muhan.ai
RESEARCH NODE / ACTIVE v 4.7 · build 2026.05

A neural
substrate
for adaptive
intelligence.

muhan.ai is a research lab building morphic neural systems — architectures that re-route, re-weight, and re-reason in real time. We treat models as living circuits, not static functions.

Open the Lab Explore Models
parameters online
0
B
active circuits
0
k
uptime
0
.97%
01 · Bento Lab

Compartments of capability

Each cell is an independent circuit. Hover the morph cell to watch a representation re-encode itself.

morph · representation C-01

From list to graph in 600ms.

Watch raw vectors morph into a connected reasoning graph. The same data, three syntaxes.

attention · heatmap C-02

Where the model is looking.

head 4 / layer 11 · entropy 1.84 nats

training · loss C-03

Loss curve.

epoch 042 · 0.0184

params · stack C-04

Architecture.

  • d_model4096
  • heads32
  • layers72
  • ctx1.0M
  • dtypebf16
paper · drop C-05

Latent Routing for Continuous-Depth Reasoners.

We propose a routing scheme that lets each token choose its own depth, reducing FLOPs by 38% with no quality regression on MMLU-Pro.

Yu, Han, Park arXiv 2026.0507 read draft →
02 · Model Registry

Three circuits, one substrate.

Each model exposes the same interface. Swap the weights, keep the protocol.

L muhan-L · 412B

Frontier reasoner. Long-horizon planning, agentic tool use, tight calibration on math + code.

  • ctx1.0M
  • tps184
  • MMLU92.4
→
M muhan-M · 70B

Workhorse. Vision-language fusion, retrieval, structured output. Designed for production loops.

  • ctx256k
  • tps720
  • MMLU86.1
→
S muhan-S · 8B

Edge-grade. Fast, distilled, on-device. Quantized to 4-bit with negligible quality loss.

  • ctx64k
  • tps2.4k
  • MMLU78.0
→
03 · Systems

The substrate underneath.

Routing, retrieval, eval — the plumbing that makes a model a system.

Latent Router

Token-wise depth routing — each input chooses how much compute it deserves.

Memory Lattice

Hybrid vector + graph memory. Sparse on-device, dense in the cloud, consistent everywhere.

Eval Harness

Continuous benchmarks, regression alerts, behavioral red-team — running on every commit.

Safety Gates

Layered policy + interpretability checks. Refusals are auditable, not black-boxed.

04 · Lab Log

Recent activity.

  1. 2026.05.06

    muhan-L hits 92.4 on MMLU-Pro.

    +1.8 over the previous frontier checkpoint, mostly from the new latent router.

  2. 2026.04.22

    Memory Lattice 0.9 ships.

    Graph + vector unified store, with 38% lower retrieval latency at p99.

  3. 2026.04.03

    Open weights for muhan-S.

    4-bit quantized, Apache-2.0, ~1.4 GB. Runs on a laptop GPU.

  4. 2026.03.18

    Continuous-depth paper accepted (NeurIPS).

    Camera-ready out next month, with reproducible training scripts.

05 · Contact

Plug into the substrate.

Research collaborations, evaluation partners, model-access requests.