Chatbot, Cabinet, and SAT Prep

Private AI work that changes shape with the problem.

Start in Chatbot for direct model-switching conversation, run Compare when you want multiple model answers side by side, escalate into Cabinet for structured multi-agent rigor, or move into SAT Prep (Beta) for adaptive Digital SAT practice. The main workspace stays temporary by default, with local saving only when you deliberately choose it.

Chatbot is the default path into the workspace. Cabinet stays one deliberate step away when the question needs more structure, critique, or synthesis pressure.

248library models
2workspace modes
SATprep beta
Localsave optional

Main workspace

Chatbot and Cabinet stay separate on purpose.

Chatbot

Chatbot

The direct mode for live threads, faster iteration, and future-reply model switching without rewriting what already happened.

  • Switch models mid-thread for future replies only
  • 248 mainstream model options with provider, task, and legacy filters
  • Carry context into Cabinet when the prompt needs more pressure
Learn more
Cabinet

Cabinet

The flagship structured mode for questions that benefit from critique, disagreement, planning, optional rigor layers, and a final synthesis pass.

  • Configurable 2-5 agent panel with presets and role labels
  • Optional judge scoring, optional web grounding, and a dedicated synthesis model
  • Built for tradeoffs, critiques, launch plans, and prompts where the first answer is incomplete
Learn more

Also included

SAT Prep (Beta) — adaptive and separate by design.

A 20-question diagnostic, targeted practice by section and difficulty band, three AI advisor lanes, and a live score estimate that sharpens with every session. Built for the Digital SAT. No flashcards, no busywork.

Explore SAT Prep

Simple by design

How Sparse Halo works

01

Pick the right surface

Start in Chatbot for direct work, move into Cabinet for structured pressure, or open SAT Prep for adaptive Digital SAT practice.

02

Run the right system

Switch models in a live thread, run Compare, configure a Cabinet panel with rigor layers, or work through diagnostic, practice, and advisor lanes.

03

Keep retention deliberate

Chatbot and Cabinet stay temporary by default in the browser, and durable history only appears when you deliberately save locally.

Privacy-first

Built for privacy, not surveillance

Chatbot and Cabinet are designed around temporary browser sessions, not a default permanent archive tied to your account.

Starting fresh clears the active workspace instead of extending an automatic account-wide history timeline.

If you want to keep a Chatbot or Cabinet thread, local saving is opt-in and stays on your own device.

The privacy promise here is about retention inside Sparse Halo, not a claim that outside model providers never process requests to generate replies.

Multi-model access

248 mainstream model options. One workspace.

The workspace currently spans 115 public models plus Pro-gated smart, heavy, and legacy options, all organized through the searchable provider and task picker.

Browse the public model library

OpenAI

GPT-5.4 ProHeavy
GPT-5.5Heavy
GPT-5.5 ProHeavy
o3 Deep ResearchHeavy
GPT-4.1 MiniSmart
GPT-5.2-CodexSmart
+6 more in the picker

Anthropic

Claude 4.7 OpusHeavy
Claude Opus 4.6 (Fast)Heavy
Claude 4.6 SonnetSmart

Google

Gemini 3.1 Pro PreviewSmart
Gemma 4 31BSmart
Gemini 2.5 FlashPublic
Gemini 2.5 Flash LitePublic
Gemini 3 Flash PreviewPublic
Gemini 3.1 Flash Lite PreviewPublic
+3 more in the picker

Meta

Llama 4 MaverickSmart
Llama 3.1 8B InstructPublic
Llama 3.3 70B Instruct (free)Public
Llama 4 ScoutPublic

DeepSeek

DeepSeek V4 ProSmart
DeepSeek V4 FlashPublic

Perplexity

SonarSmart
Sonar Pro SearchSmart

xAI

Grok 4.20 Multi-AgentSmart
Grok 4.3Smart

Moonshot

Kimi K2.6Smart

Qwen

Qwen3 Max ThinkingSmart
Qwen3.5 397B A17BSmart
Qwen3.5 Plus 2026-02-15Smart
Qwen3.5 Plus 2026-04-20Smart
Qwen3.5-122B-A10BSmart
Qwen3.5-27BSmart
+8 more in the picker

Mistral

Devstral 2 2512Heavy
Mistral Large 3 2512Smart
Ministral 3 8B 2512Public
Mistral Small 3.1 24BPublic
Mistral Small 3.2 24BPublic
Mistral Small 4Public

Nvidia

Nemotron 3 Nano Omni (free)Public
Nemotron 3 SuperPublic
Nemotron 3 Super (free)Public

Z.ai

GLM 5Smart
GLM 5 TurboSmart
GLM 5.1Smart
GLM 5V TurboSmart
GLM 4.7 FlashPublic

IBM

Granite 4.1 8BPublic

Nous

Hermes 3 405B InstructHeavy

Common questions

Frequently asked questions

Sparse Halo is a private AI product family with a main workspace and SAT Prep (Beta). The workspace uses Chatbot for direct model-switching conversation and Cabinet for structured multi-agent reasoning, while SAT Prep adds a separate adaptive Digital SAT practice lane.

No. Chatbot and Cabinet sessions are temporary in the browser by default. If you decide something matters enough to keep, local saving is opt-in on your own device.

Cabinet is Sparse Halo's structured workspace mode. You can configure a 2-5 agent panel, choose the debate shape, optionally enable judge scoring or web grounding, and end with a dedicated synthesis model.

Chatbot is the direct workspace mode. It keeps one live thread, lets you switch models for future replies mid-conversation, and acts as the default path into the main Sparse Halo workspace.

SAT Prep (Beta) is Sparse Halo's adaptive Digital SAT lane. It includes a 20-question diagnostic, targeted practice by section and difficulty band, three advisor lanes, and a score estimate that sharpens with use.

Sparse Halo's privacy promise is about default retention: Chatbot and Cabinet are temporary by default, and local saving is opt-in. That does not claim that outside model providers never process requests to generate replies.

Sparse Halo currently offers 248 mainstream OpenRouter model options across public fast, smart, heavy, and provider-level legacy groups. The picker supports search plus provider and task filters for GPT-5.5, Claude 4.7, Gemini 3.x, Sonar, Grok 4.3, Kimi K2.6, DeepSeek V4, Qwen3.6, and more.

Open the workspace, or head straight into SAT Prep.

Chatbot opens first, Cabinet stays ready for harder prompts, and SAT Prep gives students a separate adaptive lane when the work is academic instead of general.