Introducing Hindsight Models Hindsight memory is now 2.5x cheaper and 2x faster, with better answers. The LLM writes, a small decision model judges. Open weights on Hugging Face, or hosted on the new Hindsight Inference API.

The models: hs-decide-1-flash makes the small decisions in 167 ms. The LLM call it replaces takes 1 to 7 seconds. hs-rerank-1 achieved 56.0 across 7 reranking benchmarks (next best 51.3), 6× faster on CPU, and the default reranker in open-source Hindsight. hs-embed-1 finds memories by meaning, multilingual search in 16.5 ms on two CPU threads, at 112M parameters.

Self-hosting Hindsight? One login and one environment variable, and Hindsight Inference runs the models for you. No GPU, $5 of free credit. Or take the open weights from Hugging Face.

Launch blog: Hugging Face: Leaderboards: Benchmarks:

Nice! Going to try this with my agents.

Biggest level up for Hindsight yet

🙌

@Presidentlin

Try it in the playground, no GPU needed:

why can’t i use my own decision model? if im ok with a bit worse results!

You can! The decision model is optional, and Hindsight runs fine with just your LLM making those calls. hs-decide-1-flash is there for when you want them faster and cheaper, and since it's open weights, you can also swap in or fine-tune your own.

oh wow that is amazing

Splitting memory into a writer LLM and a small judge model makes sense for agents that run all day. What does the judge decide exactly: what to store, what to retrieve, or when to overwrite stale facts?

The LLM extracts the facts, then the decision model compares each one against what's already in memory and decides what happens to it, like adding it or overwriting a stale one. Full breakdown here:

A small judge model deciding what the big LLM gets to remember is a neat split. Curious how it holds up when memories contradict each other over time.

That's one of its core jobs: deciding when a new memory changes what we already know, and rewriting the page when it does. It scores 79.5 on that, vs 56–72 for other models its size.

Open weights give developers more flexibility.

Your stack, your call.

Vectorize unveils improved Hindsight Models delivering cheaper faster memory with enhanced answer quality

Thanks for sharing! The playground is the fastest way to see it:

Thanks for sharing! I’ll check out the playground and explore its features. Looks interesting!

Impressive efficiency gains with smarter judging model approach

Appreciate it. Letting the LLM write while a 0.8B model judges turned out to be the big unlock.

Love that efficiency win 2.5x cheaper and 2x faster with better answers is a huge leap!

Thanks! Have you tried it with your stack yet?

Huge win cheaper, faster, and better answers all at once!

Cheaper faster better answers win

Cheaper faster memory with better answers is a win

Hindsight’s small judge has the scarier job. One bad veto can make a faster LLM feel broken.

Smaller model sharper memory

How would you compare hs-decide against Jev in terms of accuracy?

hs-decide is not quite as good as Jev yet, but we are working on it. For its size, it scores competitively on industry benchmarks.

How does the embed/reranker compare to Qwen 8B embed/reranker?

Congrats on the launch! Gonna give it a try

Thank you! Let us know how it goes for you, we'd love to hear what you build.

Sure!

A small judge is useful if its decisions stay calibrated across different memory tasks.

Cheaper faster memory with a judge built in

Smaller judge model makes LLM smarter

2.5x cheaper + 2x faster + better answers is the trifecta nobody says no to. LLM writes, small model judges — clean architecture.

Writer judge setup getting cheaper and faster is great

can the decision model run fully offline, like on a phone?

It can run fully offline

@manthanguptaa 牛逼了