL) evaluation the Hikaru-1 line exists to answer, without the persona conflict we ran into with Granite.

Looking past Hikaru-1

If Hikaru-1 validates our approach, Hikaru-2 is intended to be built from scratch: pretrained on roughly 30 trillion tokens across 200+ languages, followed by extensive ERL fine-tuning and post-training. That plan reflects our current thinking, not a commitment. The Hikaru-2 concept can, and most likely will, change before we get there.