I don't mind this in the agent workflows themselves, but I think this is hideous in the code and encourages agents to become even more verbose in the future.
I have a workflow to create PRs where at the end, I spin up a subagent that goes through the PR to 1) remove the bulk of the comments and 2) attempt to normalize it for human consumption (I ask it to optimize wording for junior developers and non-native speakers).
I use Haiku for it, hoping that a dumber model will do a better job. I have noticed in brief experimentation that even Sonnet will sympathize too much with the need of excessive comments and is not a good adverserial reviewer for slop comments.
After that workflow runs I still need to go in and manually correct, but it's a lot less than vanilla Opus. It really feels like the problem of slop comments has gotten worse as models has gotten smarter.
I run various forms of workflows to run dedicated QA, code review (of various flavors) simplification and text simplification agents. Especially the simplification goes a long way to remove dumb padding, duplication and efficiency. Dedicated docs/comment simplification is also becoming more and more necessary on recent models. For things like feature development in my workflow, the majority of time the agents run and tokens spent is critiquing the code from various perspectives and it's not close.
Of course, this doesn't solve the overall issue that agents don't write code like you and still requires a lot of human attention in planning and code review out to clean up leftover issues, and e.g. challenge bad assumptions about architecture and real-world context. A human is still very much needed to cull the slop (or, more gratuitously: align the agent). But IME it does help avoid a lot of pitfalls and makes the code high quality a lot more quickly.
Props to them for including three benchmarks that actually seem to say something, instead of focusing on totally gamed benchmarks like regular SWE-Bench. That could mean this model is actually pretty close to the SOTA as the benchmarks indicate.
Most labs - including OpenAI and Anthropic, but also Google and Chinese labs - highlight their scores in benchmarks that have fixed, widely available answers. Those answers end up in the training data and so models can just regurgitate training data instead of actually doing the benchmark. As a result, most benchmarks often quoted are essentially meaningless for gauging model performance.
Terminal-Bench still publishes answers, but neither DeepSWE and SWE-Bench Pro do. Especially for DeepSWE it's been difficult for models to fake good results so far. SWE-Bench Pro does have weird outliers like good performance for e.g. the atrocious Muse Spark, but it also doesn't provide answers for the training data.
So either they're good, or they found a way to game DeepSWE. Given that the Cursor team previously published the well-received Composer 2.5 a good score here doesn't come out of nowhere, so this might hold up. Cursor has enormous amounts of training data to train good coding models with.
You can put many agent constraints in precommit hooks if they're static checks. I ask agents to make commits, and e.g. in a Python project have the precommit hook fire off type checks, linting and even architectural things like import boundaries (using `tach`). When an agent is prepped to make commits themselves, it will catch pre-commit failing and correct itself. The existence of static checks themselves might also help agents gain awareness of the overall verification flow including larger things like tests, but that's hard to say for certain.
Putting structural code checks in a precommit hook is arguably better than pulling it into the harness, as it will enforce those constraints no matter whether an agent or human is making the commit.
The moat right now is model performance and what that means for how many tokens and additional time you spend.
I say this as a relatively frequent user of Kimi models and generally a big fan. But on not-yet-gamed benchmarks like DeepSWE, Kimi K2.6 is beaten soundly by Claude Sonnet 4.6 ($3 / $15) and even slightly by GPT 5.4 Mini ($0.75 / $4.50).
There's no question Kimi models are very good for a lot of code tasks. They're the best quality open weight model. But to get similar overall outcomes as on Sonnet/Opus, on average you'll spend many more tokens and will have to do more managing of the model. You shouldn't look at price per token, you should look at how much you pay for the entire process.
I'm more interested in how much effort I have to put in, at least while I'm paying in the range of current subscriptions (so ~€100-€200 a month or so). If the prices go up much more than that I'll have to switch to caring more about token efficiency. But at current pricing the bottleneck is my attention, not model efficiency. As such, even a small improvement in model quality - and hence, a decrease in how much attention I have to spend on it - makes a big difference.
I personally dont put any weight to DeepSWE. Other than 5.5 being directionally the best model, it gets the others pretty wrong in my experience. FrontierCode from cognition looks interesting
Any benchmark is iffy and has weird results, but this is the best we got at the moment. Most people working with Opus and Kimi would likely tell you they're much further apart than the numbers that were quoted for Kimi K2.6, and DeepSWE seems to capture that gap better.
One major thing DeepSWE has going for it is that all other benchmarks (including those quoted by MoonshotAI on this page) don't: the other benchmarks that are completely gamed. The benchmark answers are public and part of each model's training data. This benchmark may still be iffy, but at least it's not gamed.
American labs also use gamed and cherry-picked benchmarks extensively. Anthropic used them in their Fable announcement and avoided DeepSWE because it doesn't beat GPT-5.5 in that one. Google's numbers for Gemini 3.5 Flash recently did not at all line up with people's subjective experience using these models, and this also happened with Gemini 3.1 Pro before it.
Everybody has incentives to manipulate benchmark results to show their models in the best light.
It's for predictability in upgrades. Homebrew allows you to separate system packages (from apt or dnf) from user packages (from homebrew) [1]. Running apt upgrade or dnf upgrade can render your system unbootable if you're unlucky (or unstable or degraded if you're less unlucky). Running brew upgrade can at worst break some of your own user's setup or tools.
Since everybody runs their own unique permutation of apt or dnf packages, adding as little as possible will keep you as close as possible to what distro maintainers test. There's even OSes like Fedora Silverblue or Bluefin or SteamOS that ship with a fully baked _image_ - where installing system level packages is strongly discouraged - which helps ensure predictability and stable upgradeability.
Homebrew packages also tend to be more recent (this depends on your distro of course) and don't require elevated permissions to install.
[1]: Other unprivileged package managers like Mise or Nix do the same of course
There will be more of this going forward, I think. Systemd is really not just an init system, it's a full cohesive management system for Linux distros and they've never pretended otherwise. A modular one but still a comprehensive one. Because of that its mere existence is an affront to many people with traditional opinions on Linux and Unix.
systemd-appd sounds like it could make some inroads in the threat model that Windows and Linux still have in 2026 (and macOS is still reeling from): anything that runs as my user, can access anything running as _my_ user. I don't think this threat model was tenable in 2016, much less in 2026. But moving away from that also breaks with the Unix tradition.
Systemd as the system management layer is becoming a centerpoint for moving Linux forward, on servers but especially so on the desktop, and it does so at the cost of breaking with traditional views. It's kind of hard to watch: I want Linux to move forward, and there's just a lot of good ideas there. But it will be painful for a large Linux community to break with traditions.
> systemd-appd sounds like it could make some inroads in the threat model that Windows and Linux still have in 2026 (and macOS is still reeling from): anything that runs as my user, can access anything running as _my_ user. I don't think this threat model was tenable in 2016, much less in 2026.
Are we talking about open source operating systems hero or app delivery mechanisms?
> Systemd is not an init system: it's a full cohesive management system for Linux distros.
Exactly. If you look back at the old discussions, you see how people tried
to claim systemd is merely an init system, but it never was. So all comparisons to e. g. sysinit and what not, were unfair. Dishonest. The systemd devs were not interested in fair discussions. They wanted more control. And they very ruthlessly went forward with it - also thanks to corporate support. Just look at Poettering censoring discussions and stopping them whenever he could.
> But moving away from that also breaks with the Unix tradition.
Systemd never cared about UNIX. Poettering does not even understand UNIX on top of that.
> Systemd as the system management layer is becoming a centerpoint for moving Linux forward
Forward to ...? I don't really see it as moving "forward". I see it as more top-down control singularized into one crew that manages the software here.
> on servers but especially so on the desktop, and it does so at the cost of breaking with traditional views
Well, I would not call it "traditional", as the name is loaded. I see it more as a way to gain more control over the whole ecosystem. We see the same happen with wayland, but on a smaller scale, as wayland does not try to integrate a billion features and functionality.
> It's kind of hard to watch: I want Linux to move forward, and there's just a lot of good ideas there. But it will be painful for a large Linux community to break with traditions.
I don't like systemd, but I view this more realistic. I saw how the non-systemd distributions struggled and eventually most went extinct or were converted into systemd. Only few remain strong, and those few are often also dead - like slackware. And yes I know the spin-offs, but seriously, slackware is a dead man walking. Void is not dead, but yikes, it's not moving forward either.
It is not only systemd though. The whole linux stack got a lot bigger and more complicated. Nowadays you often need python, meson, llvm, mesa and so forth to compile things. Everything got bigger too. A lot of software was abandoned downstream, such as fluxbox - may be irrelevant to most folks, but this is one example of sooo many more. At the base of this problem sits the funding issue. Corporations have a lot more net-control over the ecosystem nowadays. Due to the funding. I think we need to solve this issue of funding, because otherwise we'll end up with systemd-like projects sitting at the key areas.
If two things hold up - 1) this is actually a 2-300B parameter model and 2) this is actually competitive with frontier OpenAI and Anthropic models (and not just benchmaxing), the implications are pretty big. It would mean you could run "frontier level" performance in one box at home.
300B models at least fit in a single maxed out Mac Studio or a small stack of DGX Sparks or AMD Strix Halo boxes.
For comparison, DeepSeek V4 Flash is all the rage now for small efficient models. It's very good for its size but far from the performance of the latest GPT Pro and Opus models. The vanilla variant has 284B parameters. It fits on both 256GB and 512GB Mac Studios and hits about 20-30 tokens/second.
The implication of all this here is that you could have a (somewhat sluggish) Opus in a small box at home. At least once competing models and hardware to run them will be available (high end Mac Studios have been discontinued).
Something tells me that this means that Google's performance numbers here are inflated.
Opus is estimated to be around 4T parameters, and 5.5 around 9T. [1] And while 3.5 at least qualifies to be in the same neighborhood, which is stunning if these numbers are all true, it may be that closing that last ~10% difference needs 50x more parameters.
Their methods are only calibrated on open models (of course) and they admit very broad confidence bounds. You can also just see from comparing their estimates of the same models at different reasoning levels that there are major confounders to this. I would err on the absolute lowest side of their estimates for frontier models (e.g. 3T for GPT-5.5, 1.5-2T for Opus 4.5+).
> the implications are pretty big. It would mean you could run "frontier level" performance in one box at home.
That wouldn't surprise me at all actually, models like Qwen3.6-35B are comparable to frontier level models from a year ago and I wouldn't be surprised if we had self-hostable open weight models matching Opus 4.7 in a year. Assuming that Google has one year of advance against Chinese lab isn't far fetched given how much resources they have compared to their Chinese competitors.
I think there was a leap around Opus 4/4.1 that hasn't quite been equalled by self hostable models yet. Perhaps full Kimi K2.6 and Deepseek V4 Pro can achieve Opus 4.1 levels (it's hard to compare anyway, benchmarks are largely a game nowadays), but both of these are also north of 1000B parameters and therefore really impractical to run at home for the foreseeable future.
It's not yet obvious to me that you can achieve the breakthrough performance of say Opus 4.1/4.5 in a number of parameters you can swing at home.
> It's not yet obvious to me that you can achieve the breakthrough performance of say Opus 4.1/4.5 in a number of parameters you can swing at home.
People used to believe the same about GPT-4, and I'm not convinced this is going to be different this time.
You do need a very big model if you want something that remembers random trivia about everything, but I'm not convinced this is needed to do meaningful work.
> 300B models at least fit in a single maxed out Mac Studio or a small stack of DGX Sparks or AMD Strix Halo boxes.
I run 2.54 BPW 397B Qwen 3.5 GGUF on a 128G mac studio at 20 tokens/second generation and 200 tokens/second processing. I'm not suggesting it matches the performance of the full BF16 model, but I did run some benchmarks locally and the results were pretty good:
- MMLU: 87.96%
- GPQA diamond: 86.36%
- IfEval: 91.13%
- GSM8k: 92.57%
So I think we have been at the "frontier capabilities at home" for a few months now.
Since I started using Qwen-3.6 35B A3B, I believe frontier like capability will be more than enough in these smaller models within a year or two, at least for coding. They don't need to memorize facts into their weights, which likely has very interesting implications that I'm not going speculatively decode
TurboQuant reduces the runtime memory needed for the model's KV cache.
This reduces both the memory bandwidth needed for inference (at the cost of slightly increasing the amount of compute needed), and the amount of VRAM used overall, meaning more VRAM can be allocated for more weights on the same hardware.
You were replying to a comment estimating model params from hardware. I am saying the param count could be higher for the same hardware.
Kimi K2.5 has also been superseded by a finer tuned Kimi K2.6 three weeks ago. Moonshot's Kimi models appear to be the favored Chinese model, at least for coding, and not Deepseek V4. z.AI's GLM 5.1 is also worth mentioning as rather competent for coding, also released in April.
Those models too will not be beating US AI labs by your metrics (although for coding, Kimi K2.6 might beat the very uneven Gemini depending on the situation), but in your critism at least consider the state of the art in your comparisons.
I have been using Deepseek v4 pro for personal projects and home infra related work for last couple of weeks. It's quality of work is not bad at all, it is fairly fast and given the fraction of the cost compared to Claude, I can keep going which makes it a very compelling option. Looking forward to trying out Kimi 2.6, thanks for the recommendation.
Even without the discount, I'll have to think about whether I need the 100 EUR tier of Anthropic Max, or whether downgrading to Pro and using DeepSeek is good enough. And they're also up on OpenRouter and other places.
Been using those models, not quite comparable with Opus 4.6/4.7 but with max reasoning, pretty good for a variety of dev tasks! Only big problem is no ability to process images, so can't really do browser use for some semi-automated testing, I'd have to write Playwright tests even when I don't want to.
I've been using OpenCode Go ($10/month) for personal projects (I have Claude subscription for $DAYJOB) and for the tinkering around that I do for myself the quality of the open weight models and the limits of the OpenCode plan are sufficient. I agree that for a lot of dev tasks they're quite good!
I've been using Deepseek 4 Pro (instead of Sonnet 4.6) as the developer LLM (Opus is the planner) and it's been great. Not super fast, with all the reasoning, but has been writing good code, and I think I paid $5 so far (whereas with Sonnet I'd have run out of the weekly limits on Max for weeks now).
Definitely recommended, though it's crucial that you have GPT 5.5 review the code afterwards.
Apple is presumably leaving a lot of money on the table by not trying to sell Apple Silicon for AI inference and training. They're the only ones who can attach reasonably large GPUs (M3 Ultra) to very large amounts of cheaper memory (512GB SO-DIMM per GPU). Apple could e.g. sell server SKUs of Mac Studios, heck they can sell M3 Ultra chips on PCIe cards. And they could further develop Apple Silicon in that direction. Presumably they would be seen as a very legit competitor to Nvidia that way, perhaps moreso than Intel and AMD. I'd assume that in the current climate this would be extremely lucrative.
Now, actually doing this would disrupt Apple's own supply chain as well as force it to spend significant internal resources and cultural change for this kind of product line. There's a good argument to be made it would disproportionally negatively affect its Mac business, so this would be a very risky move.
But given that AI hardware is likely much higher margin than the Mac business an argument could probably (sadly) be made that it'd be lucrative for them to try it. I personally don't think Apple is inclined to take this kind of risk to jeopardize the Mac, but I'm sure some people at Apple have considered this.
I have a workflow to create PRs where at the end, I spin up a subagent that goes through the PR to 1) remove the bulk of the comments and 2) attempt to normalize it for human consumption (I ask it to optimize wording for junior developers and non-native speakers).
I use Haiku for it, hoping that a dumber model will do a better job. I have noticed in brief experimentation that even Sonnet will sympathize too much with the need of excessive comments and is not a good adverserial reviewer for slop comments.
After that workflow runs I still need to go in and manually correct, but it's a lot less than vanilla Opus. It really feels like the problem of slop comments has gotten worse as models has gotten smarter.
reply