Ohmygoodness, the results look amazing. I felt so sad for them when their first weights dropped, a lot of other great models came out around then and they didn't get the chance to shine even if for a week or 2. But looking at these results, this is really good stuff and a decent alternative to DSv4.1-F1017. Also the best and most detailed comparison I have seen from a lab across models.
The best open models today have been “good enough” to do most knowledge work for a few months. This might be the first US one good enough too. That’s a big deal.
MoA schemes that use proposers/task-separation/supervisory roles can still benefit from very fast diffusion models like mercury, the overall speed doesn't have to be bottle-necked by the slowest model -- that depends on the arrangement it's deployed in.
but to your point , yeah diffusion models can be kind of brain dead without proper coaxing.
my ideal arrangement is a smarter/bigger model supervising lots of proposals from diffusion models. The diffusion models spit out a big proposal each, the bigger model can vet something as minor as the headline or filename for each proposal, time still ends up being saved.
It's not as good as the current frontier, to be fair. I don't use it to help me think through architecture, for example. But I have found to work well for at least like half of coding tasks once architecture and patterns are in place. Its advantage is that it's very very fast.
Sorry to put it bluntly, but, Mercury is sh*t in comparison. And I have genuinely tried a few times to use it since the speed is so good. But nop.. it just isn't there.
Things have been remarkably "not terrible" since approximately the release of Laguna S 2.1 a while back in early summer 2026, which is now outdated within the past 6 month time frame by other, better non CN open weight models.
I'm not saying that any of them are to the level as a deepseek v4.1 flash 1017 or similar but they're catching up. The most recent huge mistral release is also grounds for an optimistic view.
If you cant beat Deepseek on price im not using it. I actually dont care about beating DS in capabilities, I dont think you need more intelligence than Deepseek for 99% of the things people work on, and I dont want LLMs to do everything, that kind of defeats the point. I just need a smart fuzzer and to offload some tedious busy work.
So why should I pay for 3-20x more for a few small percentage points in preformance. (I would prefer DS did some things better, but largely I can force my desired behavior), but I could only have access to Deepseek 4.1 and still beat almost everyone using opus/fable/astra on a productivity basis, I need automation tools not "do my entire job so Im worthless tools", and DS has already more than gotten there for me.
It doesn't beat DS v4.1, and they have so much more compute capacity than any Chinese company, yet still cannot even beat MiniMax M3.1 Flash (220B) despite having 2.7x more parameters (596B).
The number HF calculates from the safetensors metadata is often wrong (e.g. counting u8-packed 4-bit weights as half a weight each) so you need to take the number from the actual model card.
If you’re going to do the cost math that way, be prepared to put a price on ~the total sum of all works created by humanity, which the frontier labs appropriated for their models.
In their last technical report (for the original Inkling model), they mentioned 'To bootstrap post-training, we ran an initial SFT on synthetic data generated by open-weights models including Kimi K2.5. The bootstrap accounts for a small fraction of compute, with the majority being employed for large-scale RL on synthetic and human-created environments.' and in their most recent one they said 'The H3 and H4 mixtures consisted mostly of trajectories sampled from earlier models including Kimi K3 and our own models, filtered for correctness and alignment.' So it does seem like traces from larger models do contribute. I think Microsoft is one of the few companies that has actively made an effort to avoid distillation with their MAI series.
The most important thing here is training was entirely nvfp4 - this is best model by far that has ever been trained in that precision and explains why they were able to do it so fast
Very refreshing to see benchmarks where the reporting team is not the clear winner. They are claiming to be better than GLM 5.3 flash, which is impressive if it holds up outside of benchmarks.
Given the time to market and the date in which these labs were founded, I find it far more likely that the lag was representative of them establishing proper internal (RL etc) tooling.
Would this have been SOTA at the beginning of the year? Yes. Where you happy with SOTA in january? Why can't we then be happy with this? I mean, perhaps Inkling 3 or 4 Small will crush Opus 5.5 and then we will ask if it matters since it's not Opus 7... Everyone can't be SOTA. With these open weight models, I simply ask. Is it useable, is it useful? Can I get very useful work out of it at a reasonable cost to other open models and cloud models? If the answer is yes, then it's a go.
This isn't small at all, for most budgets outside of enterprise, and if you have an enterprise budget why not throw a few more machines in and run a larger and better model?
I feel like this is a dumb question, but why are we spending any effort on wine comparisons? Is there a way to get rid of experts on these to just be ones for code?
It's not like they're training on that. That's just a good showcase of how the model is good at generalizing & understanding (not benchmaxxed). You might want something similar to these experts for tasks like detecting anomalies or discoloration or whatever.
Just looking through the benchmarks, the model seems to be very high on visual/spatial capabilities. So I assumed they just tried some hard tasks on it and was surprised by wine identification and thought it was surprising that it could identify wines.
you want your coding models to have good taste, so we have fed them a vast dataset of sommelier wine gradings.
we've also contracted hundreds of local record store guys with deeply knowledgeable music taste spanning post-punk, classic jazz, early hip-hop, and deep-cut experimental vinyl. This has resulted in the model returning judgmental responses up to 23% more often which we are working on addressing.
Mistral 4 IMHO is not that good. I have run the full NVFP4 weight, and I couldn't stomach it after 3 days of driving it. I went back to other models, I have more fun bouncing between DeepSeek, GLM, MiMo, Hy4, MiniMax & Qwen. If Thinkingmachine's self reported benchmark is true, I'm certain it's going to get into heavy rotation of my models.
I played around with Le Chonk on a toy project and it was ... very okay. Felt like a throwback to the Opus 4 era. It committed when I had told it not to, and it spent a lot of time arguing with itself going in circles.
I feel like the gap between Opus and Grok is about the same as the gap between Grok and Mistral. I didn't revert its code or anything, but I also didn't refill the initial OpenRouter budget when it ran out.
At the same time, I guess Opus 4 level is pretty good for a Western open weights model. 2027 is going to be amazing.
> We even tested Inkling 2 Small against some of our researchers in a wine tasting event. It outperformed them using vision alone, correctly identifying three wines from a single picture in 19 out of 20 trials.
So the legendary skill of Charles Duchemin is now widely available? What a time to be alive.
You could probably run a Q2-Q3 with 2x DGX Sparks. So...$12k at today's prices. Plenty of folks are running DS 4.1 Flash with Strata and some GPU + bulk in system memory. Maybe more like $3k-$4k? But speeds come down quite a bit.
Do you mean to ask why non-American companies would trust anything coming out of the USA? Because American companies tend to have a lot of trust for other American companies.
Even when models score about the same, if you play with them long enough, you will discover capabilities that are unique to them. I have never see any 2 models that are comparable across every single area.
I don’t like it. Poor form to squat on an old companies name like that, one which has a rich legacy in AI. I know the trademark was cancelled so all I can say is I don’t like it.
Looks awesome, but dang a bit too big for my machine.
… I can’t wait for custom resizable models!
FWIW: like the Sims where you have certain levels where you move one up and others non-linearly go down e.g. Speed, Intelligence, Conciseness, determinism, trainability
Ohmygoodness, the results look amazing. I felt so sad for them when their first weights dropped, a lot of other great models came out around then and they didn't get the chance to shine even if for a week or 2. But looking at these results, this is really good stuff and a decent alternative to DSv4.1-F1017. Also the best and most detailed comparison I have seen from a lab across models.
The best open models today have been “good enough” to do most knowledge work for a few months. This might be the first US one good enough too. That’s a big deal.
Open weight US based Opus 5.5 within 12 months.
> Open weight US based Opus 5.5 within 12 months.
Seems unlikely. Maybe open weight at all near there? That'd be great.
Why does that seem unlikely? The difference between frontier and open weights has been much less than 12 months for a while now.
US open weights are much worse, way more than 12 months apart compared to frontier.
> Why does that seem unlikely? The difference between frontier and open weights has been much less than 12 months for a while now.
Frontier and open weights? Yes.
Frontier and _US_ open weights? No.
I'd expect a US open weights provider to suddenly get bought or go non-open if they get near Opus 5.5 anyway.
People say this about the frontier every time, and the open models catch up in a matter of months every time.
The open models catch up, the _US_ open models catch up much more slowly. Different incentives and cultures.
Inception Mercury 3 and Particle-1's model are also good.
(I am assuming Inception is an American company, not sure TBH)
Agree, Mercury 3 is one of my favorite models right now...the speed is such a quality of life improvement.
Really? I'm surprised. What do you use it for?
My read on Diffusion LLMs has been that it's cool tech, but only useful for things like like code autocomplete or editing a block of text.
They don't beat autoregressive models on quality per compute, assuming you are serving lots of users and can batch requests.
MoA schemes that use proposers/task-separation/supervisory roles can still benefit from very fast diffusion models like mercury, the overall speed doesn't have to be bottle-necked by the slowest model -- that depends on the arrangement it's deployed in.
but to your point , yeah diffusion models can be kind of brain dead without proper coaxing.
my ideal arrangement is a smarter/bigger model supervising lots of proposals from diffusion models. The diffusion models spit out a big proposal each, the bigger model can vet something as minor as the headline or filename for each proposal, time still ends up being saved.
I just use it for normal coding stuff.
It's not as good as the current frontier, to be fair. I don't use it to help me think through architecture, for example. But I have found to work well for at least like half of coding tasks once architecture and patterns are in place. Its advantage is that it's very very fast.
Sorry to put it bluntly, but, Mercury is sh*t in comparison. And I have genuinely tried a few times to use it since the speed is so good. But nop.. it just isn't there.
For my own purposes, I found Mercury Decide to be excellent. Better than Jev at what I threw at it. Inception did a really great job with it.
Thanks for your review, 15 day old HN account that likes praising Mercury services.
May I add Poolside (Laguna S) to this? Used them extensively during their free trial (sadly now apparently over).-
Muse Glimmer is pretty good.
Could you give some examples what makes them good, thanks
I give it 2.
That’d be the inflection point I’m waiting for!
Things have been remarkably "not terrible" since approximately the release of Laguna S 2.1 a while back in early summer 2026, which is now outdated within the past 6 month time frame by other, better non CN open weight models.
I'm not saying that any of them are to the level as a deepseek v4.1 flash 1017 or similar but they're catching up. The most recent huge mistral release is also grounds for an optimistic view.
It's a nice, domestic open weight model.
Without hundreds of millions in 9 days a small team is producing a model that can beat DeepSeek 4.1 and a platform to tune it.
Optimistic for the future.
Correctly if I’m wrong but isn’t inkling made by a very well funded startup from an ex - OpenAI exec iirc?
Yes, Mira Murati's Thinking Machines. They are apparently trying to raise $5B at $40B, after raising $2B at $10B last year.
If you cant beat Deepseek on price im not using it. I actually dont care about beating DS in capabilities, I dont think you need more intelligence than Deepseek for 99% of the things people work on, and I dont want LLMs to do everything, that kind of defeats the point. I just need a smart fuzzer and to offload some tedious busy work.
So why should I pay for 3-20x more for a few small percentage points in preformance. (I would prefer DS did some things better, but largely I can force my desired behavior), but I could only have access to Deepseek 4.1 and still beat almost everyone using opus/fable/astra on a productivity basis, I need automation tools not "do my entire job so Im worthless tools", and DS has already more than gotten there for me.
Not beating dsv4.1 at all
It doesn't beat DS v4.1, and they have so much more compute capacity than any Chinese company, yet still cannot even beat MiniMax M3.1 Flash (220B) despite having 2.7x more parameters (596B).
I wouldn't bet on the numbers of compute reported by Chinese labs. Those Singapore Nvidia orders had to go somewhere.
> yet still cannot even beat MiniMax M3.1 Flash (220B) despite having 2.7x more parameters
MiniMax M3.1 Flash is 428B total parameters, so the size ratio is more like 1.4x. Model card here: https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash
The number HF calculates from the safetensors metadata is often wrong (e.g. counting u8-packed 4-bit weights as half a weight each) so you need to take the number from the actual model card.
Are they distilling from the big labs or Chinese labs at all? In that case isn't it $1 trillion frontier CAPEX + Inkling's costs for the total costs.
"I bought this house with my own savings plus a small million dollar loan from mum and dad."
Everyone is building on everything that comes before.
Houses are not built from scratch either. The know-how, power, water, roads, materials…
Yes, and the sum total of those costs (that makes _some_ economic sense) end up on the total house price.
Is the same true for distillation?
If you’re going to do the cost math that way, be prepared to put a price on ~the total sum of all works created by humanity, which the frontier labs appropriated for their models.
In their last technical report (for the original Inkling model), they mentioned 'To bootstrap post-training, we ran an initial SFT on synthetic data generated by open-weights models including Kimi K2.5. The bootstrap accounts for a small fraction of compute, with the majority being employed for large-scale RL on synthetic and human-created environments.' and in their most recent one they said 'The H3 and H4 mixtures consisted mostly of trajectories sampled from earlier models including Kimi K3 and our own models, filtered for correctness and alignment.' So it does seem like traces from larger models do contribute. I think Microsoft is one of the few companies that has actively made an effort to avoid distillation with their MAI series.
domestic -> people live in antarctica. No domestic LLM for them.
No domestic good LLM for me, either.
The most important thing here is training was entirely nvfp4 - this is best model by far that has ever been trained in that precision and explains why they were able to do it so fast
Oh nice they’ll run well on sparks
edit: nm I see its a bit big
Very refreshing to see benchmarks where the reporting team is not the clear winner. They are claiming to be better than GLM 5.3 flash, which is impressive if it holds up outside of benchmarks.
Is it SOTA yet? No. Are the domestic open weights labs demonstrating incredible velocity? Undeniably.
Distillation is the tide that lifts all boats
Imagine if Mira Murati and John Schulman actually know what they're doing?
Given the time to market and the date in which these labs were founded, I find it far more likely that the lag was representative of them establishing proper internal (RL etc) tooling.
Seems to be pretty close to SOTA on design. I’m excited to try it out with my agentic design system.
Would this have been SOTA at the beginning of the year? Yes. Where you happy with SOTA in january? Why can't we then be happy with this? I mean, perhaps Inkling 3 or 4 Small will crush Opus 5.5 and then we will ask if it matters since it's not Opus 7... Everyone can't be SOTA. With these open weight models, I simply ask. Is it useable, is it useful? Can I get very useful work out of it at a reasonable cost to other open models and cloud models? If the answer is yes, then it's a go.
I can’t seem to find the cost? The craziness of deepseek flash 1017 is how ridiculously cheap it is for how capable it is.
AA lists it as $0.30/0.03/1 and cost per task is $1.18, about double that of deepseek
Wow, that's pretty good.
It’s about 2-3X more expensive so about the same token efficiency as 1017. I can see this being a good option for the US.
As someone who knows nothing about wine except I like drinking it - that wine test is fascinating.
Damn their last "Inkling Small" model was only 266B and now this is more than double.
This isn't small at all, for most budgets outside of enterprise, and if you have an enterprise budget why not throw a few more machines in and run a larger and better model?
This is amazing! Can't wait to trying this out. Also, crazy that I found this on LinkedIn with more upvotes and discussion than the HN thread has.
I feel like this is a dumb question, but why are we spending any effort on wine comparisons? Is there a way to get rid of experts on these to just be ones for code?
What if your coding project is related to wine? So you would want facts about wine to be inside the weights, not through web search.
If it was about wine, why would I want one with a training data cutoff? I'd much rather have one that can query a database about wine
It's not like they're training on that. That's just a good showcase of how the model is good at generalizing & understanding (not benchmaxxed). You might want something similar to these experts for tasks like detecting anomalies or discoloration or whatever.
Source they didn't train on it?
Just looking through the benchmarks, the model seems to be very high on visual/spatial capabilities. So I assumed they just tried some hard tasks on it and was surprised by wine identification and thought it was surprising that it could identify wines.
you want your coding models to have good taste, so we have fed them a vast dataset of sommelier wine gradings.
we've also contracted hundreds of local record store guys with deeply knowledgeable music taste spanning post-punk, classic jazz, early hip-hop, and deep-cut experimental vinyl. This has resulted in the model returning judgmental responses up to 23% more often which we are working on addressing.
Looks better than mistral 4 on all benchmarks ?
Mistral 4 IMHO is not that good. I have run the full NVFP4 weight, and I couldn't stomach it after 3 days of driving it. I went back to other models, I have more fun bouncing between DeepSeek, GLM, MiMo, Hy4, MiniMax & Qwen. If Thinkingmachine's self reported benchmark is true, I'm certain it's going to get into heavy rotation of my models.
I played around with Le Chonk on a toy project and it was ... very okay. Felt like a throwback to the Opus 4 era. It committed when I had told it not to, and it spent a lot of time arguing with itself going in circles.
I feel like the gap between Opus and Grok is about the same as the gap between Grok and Mistral. I didn't revert its code or anything, but I also didn't refill the initial OpenRouter budget when it ran out.
At the same time, I guess Opus 4 level is pretty good for a Western open weights model. 2027 is going to be amazing.
Interesting, I sent the same wine testing example to Claude Opus 5.5 medium, and it failed, where Inkling 2 Small succeed.
> We even tested Inkling 2 Small against some of our researchers in a wine tasting event. It outperformed them using vision alone, correctly identifying three wines from a single picture in 19 out of 20 trials.
So the legendary skill of Charles Duchemin is now widely available? What a time to be alive.
> Small: 594 billion parameters
cries in GPU racks
You could probably run a Q2-Q3 with 2x DGX Sparks. So...$12k at today's prices. Plenty of folks are running DS 4.1 Flash with Strata and some GPU + bulk in system memory. Maybe more like $3k-$4k? But speeds come down quite a bit.
Q2. Woof.
It actually benchmarks quite well. Honestly not practically worse than the Q4 native weights.
I couldn't pay attention on the content, i was playing with dots
They should buy the pleias team and just mog everyone.
I like the presentation on their website! Super well done
"Sorry, you have been blocked"
*Shrugs*
DeepSeek would never block me
So basically comparable to DS4.1? But why bother to make something we already have?
Because its american.
lol
Sovereign open source AI is a real thing, lots of companies want to own their own intelligence but also don't trust the models coming out of China.
Why would you trust anything coming out of USA?
The vast majority of users on this website wouldn't be trusting something coming out of the USA, they'd be trusting something in the USA.
Do you mean to ask why non-American companies would trust anything coming out of the USA? Because American companies tend to have a lot of trust for other American companies.
Why would you trust China on anything?
https://www.hirundo.io/blog/glm-revolut
I would hope that the American models aren't benchmaxxed to hell like all of the Chinese ones are.
Why?
Even when models score about the same, if you play with them long enough, you will discover capabilities that are unique to them. I have never see any 2 models that are comparable across every single area.
I don’t like it. Poor form to squat on an old companies name like that, one which has a rich legacy in AI. I know the trademark was cancelled so all I can say is I don’t like it.
Looks awesome, but dang a bit too big for my machine.
… I can’t wait for custom resizable models!
FWIW: like the Sims where you have certain levels where you move one up and others non-linearly go down e.g. Speed, Intelligence, Conciseness, determinism, trainability