Noema Blueprint / 01

Make a Project
for what comes next.

Choose its purpose, device, and local model.

No files · no prompts · no account

Shape your Noema Project here.Manual mode

Choose a purpose

Choose a device

Set the runtime

Local-only by design.Files, prompts, and conversations stay out of the link. Add them yourself in Noema.

Live Project

General Project

Ready

iOS

Destination

iPhone.

Noema 1.5 2B · Balanced · local-only

Noema checks the actual device

ProjectGeneral ProjectFresh chat in a local Noema Project

ModelNoema 1.5 2B2B · Noema

RuntimeBalanced16,384 tokens · No overrides

PrivacyLocal-onlyFiles and conversations stay in Noema

Model catalogPublic catalog IDs only

Independent model settings

Tune its voice.

VariationFocused to expressive0.70FocusBreadth of likely words0.95Repetition guardNeutral at 1.001.00

Fine controlsChatGPT can change any single value through WebMCP

Top-kMin-pRepeat windowPresenceFrequencySeed

Runtime overridesSafe GGUF controls · stored only when changedNo overrides

These are advanced model-loading and inference controls. Leave one blank to make no Blueprint override; Noema keeps the installed setting or resolves it on the actual device.

Compute shape

Bounded controls for how the GGUF runtime uses the device.

GPU layers-1 means all / automaticCPU threadsBlank makes no Blueprint overrideEvaluation batchMultiple of 32Physical batchMultiple of 32 · no more than 512 resolvedActive MoE expertsBlank makes no Blueprint override

Loading & memory

Choose No override, On, or Off explicitly for each behavior.

Vision projectorLoad it when the model includes one

KV-cache offloadMove cache work to supported acceleration

Unified KV cacheShare one local cache layout

Keep in memoryKeep weights resident while active

Memory mappingMap model weights from local storage

Skip warmupStart without a separate warmup pass

Preserve thinkingKeep thinking output for supported models

Flash AttentionV-cache quantization applies when enabled

Cache precision

Conservative allowlists only—never raw paths or model identifiers.

K-cache quantIQ4_NL is intentionally unavailable for KV-cache quantRelevant when Flash Attention is enabledOverfit handlingOff or automatic only

Speculative MTP

Draft controls become relevant only when MTP is selected.

Speculative modeOff or MTP onlyMTP draft maximumMaximum draft tokensMTP draft minimumNo more than 2 resolvedMTP probability floorMinimum accepted draft probability

MTP auto-tuneLet Noema resolve draft bounds on this device

Final movement

Send it to Noema.

02 / 02

When opened in Noema, this compact link creates or reuses your General Project and opens a fresh chat. Files and prompts remain yours to add.

···Preparing your code

When you are ready, reveal the handoff.

Open in Noema

No login. No pairing. The link carries only this Blueprint.