GLM-5.1: Towards Long-Horizon Tasks
https://z.ai/blog/glm-5.1Contrary to the model card, its one-shot performance is more impressive than its agentic abilities. On both metrics, GLM 5.1 is competitive with frontier models.
But keeping in mind this is an open source model operating near the frontier, it's nothing short of incredible.
I suspect 2 issues with the model are keeping it from fully realizing its potential in agentic harnesses: - Context rot (already a common complaint). We are still working on a metric to robustly test and visualize this on the site. - The model was most likely overtrained on standardized toolsets and benchmarks, and isn't as adaptive in using arbitrary tooling in our custom harness simulations. We've decided to commit to measuring intelligence as the ability to use custom, changing tools, instead of being trained to use specific tools (while still always providing a way to run local bash and other common tools). There are arguments to be made for either, but the former is more indicative of general intelligence. Regardless, it's a subtle difference and GLM 5.1 still performs well with tooling in our environments.
Crazy week for open source AI. Gemma 4 has shown that large model density is nowhere near optimized. Moats are shrinking.
If there are more representations of model performance you'd like to see, I'm actively reading your feedback and ideas.
I've been testing it for awhile now since it seemed to have potential as a local model.
With this new update it still cannot parse simple, test PDFs correctly. It inconsistently tells me that the value in the name field in the document is incorrect, and has the name reversed to put the last name first. Or that a date is wrong as it's in the past/future, when it is not. Tons of fundamental errors like that.
Even when looking at the thinking process there are issues:
I used a test website for it to analyze and it says that the sites copyright year states 2026 which is in the future and to investigate as it could be an attack, but right after prints today's correct date.
I'm in the process of trying to get it uncensored. Hopefully that will create some use out of z.ai
Edit: by the way, which is the best uncensored model at the moment?
Meanwhile we're even seeing emerging 'engram' and 'inner-layer embedding parameters' techniques where the possibility of SSD offload is planned for in advance when developing the architecture.
(1) OpenAI & Anthropic are absolutely cooked; it's obvious they have no moat
(2) Local/private inference is the future of AI
(3) There's *still* no killer product yet (so get to work!)1) OpenAI and Anthropic are killing it, and continue to do so, their coding tools are unmatched for professionals.
2) Local models don't hold a candle to SOTA models and there's nothing on the horizon that indicates that consumers will be able to run anything close to what you can get in a data center.
3) Coding is a killer product, OpenAI and Anthropic are raking in the cash. The top 3 apps are apps in the app store are AI. Everyone who knows anything is using AI, every day, across the economy.
On (2), I agree with you for local models. BUT, there are also the open source Chinese models accessible via open-router. Your argument ("don't hold a candle to SOTA models") does not hold if the comparison is between those.
On (1), I agree more with the grandparent than with your assessment. Yes, OpenAI and Anthropic are killing it for now, but the time horizon is very short. I use codex and claude daily, but it's also clear to me that open source is catching up quickly, both w.r.t. the models and the agentic harnesses.
(1): You don't have to be an Ed Zitron disciple to infer that OpenAI and Anthropic are likely overvalued and that Nvidia is selling everyone shovels in a gold rush. AI is a game-changing technology, but a shitty chat interface does not a company make. OpenAI and Anthropic need to recoup astronomical costs used in training these models. Models that are now being distilled[1] and are quickly becoming commoditized. (And frankly, models that were trained by torrenting copyrighted data[2], anyway.) Many have been calling this out for years: the model cannot be your product. And to be clear, OpenAI/Anthropic most definitely know this: that's why they've been aquihiring like crazy, trying to find that one team that will make the thing.
(2): Token prices are significantly subsidized and anyone that does any serious work with AI can tell you this. Go use an almost-SOTA model (a big Deepseek or Qwen model) offered by many bare-metal providers and you'll see what "true" token prices should look like. The end-state here is likely some models running locally and some running in the cloud. But the current state of OpenClaw token-vomit on top of Claude is fiscally untenable (in fact, this is why Anthropic shut it down).
(3): This is typical Dropbox HN snark[3], of which I am also often guilty of. I really don't think AI coding is a killer product and this seems very myopic—engineers are an extreme minority. Imo, the closest we've seen to something revolutionary is OpenClaw, but it's janky, hard to set up, full of vulnerabilities, and you need to buy a separate computer. But there's certainly a spark there. (And that's personally the vertical I'm focusing on.)
[1] https://www.anthropic.com/news/detecting-and-preventing-dist...
[2] https://media.npr.org/assets/artslife/arts/2025/complaint.pd...
Landing a man on the moon is way more impressive. Finding several vaccines for a once in a century pandemic within a year of its outbreak is and achievement that in its impact and importance dwarfs what the entire LLM industry put together has achieved. The near-complete eradication of polio, once again, way more important and impactful.
We probably talk abuot a year of progress diffeerence.
Its also still quite expensive for an avg person to consume any of it. Either due to hardware invest, energy cost or API cost.
Also professionally I don't think anyone will really spend a little bit less money of having the 3th quality model running if they can run the best model.
I'm happy that we reach levels were this becomes an alternative if you value open and control though.
I'd like to think the superior product wins. But Windows still thrives despite widespread Linux availability. I think sometimes we can underestimate the resilience of the tech oligopolies, particularly when they're VC-funded.
If I want to switch from Windows to Linux, I have to reconsider a whole variety of applications, learn a different UX, migrate data, all sorts of annoyances.
When I switch between Codex and Claude Code, there is literally no difference in how I interact with them. They and a number of other competitors are drop in replacements for each other.
That's because by most metrics Linux is inferior is Windows.
That's a valuable guarantee. So valuable, in fact, that you won't get it from Anthropic, OpenAI, or Google at any price.
(2) is probably true but with caveats. Top-tier models will never run on desktop machines, but companies should (and do) host their own models. The future is open-weight though, that much is for sure.
(3) This is so ignorant that others have already responded to it. Look outside of your own bubble, please.
Sorry, but you don't know that
Every time I asked a question it generated an interactive geometry graph on the fly in Javascript. Sometimes it spent minutes compiling and testing code on the server so it could make sure it was correct. I was really impressed.
Anyway I couldn't really learn anything since when the code didn't work I wasn't sure if I had ported it wrong or the AI did it wrong, so I ended up learning how to calculate SDF and pixel to hex grid from tutorials I found on google instead.
I think big corporations will continue to use them no matter how cheap and good other models are. There's a saying: nobody was fired for buying IBM.
For me, Opus 4.6 isn't working quite right currently, and I often use GLM 5.1 instead. I'd prefer to use peak Opus over GLM 5.1, but GLM 5.1 is an adequate fallback. It's incredible how good open-weight models have gotten.
i have a feeling its nearing opus 4.5 level if they could fix it getting crazy after like 100k tokens.
Mid-sized models like gpt-oss minimax and qwen3.5 122b are around 6%, and gemma4 31b around 7% (but much slower).
I haven’t tried Opus or ChatGPT due to high costs on openrouter for this application.
My use cases are not code editing or authoring related, but when it comes to understanding a codebase and it's docs to help stakeholders write tasks or understand systems it has always outperformed american models at roughly half the price.
Overeager, but I was really really impressed.
It's a fun way to quantify the real-world performance between models that's more practical and actionable.
I think the model is now tuned more towards agentic use/coding than general intelligence.
[0]: https://aibenchy.com/compare/z-ai-glm-5-medium/z-ai-glm-5-1-...
During off peak hour a simple 3 line CSS change took over 50 minutes and it routinely times out mid-tool and leaves dangling XML and tool uses everywhere, overwriting files badly or patching duplicate lines into files
But it's all casual side projects.
Edit: I often to /compact at around 100 000 token or switch to a new session. Maybe that is why.
For the price this is a pretty damn impressive model.
Providers like DeepInfra are already giving access to 5.1 https://deepinfra.com/zai-org/GLM-5.1
$1.40 in $4.40 out $0.26 cached
/ 1M tokens
That's more expensive than other models, but not terrible, and will go down over time, and is far far cheaper than Opus or Sonnet or GPT.
I haven't had any bad luck with DeepInfra in particular with quantization or rate limiting. But I've only heard bad things about people who used z.ai directly.
Devil's advocate: why shouldn't they do it if OpenAI, Anthropic and Google get away with playing this game?
"I am the storm that is approaching, provoking..." : )
> "build a Linux-style desktop environment as a web application"
We all know that building a spec-compliant browser alone is a herculean task.
Would it succeed? Probably not, but it would be way more interesting, even if it didn't work.
I find things like Claude's C compiler way more interesting where, even though CCC is objectively bad (code is messy, generates very bad unoptimized code, etc) it at least is something cool and shows that with some human guideance it could generate something even better.
Excited to test this.
Everyone else isn't that far behind and they aren't all gonna just wall off their new model.
A reason that Anthropic will eventually give is 'the competition can do what Glasswing can do so what's the point limiting it'.
For short-term bugfixing and tweaks though, it does about what I'd expect from Sonnet for a pretty low price.
https://github.com/Opencode-DCP/opencode-dynamic-context-pru...
Since the entire purpose, focus and motivation of this model seems to have been "coherency over longer contexts", doesn't that issue makes it not an OK model? It's bad at the thing it's supposed to be good at, no?
So I need them to not only not devolve into gibberish, but remain smart enough to be useful at contexts several times longer than that.
I suspect that this isn't the model, but something that z.ai is doing with hosting it. At launch I was related to find glm-5.1 was stable even as the context window filled all the way up (~200k). Where-as glm-5, while it could still talk and think, but had forgotten the finer points of tool use to the point where it was making grevious errors as it went (burning gobs of tokens to fix duplicate code problems).
However, real brutal changes happened sometimes in the last two or three months: the parent problem emerged and emerged hard, out of nowhere. Worse, for me, it seemed to be around 60k context windows, which was heinous: I was honestly a bit despondent that my z.ai subscription had become so effectively useless. That I could only work on small problems.
Thankfully the coherency barrier raised signficiantly around three weeks go. It now seems to lose its mind and emits chaotic non-sentance gibberish around 100k for me. GLM-5 was already getting pretty shaky at this point, so I feel like I at least have some kind of parity. But at least glm-5 was speaking & thinking with real sentances, I could keep conversing with it somewhat, where-as glm-5.1 seems to go from perfectly level headed working fine to all of a sudden just total breakdown, hard switch, at such a predictable context window size.
It seems so so probable to me that this isn't the model that's making this happen: it's the hosting. There's some KV cache issue, or they are trying to expand the context window in some way, or to switch from one serving pool of small context to a big context serving pool, or something infrastructure wise that falls flat and collapses. Seeing the window so clearly change from 200k to 60k to 100k is both hope, but also, misery.
I've been leaving some breadcrumbs on Bluesky as I go. It's been brutal to see. Especially having tasted a working glm-5.1. I don't super want to pay API rates to someone else, but I fully expect this situation to not reproduce on other hosting, and may well spend the money to try and see. https://bsky.app/profile/jauntywk.bsky.social/post/3mhxep7ek...
All such a shame because aside from totally going mad & speaking unpuncutaed gibberish, glm-5.1 is clearly very very good and I trust it enormously.
[[you guys, please don't post like this to HN - it will just irritate the community and get you flamed]]
Interesting.
Hopefully these aren't bots created by Z.AI because GLM doesn't need fake engagement.
Being "better than Opus 4.6" is not really something a benchmark will tell you. It's much more a consensus of users liking the flavor of an answer, rather than fueling x% correct on a benchmark.