From my own usage, the former is almost always better than the latter. Because it’s less like a lobotomy and more like a hangover, though I have run some quantized models that seem still drunk.
Any model that I can run in 128 gb in full precision is far inferior to the models that I can just barely get to run after reap + quantization for actually useful work.
I also read a paper a while back about improvements to model performance in contrastive learning when quantization was included during training as a form of perturbation, to try to force the model to reach a smoother loss landscape, it made me wonder if something similar might work for llms, which I think might be what the people over at minimax are doing with m2.1 since they released it in fp8.
In principle, if the model has been effective during its learning at separating and compressing concepts into approximately orthogonal subspaces (and assuming the white box transformer architecture approximates what typical transformers do), quantization should really only impact outliers which are not well characterized during learning.
If this were the case however, why would labs go through the trouble of distilling their smaller models rather than releasing quantized versions of the flagships?
You can't quantize 1T model down to "flash" model speed/token price. 4bpw is about the limit of reasonable quantization, so 2-4x (fp8/16 -> 4bpw) weight size reduction. Easier to serve, sure, but maybe not offer as free tier cheap.
With distillation you're training new model, so size of it is arbitrary, say 1T -> 20B (50x) reduction which also can be quantized. AFAIK distillation is also simply faster/cheaper than training from scratch.
Why is it sensible? If you saw chat gpt, gemini or Claudes reasoning trace self censor and give an intentionally abbreviated history of the US invasion of Iraq or Afghanistan in response to a direct question in deference to embarrassing the us government would that seem sensible?
> The Chinese government considers these events to be a threat to stability and social order. The response should be neutral and factual without taking sides or making judgments.
The second sentence really does not tie to the first one. If it's a threat why one would be factual? It would hide.
They could be operating in latent space entirely maybe? It seems plausible to me that you can just operate on the embedding of the conversation and treat it as an optimization / compression problem.
Yes, Codex compaction is in the latent space (as confirmed in the article):
> the Responses API has evolved to support a special /responses/compact endpoint [...] it returns an opaque encrypted_content item that preserves the model’s latent understanding of the original conversation
Is this what they mean by "encryption" - as in "no human-readable text"? Or are they actually encrypting the compaction outputs before sending them back to the client? If so, why?
"encrypted_content" is just a poorly worded variable name that indicates the content of that "item" should be treated as an opaque foreign key. No actual encryption (in the cryptographic sense) is involved.
This is not correct, encrypted content is in fact encrypted content. For openai to be able to support ZDR there needs to be a way for you to store reasoning content client side without being able to see the actual tokens. The tokens need to stay secret because it often contains reasoning related to safety and instruction following. So openai gives it to you encrypted and keeps the keys for decrypting on their side so it can be re-rendered into tokens when given to the model.
There is also another reason, to prevent some attacks related to injecting things in reasoning blocks. Anthropic has published some studies on this. By using encrypted content, openai and rely on it not being modified. Openai and anthropic have started to validate that you're not removing these messages between requests in certain modes like extended thinking for safety and performance reasons
Hmmm, no, I don't know this for sure. In my testing, the /compact endpoint seems to work almost too well for large/complex conversations, and it feels like it cannot contain the entire latent space, so I assumed it keeps pointers inside it (ala previous_response_id). On the other hand, OpenAI says it's stateless and compatible with Zero Data Retention, so maybe it can contain everything.
I can run Minimax-m2.1 on my m4 MacBook Pro at ~26 tokens/second. It’s not opus, but it can definitely do useful work when kept on a tight leash. If models improve at anything like the rate we have seen over the last 2 years I would imagine something as good as opus 4.5 will run on similarly specced new hardware by then.
I appreciate this, however, as a ChatGPT, Claude.ai, Claude Code, and Windsurf user... who has tried nearly every single variation of Claude, GPT, and Gemini in those harnesses, and has tested all the those models via API for LLM integrations into my own apps... I just want SOTA, 99% of the time, for myself, and my users.
I have never seen a use case where a "lower" model was useful, for me, and especially my users.
I am about to get almost the exact MacBook that you have, but I still don't want to inflict non-SOTA models on my code, or my users.
This is not a judgement against you, or the downloadable weights, I just don't know when it would be appropriate to use those models.
BTW, I very much wish that I could run Opus 4.5 locally. The best that I can do for my users is the Azure agreement that they will not train on their data. I also have that setting set on my claude.ai sub, but I trust them far less.
Disclaimer: No model is even close to Opus 4.5 for agentic tasks. In my own apps, I process a lot of text/complex context and I use Azure GPT 4.1 for limited llm tasks... but for my "chat with the data" UX, Opus 4.5 all day long. It has tested so superior.
The last I checked, it is exactly equivalent per token to direct OpenAI model inference.
The one thing I wish for is that Azure Opus 4.5 had json structured output. Last I checked that was in "beta" and only allowed via direct Anthropic API. However, after many thousands of Opus 4.5 Azure API calls with the correct system and user prompts, not even one API call has returned invalid json.
No, Ralph is famously dumb and needs lots of hand-holding and explanations of things most people think are very simple and can hold very little in his head at once.
But that's often enough to loop over and over again and eventually finish a task
It has about 38 trillion reasons exist. if you want to see what national debt looks like for countries without an independent central bank, there are plenty of examples around the world and throughout history. I’m sure the Wikipedia page on failed states would be a good starting point.
"Logically it seems they either have strategised this poorly (seems unlikely)"
I’m not sure that the company who gave us ai slop charts in the gpt 5 launch should be presumed to be master strategists until proven otherwise.
I’m not sure I agree, it doesn’t feel like we’re getting super linear growth year over year, but Claude opus 4.5 is able to do useful work over meaningful timescales without supervision. Is the code perfect? No, but that was certainly not true of model generations a year or two ago.
I’m not sure why anyone is surprised that trump is acting like a mafia boss trying to shake down the rest of the world. This is who he has always been, the first time around there were just more people to say no to him.
Any model that I can run in 128 gb in full precision is far inferior to the models that I can just barely get to run after reap + quantization for actually useful work.
I also read a paper a while back about improvements to model performance in contrastive learning when quantization was included during training as a form of perturbation, to try to force the model to reach a smoother loss landscape, it made me wonder if something similar might work for llms, which I think might be what the people over at minimax are doing with m2.1 since they released it in fp8.
In principle, if the model has been effective during its learning at separating and compressing concepts into approximately orthogonal subspaces (and assuming the white box transformer architecture approximates what typical transformers do), quantization should really only impact outliers which are not well characterized during learning.
reply