Mercury 2.5 | Hacker News Reader

Mercury 2.5

inceptionlabs.ai

248 pointsTopfi54 comments

Got my hopes up when it said widely available GPUs that it would be open weights but it doesn’t seem like it sadly

I tried at it creative writing - and, with thinking off, it was considerably better than Mercury 2 and generally good in fact, not very sloppy. Now with thinking on, it got worse, began hallucinating things; this is something I've noticed with all recent models - enabling reasoning causes hallucinations in creative writing assignments.

Similar, I've found that on most recent models I've tried, reasoning will cause it to spend quite a lot of time thinking about whether even the most wholesome fictional scenario is something it's allowed to write about, which in turn causes it to start hallucinating details. For creative writing I've found the minimum of thinking is the best way.

i’ve found this for myself as well, when its time for me to write creatively

Could you explain how you define hallucination in the context of creativity? Lack of consistency aside, the two look quite similar to me - or would you consider the existence of orcs and elves in LOTR to be hallucinations, too?

> Could you explain how you define hallucination in the context of creativity

Mostly as you've mentioned, "lack of consistency", which manifests in variety of ways - presence of cellphones in historical settings (I'd call them "global inconsistency") and a character for whatever reason suddenly smoking a cigarette ("local inconsistency") I've never mentioned nor present in the context.

I like the model.

FYI:

"If you do not want us to use your User Submissions to train our models, you can opt-out by setting the ‘Improve the model for everyone’ option under User Settings in the API Platform to OFF."

I have ZDR enabled globally on openrouter, and was able to use the model.

It likely depends on how you access it. With AI and "Free usage allowance", the price tends to be your soul.

Diffusion is already being used in Drafter in many LLMs.

many people are running Qwen 3.8 27b on TPU at 130tk/s for free on Kaggle TPUs:

https://www.reddit.com/r/Qwen_AI/comments/1w6gv32/qwen3827b_...

I wonder if we are going to see boxes appear soon, which can run these models for dirt cheap.

This team is in a tough spot - they'll need to innovate and get their models in a good spot quickly. They are racing the clock against various similarly sized models being burned into ASICs and running at a similar speed (or faster).

They can run it on „commodity” hardware as gpus, that gives them ability to change direction fast without burning money on ASIC? The space moves fast so model on ASIC can be outdated in couple of months?

I’m using this model to „rerank” results from vector store. The model is provided a set of results and asked to produce a string of 1s and 0s where the offset reflects the position in the result set. The prompt goes along the line „do this set of result match the provided query X”. Works like a charm, normally I would use a small non reasoning model, but given how Mercury produces the output its blazingly fast - which is what I was optimizing for - not to increase the search latency. It helped improving our search in a way that reranker could get close to.

Inception is one of the most interesting neolabs with their diffusion-based architectures. My understanding is that their primary business is low latency voice applications but they are seriously pursuing coding.

We tested Mercury 2.5 Preview, which is nowhere close to the frontier (and not advertised as such), but it's actually usable as a general-purpose chatbot. It's comparable in problem solving ability to some last-gen open weights models, and the price and cost make it compelling. However, they have not figured out general purpose tool use and agentic coding (their model performs worse on our problems when given a custom harness). If they do, I see a lot of real-time applications that the speed and cost will enable.

Ridiculously fast on OpenRouter, just subjectively it's a really strange experience because I've never seen a model respond or execute that quickly.

Try Cerebras. When I think about how speed of generation is another variable to tweak for "intelligence", it seems like this speed is best used for searching for solutions in a problem space and then validating and discarding and keeping what is best. Being intelligent at the Fable level, but what if the Fable level machine could think at 100x? What does that mean: perhaps it means more parallel "experiments" for solutions in the token/generation/hyper-dimensions of the latent space.

Try https://chatjimmy.ai/ from Taalas. There is an emergent space for super-fast-models esp finetuned or guardrailed to solve very specific latency sensitive tasks.

Generation is insanely fast, the other side is presentation, which can be slower and more controlled. Blasting the end-user with text walls is a UX problem now.

They are comparing it to 2 and 3 version old flash/fast versions of models but purely for tok/s. Then only comparing it to Mercury 2 on intelligence. This is very misleading and I suspect this model is basically useless.

Even pretty dumb models are useful for running subtasks (especially at this sort of speed). Note that in the coding section they only mention using it as a subagent for a smarter model

Not related to the Mercury language:

https://mercurylang.org/

or the Mercurial VCS

or the planet Mercury

or the liquid metal element

Or any cats who are named 'Mercury'!

Or Freddie

Interesting model. I tried to make Mercury investigate the hardcoded prompts in my (aider-derived) agent harness and repeatedly got this error:

> server: Upstream error from Inception: I'm sorry, but I can't share details of my architecture or training process. Would you like to learn about how language models work in general instead?

It looks like an overeager IP-protection classifier. However, the model recovered and completed the turn despite the errors (three total).

The idea of getting a speed up in language models from using diffusion is compelling, but it just doesn’t seem like discrete diffusion models work as good at discrete non-diffusion. Which kind of makes sense, tokens don’t have implicit continuous representations that are easily noised. You have to learn a continuous embedding and figure out how to noise that, which is ideal for distillation (Ie the Gemma models, drafter heads) but not necessarily for pretraining.

There’s been more recent work on continuous space diffusion models for language this year. Sander Dielman has a good blog post on that.

But if I’d predict where diffusion lands in LLMs, it’ll be used in looped models like Astra. Once reasoning is happening in hidden states, we’re in a good continuous domain, perfect for diffusion. We’re going to end up swapping “looping” for predicting the models hidden states at the next “timestep” with diffusion. And that way the time to generate traces (no longer human intelligible though) will become 10x faster.

A lot of current diffusion LLMs don't convert tokens to continous space before noising them. They add discrete "noise", which is often as simple as replacing some tokens with [MASK].

The real problem is that when using fewer sampling steps than output tokens, diffusion formulations fundamentally cannot represent distributions where output tokens are heavily codependent. Autoregressive formulations don't have this problem, they can represent any distribution (ignoring limitations of the underlying model).

The benchmark comparison to other models was suspiciously missing. The self-comparison is a good representation of progress, but it is light years behind frontier models. Speed is great, but wrong is much worse than slow, IMO.

Congrats! Happy to see someone seriously pursuing this direction.

Same. Models that are fast (and thus by extension, in some direction, efficient) but not local/self-hosted IMO fill a niche for quick and cheap (and acceptable quality, of course) inference in business contexts.

It's blazing fast! Works awesome for my extension summarizing YouTube videos: https://sifttheweb.com/ . I found the real world performance be 10x faster on average than DeepSeek v4 Flash

interesting: Quality: 40% increase in intelligence from Mercury 2. Comparable to cost-optimized frontier models like GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5. available : https://openrouter.ai/inception/mercury-2.5

This should make an excellent choice for arbiter in llm-consortium, mercury-2 was pretty good. One of the main drawbacks of the multi-model system is the added latency of the llm judge, but having a model run at 1100tps goes a long a way to alleviate that.

I'd imagine at this point they are likely an acquisition target if they can get a halfway decent model. I can't imagine having diffusion sub-agents (or sub-sub-agents) in an orchestration wouldn't be beneficial.

I'm less bullish, this model and previous are halfway decent and you can do diffusion sub-agents and sub-agent-agents today, and there hasn't been a sea change, or anything noticeable or well-known. Economically, there's ~no moat, diffusion models aren't a mysterious untame-able force.

Anyone here use Mercury 2.0? Curious what your experience with the model is.

I used it for testing my voice agent. It was basically what I expected. Good fast model but "generic" or "vanilla" is how i would describe its personality emulation capability as. Gemma models still outperform it in that department. As far as technicals, one thing i found annoying is cash use was not that good, it missed more then i liked, i contacted support and they were fast and responsive and said they were working on that issue, maybe they solved it with 2.5? Anyways, im prolly gonna try 2.5 again see if anything different, but cant deny the speed, thats the biggest thing this company has going for this offering as if you are in the business of classical cascaded voice agent systems, latency is number one priority and this thing is fast....

latency, instruction adherence, reliable tool cools, conversationality are all in tension.

it's great when you can get a 170ms ttft. but if you have 700 ms endpointing on the stt side and 300ms ttfb on the voice side, then you haven't really made something super snappy.

Do diffusion models let me specify how long i want the answer beforehand? That seems like it could be useful.

I like the text output on logical and historical content that I sampled so far

This token speed is too fast.

the output is good and fast. like it, not only fast, but a new architectures.

I love this it is so fucking fast!

I thought this was going to be about a boat motor.

I thought it was going to be about Mercury[1] the programming language, and was confused about the domain name.

[1] - https://www.mercurylang.org/

Oh great a new model annoncemzzzzzz ZZZZZZZZZ

Not just a model - It's a diffusion model. Instead of Next Token prediction it builds the entire page at once and then denoises it. Kind of mind blowing when you watch it happen.

I would have missed this, thanks

Ok well that is actually interesting, thank you. Instead of OpenAI beta 6.3pre3 “fairyfloss”

Where can you watch it happen? Is there a video visualizing the diffusion process on text?

This is an older post from June about a Google model, but the same general process applies and they have a couple demo videos:

https://blog.google/innovation-and-ai/technology/developers-...

Just go type questions into the playground - leave "diffusion effect" on - you can see it for every question. https://chat.inceptionlabs.ai/

I think we should be happy when this happens as it encourages more players to be active in AI, and us not rely on only two companies

Reply on news.ycombinator.com