Spend full attention where it earns its cost.
X1 interleaves gated linear attention with full attention. The 24-query, 4-key/value head layout keeps the model dense while reducing the memory pressure of long sequences.
Research
The decisions inside X1.
X1 is a released model and a concrete research object: 64 layers, hybrid attention, controllable reasoning, long context, and one evaluation record.
We publish the decisions behind the model: how it learns, how it runs, and where it falls short.
Effort is a control, not a personality.
The chat template exposes xhigh, medium, low, and no-thinking modes. Hard research and coding tasks can use a long budget; interface generation can finish quickly at low effort.
Keep the visual path intact.
The release is natively multimodal. Even in the FP8 build, the vision tower and output head stay bfloat16 so a smaller package does not quietly become a text-only model.
Evaluation
A benchmark number needs its conditions.
X1 was evaluated with vLLM, bfloat16 weights, FP8 KV cache, and a single RTX PRO 6000 Blackwell at 131,072-token context.
1.0temperature
0.95top-p
20top-k
xhighreasoning effort
We remove reasoning traces before scoring, so the result reflects the final answer.