11 ms·
GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models [pdf]
- ttul 1y agoThis feels like the first open model that doesn’t require significant caveats when comparing to frontier proprietary models. The parameter efficiency alone suggests some genuine innovations in training methodology. I am keen to see some independent verification of the results and to see how if does on Aider’s LLM Leaderboard.
- lumost 1y agoWhy was qwen3 omitted from the coding benchmark but not other benchmarks?
- Reubend 1y agoFantastic release, and it's under the Apache license too. I'm so happy that we've got open source models pushing the envelope.
- darknoon 1y agoIt's ok, somewhere between a qwen 2.5 VL and the frontier models (o3 / opus 4) on visual reasoning
- reissbaker 1y agoI've been playing around with GLM-4.5 as a coding model for a while now and it's really, really good. In the coding agent I've been working on, Octofriend [1], I've sometimes had it on and confused it for Claude 4. Subjectively, my experience has been: 1. Claude is somewhat better at whole-codebase tasks, where you need to reason over a bunch of context and consider system interactions. 2. GLM-4.5 is somewhat better at being "honest" — i.e. I rarely see it doing the things Claude does like making broken tests pass by changing the test instead of fixing the bug. Both are quite good though, and GLM-4.5 has found bugs that both Claude 4 Sonnet and 4.1 Opus have failed to catch. In general I think Claude wins a little more frequently on debugging tasks than GLM-4.5, but it's close. Compared to GPT-5, both Claude and GLM feel like they're more consistent, although GPT-5 sometimes has long brilliant runs where it nails everything with subjectively higher code quality than either of the latter. However, once GPT-5 goes off the rails, it's hard to get it back on track, so it can be a bit frustrating to work with in comparison. 1: https://github.com/synthetic-lab/octofriend https://github.com/synthetic-lab/octofriend
- nico 1y agoHow are you using glm-4.5? Are you consuming the api or running something like glm-4.5 air locally?
- reissbaker 1y agoI run a privacy-focused inference company, Synthetic [1], and I use our API of course :P I actually like GLM-4.5 enough that it's currently our default recommended model for new users. But yes, otherwise I'd use the official zai API most likely, or Fireworks. GLM-4.5-Air is quite good for a local model but GLM-4.5 is better; up to you if the tradeoff is worth it — there's definitely value in the data not ever leaving your machine, but it's not going to be as strong of a model. 1: https://synthetic.new https://synthetic.new
- azinman2 1y agoI’m curious for your service, if it’s centered around privacy, why is the data stored for 14 days at all? My understanding with fireworks is that it’s 0 logging — nothing to store. To me that’s private.
- reissbaker 1y agoGreat question! We actually don't store prompts or completions for the API at all — but legally we say 14 days so that if someone accidentally commits a log statement, we're not in breach as long as we catch it quickly and revert.
- mrtesthah 1y agoAmazing! So I’m assuming that because it’s privacy focused, you accept payment in cryptocurrencies like Monero and Zcash?
- reissbaker 1y agoWe accept USDC and USDP crypto payments via Stripe. We don't currently support Monero or Zcash — right now all our payments are via Stripe since it simplifies security + compliance for us. It would be a pretty neat feature to build though.
- starchild3001 1y agoReally appreciate the depth of this paper; it's a welcome change from the usual model announcement blog posts. The Zhipu/Tsinghua team laid out not just the 'what' but the 'how,' which is where the most interesting details are for anyone trying to build with or on top of these models. The post-training methodology (Sec 3) is what really stands out to me. The idea of creating specialized 'expert models' for reasoning, agents, and chat, and then distilling their capabilities into a final unified model is a fascinating approach. It feels like a more structured way to solve the "jack of all trades, master of none" problem that can plague generalist models. Instead of just mixing all the data, they're essentially having a generalist learn from a committee of specialists. A couple of the findings from their RL experiments are pure gold for anyone working in this space. The counter-intuitive result that a single-stage RL process at the full 64K context length outperforms a progressive, multi-stage approach (Fig 6) is a fantastic lesson. I've seen teams assume the opposite would be true. Also, the pragmatic choice to use an XML-like template for function calls to avoid JSON escaping hell (Fig 4) may be a small but brilliant engineering decision that makes a huge difference in practice. Wrangling escaped code inside JSON turns out to be a mess. The performance on SWE-bench is impressive, putting it in the same league as much larger or proprietary models. What I’d love to see, and maybe others here have thoughts, is whether this hybrid training recipe holds up outside ARC-style evals. For example, do the agentic improvements transfer to messier, real-world workflows where APIs are undocumented, partial failures are common, and user input is full of ambiguity?
- algo_trader 1y agoAre all these "post/mid-training tweaks" important if you have a specific domain with abundant/verified/synthesis data and labels? Can a small team working on ASI/domain-specific stick to scaling 2024-era best practices training stack? Or will they miss massive improvements?
- starchild3001 1y ago> Are all these post/mid-training tweaks important with abundant, verified, synthetic domain data? No. Many are aimed at cleaning/aligning noisy, mixed-domain data. With abundant, high-quality domain data, you can skip most of the complexity and focus on direct SFT/RL on your corpus. > Can a small team stick to scaling 2024-era best practices? 2024 was the year of SFT. I believe fitting reasoning traces to your final responses via RL is the technique-du-jour of 2025. Jumping from SFT to RL training might be biggest gain here if RL can be applied to your problem (e.g. math, coding etc).
- sagarpatil 1y agoI’ve been using it and I think it’s on par with sonnet.
- chvid 1y agoThis is a great model for software development - probably the best of the freely available ones.
- natrys 1y agoYep I think it's the best, period. Qwen3-coder perhaps took the limelight but the GLM models perform and behave better in agentic loops. I cannot believe they had gone from a 32B frontend focused GLM-4 to these beasts that can challenge Claude, in a matter of months.
- buyucu 1y agoHuge respect to the open-source culture in China. The Chinese are really leading the world in democratizing AI.
- quantumHazer 1y agoIt's open-weight, not open source though
- buyucu 1y agoClose enough for now. I would love if these companies also published their datasets, but publishing datasets is just asking for legal trouble these days. OpenAI, Anthropic and others are not even publishing their weights.
- almaight 1y agoso sarcastic when it first came out, but now you are so polite.
- promiseofbeans 1y agoThe sheer number of things "they observed" in this paper that could be whole papers in themselves is astounding! Lots of great stuff in here around training processes and data collection+synthesis. Does anyone have any background information on the authors? Have they published similarly impressive works in the past?
- jononor 1y agoSeems like we may get local, open, workstation-grade models that are useful for coding in a few years. By workstation-grade I mean a computer around 2000 USD, and by useful for coding I mean around Sonnet 4 level. Current cloud based models are fun and useful, but a tool that is / will be so core to the developer experience, I want to be able to run locally.
- cft 1y agoThis will be essential for the open source. Otherwise open source development will become unsustainable. I'm actually a little bit more optimistic. I think we will get something more than Sonnet 4 level in two years, that can be run on a $2,000 machine.
- unshavedyak 1y agoI've been waiting to do exactly this. Improved latency and no throttling is super appealing to me.
- orangebread 1y agoI would say by end of year, not 2 years.
- jononor 1y agoYeah, if we are lucky it might happen that fast!
- khalic 1y agoImpressive work! Can’t wait to spin it up on HF
- sriku 1y agoOk so I asked "how many bs in bluberry" and it "thought" the following - ----- First, the question is: "How many bs in bluberry?" That seems like a typo. I think it's meant to be "blueberry," but it's spelled "bluberry" with one 'e' missing. Also, "bs" probably refers to the letter 'b'. So, rephrasing the question: How many times does the letter 'b' appear in the word "blueberry"? But the word given is "bluberry," which might be a misspelling of "blueberry." Let me confirm the correct spelling. The standard spelling is "blueberry," which has two 'e's: b-l-u-e-b-e-r-r-y. The question says "bluberry," which is missing the second 'e', so it's spelled b-l-u-b-e-r-r-y. That might be intentional or a mistake. I should answer based on the word as given. The question is: "How many bs in bluberry?" So, I need to count the 'b's in "bluberry." Let me write it out: B-L-U-B-E-R-R-Y. Now, I'll go through each letter: Position 1: B (that's one b) Position 2: L (not b) Position 3: U (not b) Position 4: B (that's the second b) Position 5: E (not b) Position 6: R (not b) Position 7: R (not b) Position 8: Y (not b) So, there are two 'b's in "bluberry." But is "bluberry" the correct word? It's probably a misspelling of "blueberry." In "blueberry," it's B-L-U-E-B-E-R-R-Y. Position 1: B (one) Position 2: L Position 3: U Position 4: E Position 5: B (two) Position 6: E Position 7: R Position 8: R Position 9: Y "Blueberry" has nine letters, and there are two 'b's: one at the start and one after the 'e'. In "bluberry," as given, it's eight letters: B-L-U-B-E-R-R-Y, with two 'b's. Perhaps the question is tricking me with the misspelling. But I should answer based on what's written. Another thought: "bs" might be plural for 'b', meaning how many 'b's, or it could be something else. But in context, it's likely about the letter. Maybe it's a riddle. "Bluberry" sounds like "blueberry," but spelled differently. Perhaps "bs" refers to something else, like the fruit or something, but that seems unlikely. The question says "bs," which is probably "b's" meaning the letter b. Let me read the question carefully: "How many bs in bluberry?" It might be a play on words. "Bluberry" could be intended as "blueberry," and "bs" might be misinterpreted. Another idea: in some contexts, "b" could stand for something, but here it's probably straightforward. Perhaps it's about the sound. But the question is written, so it's about spelling. I think I should just count the 'b's in the given word. So, in "bluberry": letters are B, L, U, B, E, R, R, Y. So, B appears twice. Therefore, there are two 'b's. But to be thorough, let's consider if "bluberry" is a word. It's not standard; it's a misspelling. However, for the purpose of this question, I'll use it as is. Perhaps the question has a typo, and it's meant to be "blueberry." But since it's written as "bluberry," I'll go with that. Final answer should be 2. The question says "bs," which might be lowercase, but in counting, case usually doesn't matter, and here it's all lowercase or mixed? The word is "bluberry," which is lowercase, and "bs" is lowercase, so it's fine. In English, when counting letters, we don't distinguish case unless specified. So, I think two is correct. But let me double-check the spelling. The user wrote "bluberry," which has one 'u' and no 'e' after 'u', but in "blueberry," it's "blue" with 'e'. In "bluberry," it's "blu" without 'e', then "berry." But in terms of letter count, it's clear. Perhaps it's "blueberry" misspelled, and I should correct it, but I think for accuracy, I'll use the given spelling. Another thought: the question might be "how many b's in blueberry?" but written with a typo. But since it's "bluberry," I'll answer based on that. So, response should be: There are two 'b's in "bluberry." To make it clear, I can say: The word "bluberry" contains the letter 'b' twice. But the question is "how many bs," so I can just say 2. Perhaps it's a trick with the word "berry" having one 'b', but "blu" has one, so total two. "Blu" has one 'b', "berry" has one 'b', but in "bluberry," it's combined, so still two 'b's. Yes. I think I'm overthinking it. Final answer: 2 -----
- throwdbaaway 1y agoSo GLM-4.5 series omits the embedding layer and the output layer when counting both the total parameters and the active parameters: > When counting parameters, for GLM-4.5 and GLM-4.5-Air, we include the parameters of MTP layers but not word embeddings and the output layer. This matches with the calculation I did for GLM-4.5 (355B A32B): In [14]: 356732107008 - (775946240 * 2) # token_embd / output are 775946240 each. assume omitted Out[14]: 355180214528 In [15]: 356732107008 - 339738624000 - (775946240 * 2) # parameters that are always active Out[15]: 15441590528 In [16]: 339738624000 * 8 / 160 # parameters from activated experts Out[16]: 16986931200.0 Meanwhile, GPT OSS series includes both the embedding layer and the output layer when counting the total parameters, but only includes the output layer when counting the active parameters: > We refer to the models as “120b” and “20b” for simplicity, though they technically have 116.8B and 20.9B parameters, respectively. Unembedding parameters are counted towards active, but not embeddings. And Qwen3 series includes both the embedding layer and the output layer when counting both the total parameters and the active parameters. Why there is no standard in counting? Which approach is more accurate?
- atq2119 1y agoI'd say it depends. For the total parameter count, you should just count all parameters, since that's what matters for memory requirements. For activated parameters: All unembedding parameters are used in every inference step during token generation, but only one column of the embeddings is used (if done right). So count accordingly, since that's what matters for memory bandwidth and therefore latency.
- throwdbaaway 1y agoThat makes sense, thanks for the info. Here's a quick recap of the recent MoE models based on the criteria.. correct activated params: * DeepSeek V3/R1 series * Kimi K2 * GPT-OSS series undercount activated params: * GLM-4.5 series overcount activated params: * DeepSeek V2 series * Qwen3 series * Ernie 4.5 series * Hunyuan A13B
- matijsvzuijlen 1y agoFor people who, like me, prefer to read the abstract first: https://www.arxiv.org/abs/2508.06471 https://www.arxiv.org/abs/2508.06471
- acqbu 1y agoKimi K2 has been the best so far in my experience
- rsanek 1y agoWhat’s up with the charts? In their first figure they show sonnet 4 as having a swebench score of 53; in the next figure it’s close to 70. In reality it’s closer to the latter. [0] [0] https://www.swebench.com/ https://www.swebench.com/