users in ~2 months
ChatGPT · 2023Kraxys · Internal research project
Intelligence at a constant cost.
STATE · 24 × 24 FIXED
TOKENS IN 0MEMORY Δ 0 B
01 / Thesis
Price is the feature.
Every time intelligence got cheaper, people used far more of it.
tokens on day one, free
Ox Alpha · OpenCode · 2026 ↗of paid teams in two weeks
Jev · Vercel AI Gateway · 2026 ↗02 / Architecture
Context grows. Cost doesn’t.
Current and archive matrix memories condition the vector recurrence before its update. A direct matrix path is retained.
0×less memory
at 1M tokens of context
at 1M tokens of context
Transformer · KV cache Modus_X · bounded state
Illustrative · context length →Illustrative: 8B-parameter transformer, 32 layers, grouped-query attention (1,024-dim KV), bf16 — versus a 32 × 1,024² matrix state. Analytical estimate, not a v3 benchmark. Bounded state does not imply lossless recall.
03 / Evidence
What’s real today.
47.46MParameters
3/3Dense validation wins
0.0212Median BPC gain
2.923×Median TPU update speedup
Shown
- Local K/V context improves dense enwik8 validation in all three paired seeds at 25k updates and 102.4M characters per arm.
- Median gain: 0.021213 BPC over the calibrated MemoryFeedback control. Adds 24,576 parameters and 96 KiB of FP32 state per sequence.
- Two-stage execution is 2.923× faster per median TPU training update in two continuation experiments, with effectively unchanged validation BPC.
Open
- Head-to-head v3 comparisons with modern Mamba and Transformer baselines.
- Generalization beyond enwik8 to real developer workloads.
- Trained quality at 1B scale.
04 / Economics
No meter. No ceiling.
When the next conversation costs almost nothing, you can stop charging for it.
Conceptual · Jevons paradox · not to scale
05 / Path
Useful first. Bigger later.
01 · Done
Research
Bounded-memory architecture. Published.
02 · Next
Daily driver
One model, tuned for everyday work.
03 · Planned
Free, unlimited
Served with no meter.
04 · Planned
Scale
Bigger models, informed by real use.
Built before asked.
We’re speaking with a small number of investors and research partners.
Start a conversation