Think in branches,
not threads.

Pine puts your AI chats on an infinite canvas. Highlight any part of an answer to branch it into a new conversation that keeps the context.

pine.app

Welcome to Pine

the canvas for

everyone who asks one more question
and hates losing the thread.

// context.txt
paste the earlier answer
here, then ask again…
⌘C ⌘V

// notes.md
- branch idea: GQA?
- ask about batching
- where was that link

A good question never comes alone.

But chat apps keep every follow-up in one long, scrolling line.

pine

Time to branch out.

how it works

One question becomes a map.

01 Ask
Start a chat with any model. It lives in a window on the canvas.
02 Highlight
Select the part of the answer you want to dig into.
03 Branch
A new window opens beside it, already carrying the context.
04 Mix models
Every window has its own model. Get a second opinion in the next branch.
05 Zoom out
Your rabbit hole is now a map, and it saves itself as you go.

Your rabbit hole is now a map, and it saves itself as you go.

Faster inferenceGPT-6.1 Sol

How do I make LLM inference faster?

Decode is limited by memory bandwidth, so read fewer bytes per token: use quantized weightsBranch, a smaller KV cache and every batch you can fit.

QuantizationGPT-6.1 SolClaude Opus 5.5

from “quantized weights”

Does 4-bit quantization hurt quality?

A little. Good 4-bit methods keep most of the accuracy and cut weight memory about 4×.

BatchingGemini 4

from “every batch”

Why do bigger batches help?

Each weight read is shared by every request in the batch, so throughput climbs.

Speculative decodingDeepSeek V4

A small model drafts tokens; the big one checks them.

Prefill vs decodeKimi K3

Prefill is compute-bound; decode is memory-bound.

GPTQ vs AWQGPT-6.1 Sol

Two ways to choose which weights to protect.

FP8 inferenceClaude Opus 5.5

Newer GPUs run 8-bit math almost for free.

saved · 7 windows

Branches keep their context.

Highlight a phrase in any answer and branch from it. The new window starts with the conversation up to that exact message, so you never re-explain.

The main thread stays where you left it, one glance away on the canvas.

Start branching

pine.app · branch

…each decode step re-reads the whole KV cache, so long chats get slower per token.

from “KV cache”

Why does the KV cache grow with context length?

Every branch knows where it came from.

Chat is a line. Thinking isn't.

Every side question pushes the main one further up the scroll. In Pine each tangent grows into its own branch, so you can go deep without losing your place.

See it play

pine.app · compare

a regular chatscroll, scroll… lost it

How does LLM inference work?

Prefill the prompt, then decode token by token…

Why is decode memory-bound?

Each step re-reads the weights and KV cache…

Wait, back to batching. Why does it help?

One weight read is shared across requests…

And what does 4-bit quantization cost?

A little accuracy for ~4× less weight memory…

What was the KV cache formula again?

2 × layers × KV heads × head size per token…

How does LLM inference work?

Prefill the prompt, then decode token by token…

Why is decode memory-bound?

Each step re-reads the weights and KV cache…

Wait, back to batching. Why does it help?

One weight read is shared across requests…

And what does 4-bit quantization cost?

A little accuracy for ~4× less weight memory…

What was the KV cache formula again?

2 × layers × KV heads × head size per token…

the same chat in pineevery thread, in view

The same rabbit hole, mapped.

models

Every frontier model, on one canvas.

Pick a model per window — 10 of them from OpenAI, Anthropic, Google, xAI, DeepSeek, Moonshot AI. Ask one, branch, and ask another the follow-up. No new tab, no copy-paste.

Start here Learn more

pine.app · model

LLM inferenceGPT-6.1 Sol

GPT-6.1 SolOpenAI
GPT-6 LunaOpenAI
Claude Opus 5.5Anthropic
Claude Sonnet 5.5Anthropic
Claude Haiku 5.5Anthropic
Gemini 4 ArgonGoogle
Gemini 3.8 FlashGoogle

Your next idea has branches.

Follow every one of them. Free to start, no card needed.

Start branching Log In

Think in branches,
not threads.

Pine puts your AI chats on an infinite canvas. Highlight any part of an answer to branch it into a new conversation that keeps the context.

pine.app

Welcome to Pine

the canvas for

everyone who asks one more question
and hates losing the thread.

// context.txt
paste the earlier answer
here, then ask again…
⌘C ⌘V

// notes.md
- branch idea: GQA?
- ask about batching
- where was that link

A good question never comes alone.

But chat apps keep every follow-up in one long, scrolling line.

pine

Time to branch out.

how it works

One question becomes a map.

01 Ask
Start a chat with any model. It lives in a window on the canvas.
02 Highlight
Select the part of the answer you want to dig into.
03 Branch
A new window opens beside it, already carrying the context.
04 Mix models
Every window has its own model. Get a second opinion in the next branch.
05 Zoom out
Your rabbit hole is now a map, and it saves itself as you go.

Your rabbit hole is now a map, and it saves itself as you go.

Faster inferenceGPT-6.1 Sol

How do I make LLM inference faster?

Decode is limited by memory bandwidth, so read fewer bytes per token: use quantized weightsBranch, a smaller KV cache and every batch you can fit.

QuantizationGPT-6.1 SolClaude Opus 5.5

from “quantized weights”

Does 4-bit quantization hurt quality?

A little. Good 4-bit methods keep most of the accuracy and cut weight memory about 4×.

BatchingGemini 4

from “every batch”

Why do bigger batches help?

Each weight read is shared by every request in the batch, so throughput climbs.

Speculative decodingDeepSeek V4

A small model drafts tokens; the big one checks them.

Prefill vs decodeKimi K3

Prefill is compute-bound; decode is memory-bound.

GPTQ vs AWQGPT-6.1 Sol

Two ways to choose which weights to protect.

FP8 inferenceClaude Opus 5.5

Newer GPUs run 8-bit math almost for free.

saved · 7 windows

Branches keep their context.

Highlight a phrase in any answer and branch from it. The new window starts with the conversation up to that exact message, so you never re-explain.

The main thread stays where you left it, one glance away on the canvas.

Start branching

pine.app · branch

…each decode step re-reads the whole KV cache, so long chats get slower per token.

from “KV cache”

Why does the KV cache grow with context length?

Every branch knows where it came from.

Chat is a line. Thinking isn't.

Every side question pushes the main one further up the scroll. In Pine each tangent grows into its own branch, so you can go deep without losing your place.

See it play

pine.app · compare

a regular chatscroll, scroll… lost it

How does LLM inference work?

Prefill the prompt, then decode token by token…

Why is decode memory-bound?

Each step re-reads the weights and KV cache…

Wait, back to batching. Why does it help?

One weight read is shared across requests…

And what does 4-bit quantization cost?

A little accuracy for ~4× less weight memory…

What was the KV cache formula again?

2 × layers × KV heads × head size per token…

How does LLM inference work?

Prefill the prompt, then decode token by token…

Why is decode memory-bound?

Each step re-reads the weights and KV cache…

Wait, back to batching. Why does it help?

One weight read is shared across requests…

And what does 4-bit quantization cost?

A little accuracy for ~4× less weight memory…

What was the KV cache formula again?

2 × layers × KV heads × head size per token…

the same chat in pineevery thread, in view

The same rabbit hole, mapped.

models

Every frontier model, on one canvas.

Pick a model per window — 10 of them from OpenAI, Anthropic, Google, xAI, DeepSeek, Moonshot AI. Ask one, branch, and ask another the follow-up. No new tab, no copy-paste.

Start here Learn more

pine.app · model

LLM inferenceGPT-6.1 Sol

GPT-6.1 SolOpenAI
GPT-6 LunaOpenAI
Claude Opus 5.5Anthropic
Claude Sonnet 5.5Anthropic
Claude Haiku 5.5Anthropic
Gemini 4 ArgonGoogle
Gemini 3.8 FlashGoogle

Your next idea has branches.

Follow every one of them. Free to start, no card needed.

Start branching Log In

Think in branches,not threads.

Time to branch out.

One question becomes a map.

Branches keep their context.

Chat is a line. Thinking isn't.

Every frontier model, on one canvas.

Your next idea has branches.

Think in branches,not threads.

Time to branch out.

One question becomes a map.

Branches keep their context.

Chat is a line. Thinking isn't.

Every frontier model, on one canvas.

Your next idea has branches.

Think in branches,
not threads.

Think in branches,
not threads.