Compute efficiency.
Compute efficiency relative to Qwen3-0.6B under the stated workload.
VS QWEN3-0.6BA meaning-first
language model.
A language model built around a representation field, connecting context with expression. The first release from a different approach to intelligence.
Compute efficiency relative to Qwen3-0.6B under the stated workload.
VS QWEN3-0.6BTrained on fewer tokens. Outperforming competitors trained on 28T.
TRAINING TOKENS7.54× memory efficiency relative to Qwen3-0.6B.
VS QWEN3-0.6B240M parameters. Ahead of LFM 700M in the evaluated comparison.
PARAMETER CLASSEFFICIENCY WORKLOAD32,768 input tokens + 256 output tokens · BF16 · batch 1 · KV-cached decoding.
Task leader Sphota 240M
Instruction following · 1 of 15 tasks.
Each bar shows this task’s reported score.
Each stage compares the reported scores for one task. The final Overall stage uses the supplied summary scores.
Maya is the parent of Leo. Ben is the parent of Alice. George is the parent of Jacob. Alice is the parent of David.
Who is the grandparent of David? Answer with only the name.
Ben
David
George
David's grandparent is Alice.
Actual CPU benchmark outputs; each model received the same question.
Greedy decoding · BF16 · 8 CPU cores · batch 1 · max 48 new tokens · no web search · no added system prompt.
Native Hugging Face chat templates for the baselines; Qwen thinking off. Sphota uses a Question/Answer wrapper.
Explore what a different approach to language can make possible.