Meet dots3-note preview — an open-weight multimodal model built for long-horizon agency in real life. → 280B total / 16B active parameters → 512K context window → Native text, vision, and speech understanding Designed to reason, use tools, self-evaluate, and adapt as tasks evolve over time. dots studio

One of the biggest challenges in long-horizon RL is knowing whether an agent is truly making progress before the task is finished. That's the idea behind TEMPO. Instead of relying only on the final reward, the model alternates between actor and critic, reasoning about its current state and estimating the remaining value throughout the task.

dots3-note preview is also built to learn in unfamiliar environments. It can explore, form hypotheses, test them, update its own memory, and apply what it learns to future tasks. Combined with multimodal perception and tool use, it's designed to move beyond understanding information to acting on real-world problems.

Real-life tasks don't happen in a single conversation. User intent evolves, plans change, and external conditions like weather, inventory, or travel schedules can shift at any time. dots3-note preview is designed to maintain context and adapt across these long, evolving interactions.

The release also introduces VibeSearchBench and VibeLifeBench, two open benchmarks designed to evaluate long-horizon AI agents in more realistic scenarios. Tech blog: API: @dotsstudioai

@iam_chonchol whoa, that's a massive context window! curious how it handles real-time adjustments in long tasks.

Long tasks need memory that can actually adapt.

Long-term memory becomes valuable when it improves future tasks.

Nice, long horizon agency is the honest test for a preview like this, and it is where we found a floatless architecture rather than a traditional transformer paid off for EvaEsi.

Multimodal understanding plus tool use is a strong combination.

native text vision and speech support makes the model much more useful for work

Remembering what it learns could change long-term tasks.

Self-evaluation could make agents much more reliable.

Long-horizon agents need to handle change, not just context.

This feels built for tasks that don't end in one prompt.

Changing plans without losing the bigger picture is key.

Progress tracking during a task is a hard problem.

Adapting as conditions change is a huge advantage.

Adapting to real-world variables is where agents get interesting.

The focus on unfamiliar environments is a strong signal.

Progress tracking could be the missing piece for making long-horizon agents truly reliable. 🔥

The 512K context window is seriously interesting.

Learning from unfamiliar environments is the exciting part.

This is moving agents closer to real-world usefulness.

Amazing

Benchmarks like these could make agent progress easier to measure.

The long-horizon agent focus is what makes this interesting.

The actor-critic approach makes this especially interesting.

Real-world tasks rarely follow a fixed plan.

Exploring, testing, and updating memory feels promising.

Repost me plz (t-grm me: mycono)

Fantastic

长时程代理最想看连续跑几天后还稳不稳。

The combination of long context and multimodal reasoning makes this one worth watching.