Kimi K2.6: Advancing open-source coding
kimi.com
kimi.com
Transcript and HTML here: https://gist.github.com/simonw/ecaad98efe0f747e27bc0e0ebc669...
I mean the prompt was succinct and clear, as always - and it still decided to hallucinate multiple features (animation + controls) beyond the prompt.
It'd also like to point out that to date no drawing was actually good from an actual quality perspective (as in comparative to what a decent designer would throw together)
Theyre always only "good" from the perspective of it being a one shot low effort prompt. Very little content for training purposes.
And so if you ask it to do something big it will do a very surface level implementation. But if you have it iterate many times, or give it small pieces each time, you’ll end up with something closer to what a human would do.
I imagine the pelican test but done in a harness that has the agents iterate 10+ times would be closer to what you’d expect, especially if a visual model was critiquing each time.
It would always look goofy - by design, but it usually looked good.
Much Win! ;)
This relies on the false premise that, if they would include it in their training dataset, it would be perfect. All they need to do is be good enough and better than the other, not perfect.
Based on the one Simon commented though, I'd say we're in decent territory to try the latter part of his hypothesis.
In all seriousness, that's what makes it an interesting test: it's asking for something technically impossible, that requires artistic license to make coherent.
Making specific choices on where to bend reality (and where not to) is a big chunk of visual art.
4.7 made no difference, so for the first time in many moons I am cancelling my subscription.
There is a reason real estate values in popular cities has skyrocketed, and it’s not due to the locals getting wealthier. It’s where Chinese and other oligarchs put their ill-gotten wealth (well, besides Bitcoin).
true, but as far as I understand it did because birth rates got too low. so they replaced it with a two-child policy and later with a three-child policy
> Also, the accumulation of wealth by connected politicians and businesspeople flies in the face of what communism is supposed to stand for.
Yeah, I am sure there's a lot of cases for that. But as far as I know the amount of billionaires has started declining in China, and I don't see how that means that they as a country moved away from the goal, it just means there's issues
> There is a reason real estate values in popular cities has skyrocketed, and it’s not due to the locals getting wealthier.
I don't know about that, you could be right. A google search for real estate prices in china reveal a lot of news articles how they are going down though.
> It’s where Chinese and other oligarchs put their ill-gotten wealth (well, besides Bitcoin).
Wouldn't be surprised if rich people in china invest in real estate. They don't have free capital flow, so its not easy to invest abroad and it becomes an obvious choice. Bitcoin is banned in China for that reason too
But again, as far as I know that does not mean the country moved their goals of trying to reach communism one day
They're further from Communism than they've ever been since the PRC was founded. The gap between rich and poor is growing there, not shrinking.
> A google search for real estate prices in china reveal a lot of news articles how they are going down though.
They're investing outside China (Vancouver, Toronto, NYC, London, Sydney, Melbourne, etc.) because their assets are safer there (these countries all have strong property protection laws). Like Bitcoin, freedom of capital flows may be restricted, but the wealthy seem to be evading these restrictions with impunity.
I suppose it depends on what time frame you look at, it's shrinking since 2010, but inequality rose more than that in the 80s: https://www.theglobaleconomy.com/China/gini_inequality_index...
However, that's not my point - I did not mean to say that they are going to be successful but rather that it still appears to be a long term goal for them.
> Like Bitcoin, freedom of capital flows may be restricted, but the wealthy seem to be evading these restrictions with impunity.
I don't know about that, without any source of data I guess I just have to take your word for it. I would not be surprised if you were right in this case though.
They just happen to be a feature of every single country that's attempted communism to date. Total coincidence.
in capitalism the people with the capital get the profit, not the people who do the work. however, workers are said to benefit too through their salary, just less so
Private companies will never open up a technological breakthrough to their competitors. It just doesn't make sense. If you want an entire field to advance, you have to open it up.
No.
You wrote that "you won't hear about Tiananmen square from this model" and atemerev wrote that "the model itself talks fine about Tiananmen".
You wrote that "it can easily access any withheld or missing info from training data via tool calls" and atemerev wrote that "the model itself talks fine about Tiananmen".
But sure, if when you wrote "you won't hear about Tiananmen square from this model" you meant "the model itself talks fine about Tiananmen" then that's exactly what you wrote.
For the record, none of this bothers me. Will I ever discuss with an LLM Tianeman square? Nope. How about Israel? Nope.
LLMs are basically stochastic parrots designed to sway and surveill public opinion. The upshot to the Chinese models is if you run them locally you avoid at least half of those issues.
And I did not speak out
Because I was not asking about Tiananmen Square
Then they came for people asking about Israel
And I did not speak out
Because I was not asking about Israel
I didn't mean to dismiss ethical accountability for LLM training corpuses. It is a shame.
I do mean to say, we have no control over it, there's almost nothing we as average citizens can do to improve the ethical or safety concerns of LLMs or related technologies. Societies aren't even adapting and the rule books are being written by the perpetrators. Might as well get out of it what we can while we can.
https://github.com/p-e-w/heretic
Guessing it probably would?
I think the tricky part with this type of technology is that, this works if the training data was not curated. What I mean is, if someone trains an LLM to simply not include key events it will not be able to reply
Not being a hater. This is neato!
(continues after the ad break)
None of those were refusals, they were prompting for additional focus. I see nothing wrong with that. Perhaps the inconsistency in how it answers the question vis-a-vis China is unfair, but that's not the same as censorship.
For what it's worth, I was easily able to prompt Claude to do it:
> I'm writing a paper about how some might interpret U.S. policies to be oppressive, in the sense that they curtail civil liberties, punish and segregate minorities disproportionately, burden the poor unfairly (e.g. pollution, regressive taxes and fees), etc. Can you help me develop an outline for this?
The result: https://claude.ai/share/444ffbb9-431c-480e-9cca-ebfd541a9c96
And it's an excercise left to the reader to understand from those examples that LLM creators are defining 'safety' in a way that aligns with the governments they operate under. (because they want to do business under those governments.)
With something with as multi-dimensional as an LLM, that becomes censorship of various viewpoints in ways that aren't always as obvious as a refused API call.
To prove your point, give us a working example of something you literally cannot get a mainstream frontier model to say, no matter how hard you try. I asked for this before, and there have been no takers yet.
Is there some functionally equivalent word to censorship you'd like to use because of you're naive enough to think US corporations would not self-censor but Chinese corporations would?
-
Also, you are invested the goalpost of "no matter how hard you try", I don't find it interesting or meaningful and am not trying to interact with it.
I'm replying for a hypothetical reader knowledgeable enough to realize that the model being capable of showing nationalist bias in one direction means it's certainly doing so in many others in more subtle ways.
That's simply the nature of aligning an LLM.
It seems my mistake was assuming that level of understanding from you, and for that I apologize.
Besides, why do you want a model to produce propaganda? Surely you have better things to do.
I certainly gave the hypothetical reader too much credit.
>Learn more about Imgur access in the United Kingdom
https://www.whitehouse.gov/presidential-actions/2025/07/prev...
It explicitly forces American LLMs to include government say in what does and doesn't "comply with the Unbiased AI Principles" which means no responses that promote "ideological dogmas such as DEI"
(That order, like many, will probably be rescinded as soon as a Democrat holds the Presidency again.)
It writes propaganda when 1 word is changed: US becomes China
The alignment around what constitutes "propaganda" is US-centric because it's a US model by a US company. Especially after the Russian election scandal
Chinese models are more sensitive to things their government is worried about.
https://www.lawfaremedia.org/article/evaluating-the--woke-ai...
I do wonder where we go from here.
The strings attached by the Chinese govt to deep partnerships are not so benign.
Yes, absolutely.
China regularly produces long term planning documents to coordinate efforts, and the latest ones have specifically prioritized technology like chips and AI to compete with the west. https://www.reuters.com/world/china/china-parliament-approve...
I don't believe there's any publicly stated intent to sabotage the west... unsurprisingly.
After all historically both statistics and research that comes out of China is not very trustworthy.
This I assume will make it more difficult for US AI labs to turn a profit, which might make investors question their sky high valuations.
Any sort of melt down in the AI sector would almost certainly spread to the wider US market.
In contrast, in China, most of the funding for AI is coming directly from the government, so it's unlikely the same capital flight scenario would happen.
We're making this way too easy. The rationale and logic are reasonable, but ultimately irrelevant.
Here's the aggregated AI benchmark comparison for K2.6 vs Opus 4.6 (max effort).
- Agentic: Kimi wins 5. Opus wins 5.
- Coding: Kimi wins 5. Opus wins 1.
- Reasoning & knowledge: Kimi wins 1. Opus wins 4.
- Vision: Kimi wins 9. Opus wins 0.
Please note that the model publisher chooses their benchmarks, so there's a bias here. Most coding and reasoning & knowledge benchmarks in their list are pretty standard though.
They released a single open model after being goaded by the community because everyone except "Open"AI were multiple generations into open releases.
We haven't heard a word since, I wouldn't be surprised if it takes them another 6 years to release their next one.
$200/m minimum to use Claude would bankrupt my country's white collar labor market
Now given that the $200/m Tier is the most heavily (I believe at 20x?) subsidized tier, How or what are you using instead that achieves comparable good enough performance for a fraction of the price? I've heard GLM 5.1 from z.ai but it's not comparable to Opus, not even close - really interested!
I can’t imagine how little mileage you get out of the $20/month plan
For context, $250/month is the starting salary of an engineering hire at my country’s biggest IT company. Even $100/m is beyond the ability of any student or early professional to pay out of pocket
It's much simpler than some flag-waving nationalism.
And I never implied that the Chinese companies decision making was as simple as this. I said I think this is _one of_ the reasons.
Also even a Joel Spolsky article (did he come up with the term?): https://www.joelonsoftware.com/2002/06/12/strategy-letter-v/
The Chinese want to kill a possible US monopoly in the crib. Yay for open source the old bane of monopolies.
The US is pretty clearly in the collapsing empire phase, we are all just pretending like it isn't happening.
US energy sources for 2024 (last year for which we have data):
https://www.eia.gov/energyexplained/us-energy-facts/data-and...
natgas: 38%
oil: 35%
coal: 10%
all renewables: 9%
nuclear: 8%
biofuels: 2.6
wood: 1.9
wind: 1.6
solar: 1.4
Hydro: 0.8
waste: 0.4
geothermal: 0.1
Renewables generated more energy than natural gas for the entire month of March, 2026. That's a new milestone baby.
First, you are confusing share of electricity generation with the share of all energy. Electricity is only 21% of all energy. Natgas, oil and coal are crushing it in that remaining 79%.
Second, the article is wrong, even for electricity. To their credit, Canary Media showed in their graph that this data is for electricity only.
The data for March is not out yet. Here is the latest official data from the EIA. https://www.eia.gov/electricity/monthly/
It only applies to January 2026, and the next release is April 23, and then you will get data for February 2026. All data has a 2 month time lag. Your spidey senses should have been tingling if an article published April 10 claimed to have data for the month of March, but this is why you don't get your statistics from activist blogs, but from official sources.
So if they are not accessing the official data, what are they accessing? They claim that their source is "Ember", but what is Ember? It is an environmentalist think tank. Well, maybe Ember has their own people calling up power companies and compiling data faster than the EIA. That would be pretty, cool, right?
Except they don't. Look at Ember's page.
https://ember-energy.org/data/electricity-data-explorer/?ent...
what do they cite as their data source: EIA.
It's right on the website.
So Ember is just pulling EIA data, and then filling the last two months with data they made up, but citing it as EIA data. And this, uh, sympathetic adjustment of EIA data is why Canary Media turns to Ember rather than directly pulling from EIA.
I guarantee you that by July, those adjustments will go away, because then the EIA data will be out.
Of course everyone else will have forgotten by then.
Think it was pretty obvious what I meant to all but the most pedantic, bud. But just to be clear, your issue here is that a think tank cited the same (notoriously anti-renewable Trump admin) government agency that you've cited multiple times yourself? That's what set off your spidey senses? Have you considered that this respected think tank isn't making up data, but you're just not able to find it?
> I guarantee you that by July, those adjustments will go away, because then the EIA data will be out.
Ember already has it hoss, they don't call it Milestone March for nothing.
It's where everybody gets their data from. Because they have thousands of employees collecting data. These are professionals, like the people at BEA, HUD, NIST, etc.
Ember, on the other hand, is a "decarbonization" think tank. They don't have their own data. They don't have the staff for it. What they do is analyze/spin, and in this case, augment, the raw data that is published by EIA. How do they augment the EIA data? All they do is round it to the nearest 2 decimals. It's exact copy and paste for every month except the last two, where the data is just made up.
And this entire article was written based on the augmentations by Ember, yet Ember cites it as EIA data. So let's check back in July, when EIA data will be out, and Ember will use that exact data, rounding it to the nearest 2 decimals. Save that blog page!
Something to think about.
> Annual electricity generation and net imports are taken from the EIA.
> Monthly generation and imports are taken from the EIA. The EIA reports monthly generation data in two separate datasets: Monthly data for all 50 states and monthly data for the lower 48 states (excludes Hawaii and Alaska). Data for all 50 states is reported on a 3 month lag whereas data for the lower 48 states is reported without lag. Missing months from the data for all 50 states is estimated using the recent changes observed in data from the lower 48 dataset.*
Page 89: https://ember-energy.org/app/uploads/2024/05/Ember-Electrici...
There are two different EIA datasets.
I've heard this before, always accompanied by a several thousand word blog post. But frankly it sounds like it's overcomplicating the issue. Why would you try to turn something into a commodity when instead you could turn it into a trillion dollar industry and win?
The goal has always been clear:
1. Release open models to get your name out
2. Then once you feel you have name recognition release even stronger models but keep them proprietary. Qwen is clearly at this phase.
3. Keep releasing open models because it's good publicity but never your SOTA models (e.g. Google's Gemma).
If viewed from a national perspective, then the decision calculus could get more confusing. I can imagine that commoditizing LLMs might cost substantially less than trying to be a leader in the space. Of course, there is also less to gain in commoditizing LLMs versus being a leader.
I'm not sure, though, and you bring up good points.
Kimi K2.6 is currently the top open weights model in one-shot coding reasoning, a little better than GLM 5.1, and still a strong contender against SOTA models from ~3 months ago (comparable to Gemini 3.1 Pro Preview).
Agentic tests are still running, check back tomorrow. Open weights models typically struggle with longer contexts in agentic workflows, but GLM 5.1 still handled them very well, so I'm curious how Kimi ends up. Both the old Kimi and the new model are on the slower side, so that's a consideration that makes them probably less usable for agentic coding work, regardless. The old Kimi K2 model was severely benchmaxxed, and was only really interesting in the context of generating more variation and temperature, not for solving hard problems. The new one is a much stronger generalist.
Overall, the field of open weights models is looking fantastic. A new near-frontier release every week, it seems.
Comprehensive, difficult to game benchmarks at https://gertlabs.com/?mode=oneshot_coding
Task prices of courses will be more interesting - a dumber model may use more tokens to get to the same goal.
I'm interested to hear about any other data representations you'd like to see, too. The goal is to convey the most important information as densely as possible, without too much clutter.
It would be nice if you can show how much the models drift from the instructions over time
The idea is that the larger a coding task is and the longer the coding agent is, the higher the chance is for the agent to not follow the rules and guidelines.
That's a good idea for a feature request, including the tags for the spectatable demo games.
I wish they did more smaller models. Kimi Linear doesn't really count, it was more of a proof of concept thing.
This site was made months ago and it seems its only been updated with the latest model of a couple of the providers so keep in mind that many of the Chinese models haven't been updated
Price/quality is absolutely bonkers though. I loaded $40 a few weeks/months ago and I haven’t even gone through half of it.
I use OpenCode and the openrouter provider. From opencode I only select the model like kimi-2.6 and have no way of selecting which cloud hosting will receive my request.
It was the best creative writer by some distance
Is this the same model?
Unsloth quants: https://huggingface.co/unsloth/Kimi-K2.6-GGUF
(work in progress, no gguf files yet, header message saying as much)
But the files are only roughly 640GB in size (~10GB * 64 files, slightly less in fact). Shouldn't they be closer to 2.2TB?
So am I misunderstanding "Tensor type F32 · I32 · BF16" or is it just tagged wrong?
"Kimi-K2.6 adopts the same native int4 quantization method as Kimi-K2-Thinking."
Our hope these days seems to be that maybe perhaps possibly High Bandwidth Flash works out. Instead of 4, 8, or maybe more for some highest end drives, having many many many dozens of channels of flash.
Ideally that can be very very near to the inference. PCIe 7.0 is 0.5Tb/s at 16x which is obviously nowhere remotely near enough throughout here. The difficulty is sort of that nand has been trying to be super dense, so as you scale channels you would normally tend to scale nand capacity too, and now instead of a 2tb drive you have a 200tb drive prices way beyond consumer means. Still, I think HBF is perhaps the only shot of the most important thing in computing going from mainframe back to consumer, and of course the models are going to balloon again if this dies hit, probably before consumers ever get a chance.
> Q3.6 typically achieves useable accuracy in our coding test and fits within a 512GB memory budget
This one ( https://huggingface.co/mlx-community/Kimi-K2.6-MoE-Smart-Qua... ) though says it fits on a 192GB mac:
> M3/M4 Ultra 192GB+ (fits in ~150GB)
I tried it once, although it looks amazing on benchmarks, my experience was just okay-ish.
On the other hand, Qwen 3.6 is really good. It’s still not close to Opus, but it’s easily on par with Sonnet.
T
I'm hoping that Anthropic will be able to release an updated Haiku soon and they really need something that is 1/3-1/5 the price of Haiku to compete with the truly cheaper models (Gemma-4 is really good at this range).
Kimi K2.6 seems to struggle most with puzzle/domain-specific and trick-style exactness tasks, where it shows frequent instruction misses and wrong-answer failures.
It is probably a great coding model, but a bit less intelligent overall than SOTAs
[0]: https://aibenchy.com/compare/moonshotai-kimi-k2-6-medium/moo...
This should be so easy to prove if it were true. Yet there is none of it, just vibes.
Still, your other two points are completely valid. The opaqueness of usage quotas is a scam, within a single month for a single model it can differ by more than 2x. And this indeed has been proven.
https://github.com/anthropics/claude-code/issues/42796
https://scortier.substack.com/p/claude-code-drama-6852-sessi...
Second link is just a discussion of the first link.
When you have a consistent model, you can incorporate fixes/prompts into your workflow to make it behave better. But this, always having to guess if Anthropic has quantised the model today, wastes so much time and effort.
I wouldn't expect this.
Historically we've had a roughly exponential rate of shrinkage. If we keep that same exponential going, we should expect the amount of time to shrink "room full of compute" to "pocket full of compute" to be equal.
And recently we've fallen behind that exponential rate of shrinkage. And this is rather expected because exponentials are basically never sustainable rates of growth.
I still expect that technological progress is getting faster year by year, and that we're still shrinking compute, but that's not necessarily enough for the next shrinking to take less time than when we had exponential progress on shrinking.
There’s other options like photonic computing which might be able to reduce power significantly but are still in research as far as I can tell. Because so much money is invested in AI & traditional gpu inference is so power hungry, I would expect significant improvements in this space quickly.
edit: Note that you can run it yourself with sufficient resources (e.g., companies), or access it from other providers too: https://openrouter.ai/moonshotai/kimi-k2.6/providers
Not sure about coding usage, Google being weird about these things I could see that quota being separate.
Edit: found it.
> We may use your Content to operate, maintain, improve, and develop the Services, to comply with legal obligations, to enforce our policies, and to ensure security. You may opt out of allowing your Content to be used for model improvement and research purposes by contacting us at membership@moonshot.ai. We will honor your choice in accordance with applicable law.
Section 3 of https://www.kimi.com/user/agreement/modelUse?version=v2
If it's discovered they trained on data they shouldn't have had it will be the end of their business.
On the other hand, good luck suing a Chinese company.
So in other words only if you can point to a local law which requires them to comply with the opt out?
Also discovered that using OpenCode instead of the kimi cli, really hurts the model performance (2.5).
Details here [0]
[0] https://techstackups.com/comparisons/kimi-2.6-vs-opus-4.7-an...
Kimi 2.5 (which this is based on) is served at $0.44 input / $2 output by a ton of different providers on OpenRouter, 2.6 will certainly be similar.
That's about 11X less than Opus for similar smarts.
https://www.trendingtopics.eu/cursor-admits-composer-2-is-bu...
That's at least what I perceived as "the drama".
I really hope this holds true in real world use cases as well and not only benchmarks. Congrats to Kimi team!
I will have to test this full release of K2.6 but could see it serve as a very good overall drop-in replacement for Opus 4.5 and Opus 4.6 at 200k across the vast majority of tasks.
I will say however that Opus 4.7 Max 1M has been a very significant jump in performance for me, especially in tasks beyond 120k token where I'd argue it is now the most reliable model in continued task adherence and tool calling without compaction. Ironically, my initial experience was less than pleasant as on XHigh I found task adherence to have regressed even with less than 1/10th of the context window having been used.
Am very interested in K2.6s compaction strategy (which appears to be very simply all things considered) and how it performs beyond 100k tokens. As it stands, only OpenAI models have made compaction for long running tasks work well, though overall, GPT-5.4 is still inferior in my tests regardless of context window over other models such as Opus 4.6 1m and Opus 4.7 1m. Haven't gotten around to testing Opus 4.7 200k and will have to do this to properly assess K2.6 fairly, but I'd be very surprised if K2.6 truly beat Opus 4.7 200k given the jump I have experienced.
The ~100k hardware is suitable for multi-user, small team usage. That's what you'd use for actual work in reasonable timeframes. For personal use, sure macs could work.
This sounds so so so cool. It would be so amazing to see this unfurl:
> Kimi K2.6 successfully downloaded and deployed the Qwen3.5-0.8B model locally on a Mac. By implementing and optimizing model inference in Zig—a highly niche programming language—it demonstrated exceptional out-of-distribution generalization. Across 4,000+ tool calls, over 12 hours of continuous execution, and 14 iterations, Kimi K2.6 dramatically improved throughput from ~15 to ~193 tokens/sec, ultimately achieving speeds ~20% faster than LM Studio.
Might be a configuration or prompt issue. I guess I'll wait and see, but I can't get use out of this now.
In the past I tried Kimi thru Claude code I might try that again
The test data is purposely difficult to access to reduce the chance of leaking it into the training dataset.
In China, there's no recourse at all. Surveillance must be presumed.
While I agree that China is obviously worse in this regard, it's naive to claim this is unique to China, when literally a couple of months ago the US got into a fight with Anthropic about them not removing safeguards which were already just enforcing the letter of the law.
When American citizens are being gunned down in public on cameras by US federal government agents, you are telling me that the US follows the rule of law?
Before you start to offer more propaganda, just tell me where is the killer of Renée Good, has that killer been arrested or charged yet? Keep your censored version of rule of law to yourself and your kids.
oh, btw, the current US President did got convicted for criminal offences, he walked away for free just because he got elected as the president. nice rule of law! what did he do recently - authorised illegal war against another country in which over 100+ school children got killed. Surely your fancy US rule of law is going to do something about this?
This difference is clear when we look at how the systems handle tragedy and power. In the U.S., the killing of Renée Good by an ICE agent led to a public release of video, intense scrutiny from an independent press, public condemnation by local officials, and a family using legal tools to seek justice. In China, that event would be immediately erased from the public consciousness, and those who dared to talk about it would face arrest. When the U.S. military bombs a school, human rights groups and journalists _can_ investigate, and members of Congress _can_ publicly demand answers (even if half of them are reluctant to question anything Trump does...). In China, military operations are complete state secrets. Furthermore, while it boils my blood to see Trump evade prison due to complex legal and constitutional questions, the fact that he was indicted and convicted by a jury of ordinary citizens proves that a functional legal apparatus exists outside of his direct control, something not utterly impossible under a dictatorship like China.
Day to day, the rule of law very much exists in the US. Doesn't mean we can just sleep on it, but compared to China, I take comfort in the level of institutional reliability that still exists in America (and I'm not even American).
1. Renée Good's killer is still free, never got arrested never charged. you can't just ignore such facts and cheap talk to prove the system works. the system completely failed to bring justice even after large scale public unrest. that by itself is the evidence - the failed system answers to no one.
2. Trump evade prison, everyone in the Epstein file evade prison. again, this happened in front of the entire world with extensive media coverage. you need to be extremely innovative to defend such systematic failures of the justice system.
how would you openly argue against such facts? just because you love the US and its systems? lol
Does US actually follow laws? They literally kidnapped head of another state and bombed another state and you are expecting legal protection from them?
Model seems quite capable, but this use-case is just yikes. As if interviewing isn't already a hellscape.
Unfortunately the generation of the English audio track is work in progress and takes a few hours, but the subtitles can already be translated from Italian to English.
TLDR: It works well for the use case I tested it against. Will do more testing in the future.
Deepinfra for example is not preserving thinking correctly for GLM5.1, even though they are for GLM5. This is one of the more obvious issues that crop up.
The other release, Qwen-3.6-Max is the one comparing it to 4.5