Data points from my own usage, for whatever they're worth[1]. I'm just an individual contributor and hardly represent enterprise or large orgs:

I've been on Kimi, with a little DeepSeek-V4-Pro, GLM 5.2/5.3, and MiMo thrown in, for probably about a year now. It's great here!

1) For DeepSeek, I recommend their Reasonix harness strongly, due to its alignment to DeepSeek's prefix cache. It means mostly (95%+) cache hit input tokens, so very cheap large-scale code analyses and things that require mega context windows (at the cost of some attentional drift, yes). Reasonix does require that you send data to China. This is fine. I mostly use this for big, expansive ingestion of open-source codebases to figure out how something really works, usually something that documentation doesn't quite reach.

The economy of doing it this way versus American frontier model companies' token pricing cannot be overstated. I think I topped up $10 in June (2.5 months ago) and have still not burned through it, despite cycling untold tens of millions of tokens through it.

My biggest annoyance is that DeepSeek does seem to be considerably rate-limited of late, at least during working hours in Beijing, which is a range that I gather to be quite expansive there. I'm not blasting it with anything, I'm just noting that the agent takes 10-20 minutes to do stuff that takes much less time if I'm willing to pay the OpenRouter premium.

2) For most everyday stuff outside of where Reasonix + DeepSeek just makes overwhelming sense, I use OpenCode/Maki/Pi/whatever harness I feel like using today with Kimi K3, via OpenRouter. This does not require sending data to China.

I also use Kimi K3 in Zed via OpenRouter quite a bit, but sometimes like to mix it up with the other models.

3) Because I have the most experience with it, I can say with confidence that I would generally consider the SWE capabilities of Kimi to be on par with Claude, at least for the bottom 99% of purposes--and certainly, any routine business programming.

I think this has been true for a long time, well before K3. I've been using Kimi since K2.5.

4) For local hardware experiments on my MacBook Pro (M4 Max, 128 GB unified memory), Qwen3.6-35B-A3B (speed) and Qwen3.8-27B (intelligence, but slow). As has been widely noted, this amount of unified memory isn't as useful as it seems, due to memory bandwidth and decoding constraints, lack of tensor cores (on the M4 Max, anyway), etc.

A giant bag of memory isn't fast, but it'll let you load some impressively big models.

The future M5 Studio Macs will continue in this general vein, but will of course be somewhat faster, particularly due to the apparition of tensor cores in the M5 -- excuse me, "Neural Accelerators".

Still, if you really want to cook, get a real GPU. Real GPU running quantisations is still a lot better than a big slow bag of unified memory.

5) Overall, the Chinese models are simply excellent, and cater to lots of use-cases and tastes. However, I'll still tend to use Claude ($20/mo subscription) for general Q&A, whether of a technical nature or otherwise, particularly where web research and worldly knowledge is required.

[1] Disclaimer: this comment is an elaboration of https://news.ycombinator.com/item?id=49809605, which is not something I'd normally do. However, it seems a lot more relevant here than where I had originally posted it, in an article about gauging MiMo Pro v2.6 capabilities.

▲

arel-fgkl 2 hours ago | parent [-]

As I understand it, all real world use cases requires provider data centers and the local efforts are experimental.

So "open models" buy you little and the independence from the oligopoly is not achieved.

	▲	abalashov 2 hours ago \| parent [-]
		Yeah, that's broadly accurate, although details matter and local models are surprisingly capable. Local models are viable for a lot of small use-cases, but no, there's no deluding oneself that frontier quality doesn't require frontier size. However, everyone who thinks they need frontier "intelligence" for internal CRUD type tasks should look in the mirror and ask themselves if that's really true. Still, even if you need other people's industrial-class hardware, open models offer a lot more freedom and options. They allow you to use GPU capacity from entities who are not themselves building or training models and are ostensibly disinterested. There's a big range of possibilities here in terms of data sovereignty and so forth. 1) You can use OpenRouter to route your open model requests to US-based inference providers with ZDR (zero data retention), as far as you can believe anything in this world. If you look at who actually serves open Chinese models on OpenRouter, you'll see a lot of folks like Digital Ocean, etc. I suppose I can't vouch for their purity, no, but I'd much rather send data there than send it to Dario. 2) Or, you can rent GPUs from companies like Runpod or Vast.ai and serve some very sizable models to yourself (e.g. using their pre-built vLLM images). You can't serve a model like Kimi K3 to yourself that way, at least not in any economically reasonable way. However, a private H100SXM or B200 can go a long way. You could serve the big Qwens, or DeepSeek-V4-Flash--you could do a lot if you're willing to spend on a rented GPU with sizable VRAM. 3) Finally, if you have and want to spend $750K-$1MM+ (I suspect I'm low-balling at this point), and if you can get them in the current demand climate, you can absolutely buy 16 x H200SXMs, with the appropriate boards to take them, pay for 20-25 kW of cooling, etc., and run one of these models yourself, on your kitchen floor if you like. You simply cannot do that with Anthropic or OpenAI.