Skip to content
Nova LoopPrelim
How it loops Desk Receipts Model and API Open the Desk
The model

One block, run again.

Nova Loop is built on SenseNova-Loop-9B, an open-weight model that gets deeper computation from the same weights by running its middle layers twice. This page covers the model, the engine that serves it, and the API.

Engine status

EngineSenseNova-Loop-9B
Statuschecking…
Without the engineRuns still do the Prelude and every read, and return the raw readings.
Thinking modeOff by default for speed; the loop passes do the re-checking.

SenseNova-Loop-9B at a glance

ParametersAbout 9B stored, about 14.6B worth of compute per token
Layers32 distinct, 56 runs per forward pass
Looped blockLayers 4 to 27, two passes, shared weights
AttentionThree linear-attention layers then one full-attention layer, repeating
Residual streams2, mixed by learned maps (mHC, Sinkhorn-normalised)
Context512K tokens; text and images in, text out
LicenceApache 2.0, released 17 November 2026

How the loop maps to a Nova Loop run

Inside the model, a token passes layers 0 to 3 once, then the block of layers 4 to 27 twice (the second pass reads the first pass's output), then layers 28 to 31 and the output head once. Written short: head(C(B(B(A(h))))).

A Nova Loop run follows the same order one level up. The Prelude reads what the question names. The loop block is one prompt, applied again to its own draft together with a fresh read of every moving number. The Coda writes the answer, the chart and the receipt. SenseNova's own loop-scaling tests found three or four loops scored best on data analysis and coding; Nova Loop lets you pick one to four passes, or Auto.

Hosting the engine

The engine is any server that hosts the SenseNova-Loop weights behind an OpenAI-compatible /v1/chat/completions endpoint. Two ways the model's authors document:

# GGUF (Q4_K_M, about 5.8 GB) with the OpenSenseNova llama.cpp fork
llama serve -hf sensenova/SenseNova-Loop-9B-GGUF:Q4_K_M -ngl 99 -c 16384 --port 8000

# vLLM with the SenseNova Loop support from vLLM PR #64832
vllm serve /path/to/SenseNova-Loop-9B --served-model-name sensenova-loop-9b --dtype bfloat16

At 4-bit weights with an 8-bit KV cache, a 64K context needs about 10 GiB of GPU memory.

Loop API

GET /api/loop says whether the engine is on. POST /api/loop runs a question and streams one JSON event per line (NDJSON) as each step lands.

curl -N https://novaloop.pw/api/loop \
  -H 'content-type: application/json' \
  -d '{"question": "How is $NVDA doing today?", "depth": "auto"}'
startengine name and whether it is on, depth, start time (UTC)
planthe reads the Prelude picked, and whether rules or the model picked them
readone per read: tool, argument, data or error, time taken
pass-start / passone per loop pass: answer, checks, what changed, confidence, time
needextra reads a pass asked for
offlinethe engine is off: the run stops after the reads
codafinal answer, chart data, sources, run hash
savedthe run is stored; GET /api/run?h=<hash> returns it

Limits: 12 runs an hour per connection, questions up to 500 characters, depth "auto" or 1 to 4.