The terminal bench scores look weak but nice otherwise. I hope once the benchmarks are saturated, companies can focus on shrinking the models. Until then, let the games continue.
The frontend examples, especially the first one, look uncannily similar to what Gemini 3 Pro usually produces. Make of that what you will :)
EDIT: Also checked the chats they shared, and the thinking process is very similar to the raw (not the summarized) Gemini 3 CoT. All the bold sections, numbered lists. It's a very unique CoT style that only Gemini 3 had before today :)
Yeah, I think it sometimes even repeats Gemini's injected platform instructions. It's pretty curious because a) Gemini uses something closer to the "chain of draft" and never repeats them in full naturally, only the relevant part, and b) these instructions don't seem to have any effect in GLM, it repeats them in the CoT but never follows them. Which is a real problem with any CoT trained through RL (the meaning diverges from the natural language due to reward hacking). Is it possible they used is in the initial SFT pass to improve the CoT readability?
My quickie: MoE model heavily optimized for coding agents, complex reasoning, and tool use. 358B/32B active. vLLM/SGLang only supported on the main branch of these engines, not the stable releases. Supports tool calling in OpenAI-style format. Multilingual English/Chinese primary. Context window: 200k. Claims Claude 3.5 Sonnet/GPT-5 level performance. 716GB in FP16, probably ca 220GB for Q4_K_M.
My most important takeaway is that, in theory, I could get a "relatively" cheap Mac Studio and run this locally, and get usable coding assistance without being dependent on any of the large LLM providers. Maybe utilizing Kimik2 in addition. I like that open-weight models are nipping at the feet of the proprietary models.
This is true assuming there will be updates consistently. One of the advantages of the proprietary models is that the are updated often EKG and the cutoff date moves into the future
This is important because libraries change, introduce new functionality, deprecate methods and rename things all the time, e.g. Polars.
I tested the previous one GLM-4.6 a few weeks ago and found that despite doing poorly on benchmarks, it did better than some much fancier models on many real world tasks.
Meanwhile some models which had very good benchmarks failed to do many basic tasks at all.
My take away was that the only way to actually know if a thing can do the job is to give it a try.
This model is much stronger than 3.5 sonnet, 3.5 sonnet scored 49% on swe-bench verified vs. 72% here. This model is about 4 points ahead of sonnet4, but behind sonnet 4.5 by 4 points.
If I were to guess, we will see a convergence on measurable/perceptible coding ability sometime early next year without substantially updated benchmarks.
Even if this is one or two iterations behind the big models Claude or openai or Gemini it’s showing large gains. Here’s hoping this gets even better and better and I can run this locally and also that it doesn’t melt my PC.
I've been playing around with this in z-ai and I'm very impressed. For my math/research heavy applications it is up there with GPT-5.2 thinking and Gemini 3 Pro. And its well ahead of K2 thinking and Opus 4.5.
> For my math/research heavy applications it is up there with GPT-5.2 thinking and Gemini 3 Pro. And it’s well ahead of K2 thinking and Opus 4.5.
I wouldn’t use the z-ai subscription for anything work related/serious if I were you. From what I understand, they can train on prompts + output from paying subscribers and I have yet to find an opt-out. Third party hosting providers like synthetic.new are a better bet IMO.
Grok 4 Heavy wasn't considered in comparisons.
Grok meets or exceeds the same benchmarks that Gemini 3 excels at, saturating mmlu, scoring highest on many of the coding specific benchmarks. Overall better than Claude 4.5, in my experience, not just with the benchmarks.
Benchmarks aren't everything, but if you're going to contrast performance against a selection of top models, then pick the top models? I've seen a handful of companies do this, including big labs, where they conveniently leave out significant competitors, and it comes across as insecure and petty.
Claude has better tooling and UX. xAI isn't nearly as focused on the app and the ecosystem of tools around it and so on, so a lot of things end up more or less an afterthought, with nearly all the focus going toward the AI development.
$300/month is a lot, and it's not as fast as other models, so it should be easy to sell GLM as almost as good as the very expensive, slow, Grok Heavy, or so on.
GLM has 128k, grok 4 heavy 256k, etc.
Nitpicking aside, the fact that they've got an open model that is just a smidge less capable than the multibillion dollar state of the art models is fantastic. Should hopefully see GLM 4.7 showing up on the private hosting platforms before long. We're still a year or two from consumer gear starting to get enough memory and power to handle the big models. Prosumer mac rigs can get up there, quantized, but quantized performance is rickety at best, and at that point you look at the costs of self hosting vs private hosts vs $200/$300 a month (+ continual upgrades)
Frontier labs only have a few years left where they can continue to charge a pile for the flagship heavyweight models, I don't think most people will be willing to pay $300 for a 5 or 10% boost over what they can run locally.
It seems like someone at X.ai likes maxing benchmarks but real world usage shows it significantly behind frontier models.
I do appreciate their desire to be the most popular coding model on OpenRouter and offer Grok4-Fast for free. That's a notable step down from frontier models but fine for lots of bug fixing. I've put hundreds of millions of tokens through it.
"
Grok 4 Heavy wasn't considered in comparisons. Grok meets or exceeds the same benchmarks that Gemini 3 excels at, saturating mmlu, scoring highest on many of the coding specific benchmarks. Overall better than Claude 4.5, in my experience, not just with the benchmarks."
I think these types of comments should just be forbidden from Hacker News.
It's all feelycraft and impossible to distinguish from motivated speech.
I've been using Z.Ai coding plan for last few months, generally very pleasant experience. I think with GLM-4.6 they had some issues which this corrects.
Overall solid offering, they have MCP you plug into ClaudeCode or OpenCode and it just works.
I'm surprised by this; I have it also and was running through OpenCode but I gave up and moved back to Claude Code. I was not able to get it to generate any useful code for me.
How did you manage to use it? I am wondering if maybe I was using it incorrectly, or needed to include different context to get something useful out of it.
A few comments mentioning distillation. If you use claude-code with the z.ai coding plan, I think it quickly becomes obvious they did train on other models. Even the "you're absolutely right" was there. But that's ok. The price/performance ratio is unmatched.
I don't think that's particularly conclusive for training on other models. Seems plausible to me that the internet data corpus simply converges on this hence multiple models doing this.
I imagine - and sure hope so - everyone trains on everything else. Distillation - ofc if one has bigger/other models providing true posterior token probabilities in the (0,1) interval (a number between 0 and 1), rather than 1-hot-N targets that are '0 for 200K-sans-this-token, and 1 for the desired output token' - one should use the former instead of the latter. It's amazing how as a simple as straightforward idea should face so much resistance (paper rejected) and from the supposedly most open minded and devoted to knowing (academia) and on the wrong grounds ('will have no impact on industry'; in fact - it's had tremendous impact on industry; better rejection wd have been 'duh it is obvious'). We are not trying to torture the model and the gpu cluster to be learning from 0 - when knowledge is already available. :-)
I have been using 4.6 on Cerebras (or Groq with other models) since it dropped and it is a glimpse of the future. If AGI never happens but we manage to optimise things so I can run that on my handheld/tablet/laptop device, I am beyond happy. And I guess that might happen. Maybe with custom inference hardware like Cerebras. But seeing this generate at that speed is just jaw dropping.
Cerebras and Groq both have their own novel chip designs. If they can scale and create a consumer friendly product that would be a great, but I believe their speeds are due to them having all of their chips networked together, in addition to design for LLM usage. AGI will likely happen at the data center level before we can get on-device performance equivalent to what we have access to today (affordably), but I would love to be wrong about that.
Apple's M5 Max will probably be able to run it decently (as it will fix the biggest issue with the current lineup, prompt processing, in addition to a bandwidth bump).
That should easily run an 8 bit (~360GB) quant of the model. It's probably going to be the first actually portable machine that can run it. Strix Halo does not come with enough memory (or bandwidth) to run it (would need almost 180GB for weights + context even at 4 bits), and they don't have any laptops available with the top end (max 395+) chips, only mini PCs and a tablet.
Right now you only get the performance you want out of a multi GPU setup.
I tried this on OpenRouter chat interface to write a few documents. Quick thoughts: Its writing has less vibe of AI due to the lack of em-dashes! I primarily use Kimi2 Thinking for personal usage. Kimi writing is also very good, on par with the frontier models like Sonnet or Gemini. But, just like them, Kimi2 also feels AI. I can't quantify or explain why, though.
For work, it is Claude Code and Anthropic exclusively.
Cerebras is serving GLM4.6 at 1000 tokens/s right now. They're probably likely to upgrade to this model.
I really wonder if GLM 4.7 or models a few generations from now will be able to function effectively in simulated software dev org environments, especially that they self-correct their errors well enough that they build up useful code over time in such a simulated org as opposed to increasing piles of technical debt. Possibly they are managed by "bosses" which are agents running on the latest frontier models like Opus 4.5 or Gemini 3. I'm thinking in the direction of this article: https://www.anthropic.com/engineering/effective-harnesses-fo...
If the open source models get good enough, then the ability to run them at 1k tokens per second on Cerebras would be a massive benefit compared to any other models in being able to run such an overall SWE org quickly.
I am quite impressed with this model. Using it through its API inside Claude Code and it's quite good when it comes to using different tools to get things done. No more weekly limit drama of Claude also their quarterly plan is available for just $8
> Preserved Thinking: In coding agent scenarios, GLM-4.7 automatically retains all thinking blocks across multi-turn conversations, reusing the existing reasoning instead of re-deriving from scratch. This reduces information loss and inconsistencies, and is well-suited for long-horizon, complex tasks.
does it NOT already do this? i dont see the difference. the image doesnt show any before/after so i dont see any difference
Great performance for coding after I snatched a pretty good deal 50%+20%+10%(with bonus link) off.
60x Claude Code Pro Performance for Max Plan for the almost the same price. Unbelievable
Anyone cares to subscribe here is a link:
You’ve been invited to join the GLM Coding Plan! Enjoy full support for Claude Code, Cline, and 10+ top coding tools — starting at just $3/month. Subscribe now and grab the limited-time deal! Link：
When I click on Subscribe on any of the plan, nothing happens. I see this error on Dev Tools.
page-3f0b51d55efc183b.js:1 Uncaught TypeError: Cannot read properties of undefined (reading 'toString')
at page-3f0b51d55efc183b.js:1:16525
at Object.onClick (page-3f0b51d55efc183b.js:1:17354)
at 4677-95d3b905dc8dee28.js:1:24494
at i8 (aa09bbc3-6ec66205233465ec.js:1:135367)
at aa09bbc3-6ec66205233465ec.js:1:141453
at nz (aa09bbc3-6ec66205233465ec.js:1:19201)
at sn (aa09bbc3-6ec66205233465ec.js:1:136600)
at cc (aa09bbc3-6ec66205233465ec.js:1:163602)
at ci (aa09bbc3-6ec66205233465ec.js:1:163424)
A bit weird for an AI coding model company not to have seamless buying experience
GLM 4.6 has been very popular from my perspective as an inference provider with a surprising number of people using it as a daily driver for coding. Excited to see the improvements 4.7 delivers, this model has great PMF so to speak.
Appears to be cheap and effective, though under suspicion.
But the personal and policy issues are about as daunting as the technology is promising.
Some the terms, possibly similar to many such services:
- The use of Z.ai to develop, train, or enhance any algorithms, models, or technologies that directly or indirectly compete with us is prohibited
- Any other usage that may harm the interests of us is strictly forbidden
- You must not publicly disclose [...] defects through the internet or other channels.
- [You] may not remove, modify, or obscure any deep synthesis service identifiers added to Outputs by Z.ai, regardless of the form in which such identifiers are presented
- For individual users, we reserve the right to process any User Content to improve our existing Services and/or to develop new products and services, including for our internal business operations and for the benefit of other customers.
- You hereby explicitly authorize and consent to our: [...] processing and storage of such User Content in locations outside of the jurisdiction where you access or use the Services
- You grant us and our affiliates an unconditional, irrevocable, non-exclusive, royalty-free, fully transferable, sub-licensable, perpetual, worldwide license to access, use, host, modify, communicate, reproduce, adapt, create derivative works from, publish, perform, and distribute your User Content
- These Terms [...] shall be governed by the laws of Singapore
To state the obvious competition issues: If/since Anthropic, OpenAI, Google, X.AI, et al are spending billions on data centers, research, and services, they'll need to make some revenue. Z.ai could dump services out of a strategic interest in destroying competition. This dumping is good for the consumer short-term, but if it destroys competition, bad in the long term. Still, customers need to compete with each other, and thus would be at a disadvantage if they don't take advantage of the dumping.
Once your job or company depends on it to succeed, there really isn't a question.
54 comments
[ 4.2 ms ] story [ 245 ms ] threadhttps://github.com/charmbracelet/crush
https://news.ycombinator.com/item?id=44736176
EDIT: Also checked the chats they shared, and the thinking process is very similar to the raw (not the summarized) Gemini 3 CoT. All the bold sections, numbered lists. It's a very unique CoT style that only Gemini 3 had before today :)
My most important takeaway is that, in theory, I could get a "relatively" cheap Mac Studio and run this locally, and get usable coding assistance without being dependent on any of the large LLM providers. Maybe utilizing Kimik2 in addition. I like that open-weight models are nipping at the feet of the proprietary models.
This is important because libraries change, introduce new functionality, deprecate methods and rename things all the time, e.g. Polars.
I tested the previous one GLM-4.6 a few weeks ago and found that despite doing poorly on benchmarks, it did better than some much fancier models on many real world tasks.
Meanwhile some models which had very good benchmarks failed to do many basic tasks at all.
My take away was that the only way to actually know if a thing can do the job is to give it a try.
If I were to guess, we will see a convergence on measurable/perceptible coding ability sometime early next year without substantially updated benchmarks.
I wouldn’t use the z-ai subscription for anything work related/serious if I were you. From what I understand, they can train on prompts + output from paying subscribers and I have yet to find an opt-out. Third party hosting providers like synthetic.new are a better bet IMO.
Benchmarks aren't everything, but if you're going to contrast performance against a selection of top models, then pick the top models? I've seen a handful of companies do this, including big labs, where they conveniently leave out significant competitors, and it comes across as insecure and petty.
Claude has better tooling and UX. xAI isn't nearly as focused on the app and the ecosystem of tools around it and so on, so a lot of things end up more or less an afterthought, with nearly all the focus going toward the AI development.
$300/month is a lot, and it's not as fast as other models, so it should be easy to sell GLM as almost as good as the very expensive, slow, Grok Heavy, or so on.
GLM has 128k, grok 4 heavy 256k, etc.
Nitpicking aside, the fact that they've got an open model that is just a smidge less capable than the multibillion dollar state of the art models is fantastic. Should hopefully see GLM 4.7 showing up on the private hosting platforms before long. We're still a year or two from consumer gear starting to get enough memory and power to handle the big models. Prosumer mac rigs can get up there, quantized, but quantized performance is rickety at best, and at that point you look at the costs of self hosting vs private hosts vs $200/$300 a month (+ continual upgrades)
Frontier labs only have a few years left where they can continue to charge a pile for the flagship heavyweight models, I don't think most people will be willing to pay $300 for a 5 or 10% boost over what they can run locally.
I do appreciate their desire to be the most popular coding model on OpenRouter and offer Grok4-Fast for free. That's a notable step down from frontier models but fine for lots of bug fixing. I've put hundreds of millions of tokens through it.
I think these types of comments should just be forbidden from Hacker News.
It's all feelycraft and impossible to distinguish from motivated speech.
Overall solid offering, they have MCP you plug into ClaudeCode or OpenCode and it just works.
How did you manage to use it? I am wondering if maybe I was using it incorrectly, or needed to include different context to get something useful out of it.
(I know that people must pay it on privacy) but still for maybe playing around with still worth it imo
I don't think that's particularly conclusive for training on other models. Seems plausible to me that the internet data corpus simply converges on this hence multiple models doing this.
...or not...hard to tell either way.
That should easily run an 8 bit (~360GB) quant of the model. It's probably going to be the first actually portable machine that can run it. Strix Halo does not come with enough memory (or bandwidth) to run it (would need almost 180GB for weights + context even at 4 bits), and they don't have any laptops available with the top end (max 395+) chips, only mini PCs and a tablet.
Right now you only get the performance you want out of a multi GPU setup.
For work, it is Claude Code and Anthropic exclusively.
I really wonder if GLM 4.7 or models a few generations from now will be able to function effectively in simulated software dev org environments, especially that they self-correct their errors well enough that they build up useful code over time in such a simulated org as opposed to increasing piles of technical debt. Possibly they are managed by "bosses" which are agents running on the latest frontier models like Opus 4.5 or Gemini 3. I'm thinking in the direction of this article: https://www.anthropic.com/engineering/effective-harnesses-fo...
If the open source models get good enough, then the ability to run them at 1k tokens per second on Cerebras would be a massive benefit compared to any other models in being able to run such an overall SWE org quickly.
does it NOT already do this? i dont see the difference. the image doesnt show any before/after so i dont see any difference
Great performance for coding after I snatched a pretty good deal 50%+20%+10%(with bonus link) off.
60x Claude Code Pro Performance for Max Plan for the almost the same price. Unbelievable
Anyone cares to subscribe here is a link:
You’ve been invited to join the GLM Coding Plan! Enjoy full support for Claude Code, Cline, and 10+ top coding tools — starting at just $3/month. Subscribe now and grab the limited-time deal! Link：
https://z.ai/subscribe?ic=OUCO7ISEDB
Great performance for coding after I snatched a pretty good deal 50%+20%+10%(with bonus link) off.
60x Claude Code Pro Performance for Max Plan for the almost the same price. Unbelievable
Anyone cares to subscribe here is a link:
https://z.ai/subscribe?ic=OUCO7ISEDB
page-3f0b51d55efc183b.js:1 Uncaught TypeError: Cannot read properties of undefined (reading 'toString') at page-3f0b51d55efc183b.js:1:16525 at Object.onClick (page-3f0b51d55efc183b.js:1:17354) at 4677-95d3b905dc8dee28.js:1:24494 at i8 (aa09bbc3-6ec66205233465ec.js:1:135367) at aa09bbc3-6ec66205233465ec.js:1:141453 at nz (aa09bbc3-6ec66205233465ec.js:1:19201) at sn (aa09bbc3-6ec66205233465ec.js:1:136600) at cc (aa09bbc3-6ec66205233465ec.js:1:163602) at ci (aa09bbc3-6ec66205233465ec.js:1:163424)
A bit weird for an AI coding model company not to have seamless buying experience
But the personal and policy issues are about as daunting as the technology is promising.
Some the terms, possibly similar to many such services:
To state the obvious competition issues: If/since Anthropic, OpenAI, Google, X.AI, et al are spending billions on data centers, research, and services, they'll need to make some revenue. Z.ai could dump services out of a strategic interest in destroying competition. This dumping is good for the consumer short-term, but if it destroys competition, bad in the long term. Still, customers need to compete with each other, and thus would be at a disadvantage if they don't take advantage of the dumping.Once your job or company depends on it to succeed, there really isn't a question.