Lab notes · Series C · Zürich & Toronto
Models that think
in real time.
SYNAPSE builds foundation models, an inference runtime and an agent framework designed as one nervous system. Every token you see on this page is rendered by the same engine we ship.
- 0tokens / sec, single GPU
- 0k context window
- 0peer-reviewed papers
01 · Cortex
A foundation model with a sense of proportion.
Cortex-3 is a 70B mixture-of-experts model trained on 9 trillion tokens of licensed text, code and structured scientific data. It routes each token through only two experts, which is why it answers like a model four times its active size at a quarter of the cost.
- MoE 8 experts, 2 active, learned routing with load balancing
- CTX 128k context with sliding attention sinks
- EVAL 91.4 on the SYN-Bench reasoning suite
02 · Axon
Inference that never makes you wait for the first word.
Axon is our serving runtime: speculative decoding, paged KV cache and a scheduler that treats your latency budget as a hard constraint. It streams the first token in under 90 milliseconds at p99, then holds 2,000 tokens per second on a single accelerator.
- TTFT 84 ms p99 first token on Cortex-3
- BATCH Continuous batching with per-request SLOs
- EDGE Same binary from datacenter to laptop
03 · Dendrite
Agents that branch, verify and come back with receipts.
Dendrite turns a single prompt into a tree of short-lived workers. Each branch runs tools in a sandbox, writes down what it checked and reports to a verifier before anything reaches you. The result reads like one careful colleague rather than a committee.
- TREE Parallel branches with automatic pruning
- PROOF Every claim links to the tool call that produced it
- SAFE Sandboxed execution, scoped credentials
Live demo · simulated
Cortex-3 on Axon, streaming.
A replay of a real session, played back at the speed the runtime produced it. No network calls are made from this page.
Research
Selected papers
Hover or focus a title to preview the figure. All papers are released with code and evaluation harnesses.
Build on the nervous system.
API access is open to research groups and early design partners. Average time from request to first key: two business days.