Confidence intervals have nothing to do with run-to-run variation. They have little to do with anything people usually ascribe to them (https://link.springer.com/article/10.3758/s13423-015-0947-8 ), but even less with run-to-run variation (https://link.springer.com/article/10.1007/s10654-016-0149-3 misconception 22).

It is a different problem if we pick two sets from the same data distribution, A and B, and first we have a score on A, then on B. Here we re-run on precisely the same set of Terminal Bench 2.1 problems. It may be that results are so random between runs that each single task has the same probability in a Bernoulli distribution. But more likely, many problems are easy (i.e. each run will solve them consistently), many are too hard (i.e. no run is going to solve them) and only a fraction is somehow in between.

Maybe there is some good trick to find a proper distribution, but to my knowledge, we would need to run it at least two times on TB2.1 to get any more educated estimates. That said, I am open to new ideas.

That said, I consider frequentist probability a dirty trick, and that Bayesian is the proper way of doing things (vide David J.C. MacKay" Information Theory, Inference, and Learning Algorithms" and Cam Davidson-Pilon "Probabilistic Programming & Bayesian Methods for Hackers" https://www.inference.org.uk/itprnn/book.pdf, https://dataorigami.net/Probabilistic-Programming-and-Bayesi...).

I'm not that well versed in statistics, but a standard box plot is probably the best alternative

Saying there's a confidence interval for run-to-run makes no sense, from the way I understand it, because each run could have a result that's all over the place.

Let’s say that the “true” distribution of the data has mean μ=100 and standard deviation σ=15, but we don’t know that.

95% confidence interval for μ = “if we repeatedly draw N samples from the true distribution and compute a confidence interval every time, 95% of those intervals will contain μ.” That’s all that the definition of a confidence interval guarantees. It does not follow that if we take one of those intervals, it, specifically, has a 95% chance of containing μ. For a frequentist, that’s a meaningless statement (both the interval and μ are fixed so there’s no frequentist probability about it); for a Bayesian, there is no guarantee that that probability is 95%. 95% is instead the probability of “sampling data that will happen to generate an interval that contains μ”.

95% Bayesian credible interval for μ = interval that can be interpreted as having a 95% probability of containing μ, generally obtained by computing the posterior probability density distribution for μ and finding an interval that encompasses 95% of the probability mass. Conventions include highest-density intervals (HDIs), which are obtained by making sure that the PDF is equal at both bounds, and equal-tailed intervals (equal probability mass before and after the interval). With enough samples, it may become arbitrarily narrow (“we are very sure of the mean”), despite the standard deviation of 15 that is built into the “true” distribution that we are estimating, and a Jeffreys prior will happen to make it satisfy the definition of a confidence interval as well (https://sami.boo/jaynes/confidence-intervals-vs-bayesian-int... ).

Posterior predictive distribution = taking into account the uncertainty on both μ and σ, distribution of samples that would be obtained by sampling from N(μ, σ) (which, because of that uncertainty, is a https://en.wikipedia.org/wiki/Compound_probability_distribut... but may have a convenient closed form https://en.wikipedia.org/wiki/Conjugate_prior#Table_of_conju... ), from which we can likewise extract a 95% interval.

Correct, which is why it's crazy to consider a Jeffrey's credible interval as valid, but not a confidence interval.

Perhaps the confidence interval might not be valid under a frequentist interpretation, but the same math can be valid under a Bayesian interpretation even if the person publishing didn't mean it that way.

Does that mean it's valid? Well, the confidence interval is as equally valid as a lot of credible intervals, which is why the credible interval is not some magical better alternative - it's mostly just a choice of how you interpret it, it is not a different procedure that produces different numbers.

The confidence interval is “valid” for what its definition is. But:

- saying “there is a 95% chance that this interval contains the true value because it’s a 95% credible interval” is valid;

- saying “there is a 95% chance that this interval contains the true value because it’s a 95% confidence interval” is not.

It’s very much possible to reach a correct conclusion with invalid reasoning. As Jaynes put it:

“As we stress repeatedly, from a false proposition – or from a fallacious argument that leads to a false proposition – all propositions, true and false, may be deduced. But this is just the danger; if fallacious reasoning always led to absurd conclusions, it would be found out at once and corrected. But once an easy, shortcut mode of reasoning has led to a few correct results, almost everybody accepts it; those who try to warn against it are not listened to.”

> Perhaps the confidence interval might not be valid under a frequentist interpretation, but the same math can be valid under a Bayesian interpretation even if the person publishing didn't mean it that way.

Right, but if you get that by happenstance and not from the fact that it’s a confidence interval, the latter isn’t really getting you anything.

Per my first link: “One of the misconceptions regarding the relationship between Bayesian inference and frequentist inference is that they will lead to the same inferences, and hence all confidence intervals can simply be interpreted in a Bayesian way. In the case where data are normally distributed, for instance, there is a particular prior that will lead to a confidence interval that is numerically identical to Bayesian credible intervals computed using the Bayesian posterior (Jeffreys 1961; Lindley 1965). This might lead one to suspect that it does not matter whether one uses confidence procedures or Bayesian procedures. We showed, however, that confidence intervals and credible intervals can disagree markedly. The only way to know that a confidence interval is numerically identical to some credible interval is to prove it. The correspondence cannot — and should not — be assumed. More broadly, the defense of confidence procedures by noting that, in some restricted cases, they numerically correspond to Bayesian procedures is actually no defense at all. One must first choose which confidence procedure, of many, to use; if one is committed to the procedure that allows a Bayesian interpretation, then one’s time is much better spent simply applying Bayesian theory. If the benefits of Bayesian theory are desired — and they clearly are, by proponents of confidence intervals — then there is no reason why Bayesian inference should not be applied in its full generality, rather than using the occasional correspondence with credible intervals as a hand-waving defense of confidence intervals.”

> which is why the credible interval is not some magical better alternative - it's mostly just a choice of how you interpret it

You don’t get to choose to interpret a confidence interval as if it had the meaning of a Bayesian one. If you want to be able to interpret a computed interval as “there is a 95% chance that this interval contains the true parameter value”, only the Bayesian procedure guarantees it, and it can very well produce different numbers. So it is a better alternative in that sense.

https://sami.boo/jaynes/confidence-intervals-vs-bayesian-int... (+ Cauchy example below)

It’s like with test accuracy. Test accuracy is the pre-test probability that the test will give a correct result. But once you have a positive or negative result, which way it turned out plays a part in computing the predictive value. Likewise, once you have computed the interval, the specific bounds you ended up getting can affect the plausibility that they contain the true value.

The conclusion they come to is "it is possible to do better in the individual case by taking into account evidence from the sample that the confidence interval method throws away". That means the confidence interval is inefficient, not incorrect.

But once you have picked one, and you know its bounds (say, [12.1471, 13.8264]), then it’s fallacious to make the post-data reasoning that “because it was picked at random from the set of 90% confidence intervals, it has a 90% chance of containing the true parameter”.

Again, it’s like with medical tests. If a test has 90% sensitivity and 90% specificity, it has 90% accuracy (it will, in 90% of cases, produce a result that matches disease status) – a pre-data statement on the test result (/ on the confidence interval that we will compute). But it does not follow that, if you screen an asymptomatic patient with low prior probability of disease and get a positive result, they have a 90% chance of having the disease – a post-data statement on disease status, given the test result (/ on where the parameter lies, given the interval).

> The conclusion they come to is "it is possible to do better in the individual case by taking into account evidence from the sample that the confidence interval method throws away". That means the confidence interval is inefficient, not incorrect.

It means you know in the individual case that the specific confidence interval does not in fact have a 90% chance of containing the parameter.

The examples that actually ruin the number so far have been asymmetrical. Is there a way to do it with a gaussian? Especially if you're trying to make a reasonable internal?

Sorry, I don’t follow. Let’s say I want to compute a 50% confidence interval for the unknown mean of a Gaussian distribution. I sample two numbers from the distribution, get 9 and 7, compute the interval according to the trivial procedure and get (-∞, ∞). Does the interval (-∞, ∞) have a 50% probability of containing the mean of that Gaussian distribution? I would think it’s closer to 100%.

If what you are saying is “it’s meaningless to talk about the probability of that specific interval containing the unknown-but-fixed parameter” then that’s the purely frequentist view and then you also agree that it’s meaningless to say that [-2.31, 10.31] has a 90% chance of containing the location parameter of the Cauchy distribution that happened to yield the samples 3 and 5. Incidentally, what asymmetry are you referring to in the Cauchy example?

> then that’s the purely frequentist view and then you also agree that it’s meaningless

I'm saying that when you hit "meaningless" you can back up a step to where you actually had randomness and look at that distribution, which gets rid of a lot of these issues.

But after looking at these examples I think it only makes sense in limited circumstances to do that. Like in the trivial example: your final distribution isn't based on the probability of the mean being any particular number. The only probability was back a step and that was 50%.

At this point I still don't think it's objectively wrong to say a particular interval above is 95% likely, but there's too many ways to interpret the statement so nobody should say it is.

The way we're calculating that these intervals are "wrong" is by looking at all the possible parameters that could have given us the samples we got, and checking how often the range contains the parameter. That's a useful calculation but is it the one people expect? I think that depends on the situation. Treating the parameter as being the thing we sample over is misleading, but treating it as fixed is also misleading.

They're really reaching here. There's a 95% chance it's 1 and 5% it's 0, and we don't know which, so it's 0.95. Like saying the probability your uncle died of a shark attack is either 0 or 1. But in reality I don't know anything about your uncle so it's the average probability that someone picked randomly died of a shark attack, until I know it about your uncle.

> Like saying the probability your uncle died of a shark attack is either 0 or 1.

For a frequentist, it is. If you want the Bayesian interpretation of probability, you must calculate the probability that the interval you ended up getting contains the true value, and you can easily end up with a value different from 95%. You can’t just stop at “it was drawn from a pool of intervals of which 95% contain the true value” since you have more information than that (the numerical bounds).

Just like if I pick a ball from a bag of balls of which 50% are red, it has a 50% chance of being red.

  --no-mmproj-offload

I'm on Linux and using some sort of unholy mess of ROCM libraries that I don't understand.

https://huggingface.co/ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF

(I don’t know much about it, just saw a YouTube video about it last night)

https://huggingface.co/Jackrong/Qwopus3.8-27B-Flash-GGUF

Runs the 3bit model faster than the 2bit one runs on my old-ass card. Can’t vouch for its intelligence yet, but i suspect whatever loss in smarts it takes is made up for by the extra resolution.

https://huggingface.co/ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF

Luke of Luke’s Dev Lab on YouTube had a look at it. It seems to outperform the typical 3-bit quantisation but whether it outperforms the new Unsloth dynamic I don’t know.

Unfortunately, the calculus has changed and it seems cheaper to me to just use MiMo V2.5 for pennies or DeepSeek V4 Flash instead of using Qwen anymore unless I need a local model specifically for doing reverse engineering work that gets otherwise rejected.

Have you tried it with MTPLX? I get around 30 tok/s with it, also on an M1 Max with 64GB.

I tried it with the author’s 4-bit quant of Qwen 3.8 27B: https://huggingface.co/Youssofal/Qwen3.8-27B-MTPLX-Optimized... (but no need to download it manually; MTPLX will ask which one you want).

       /data/llm/llama.cpp/build/bin/llama-server
        --threads 4
        --threads-batch 8
        --batch-size 4096
        --ubatch-size 256
        --port 9999
        --temp "1.0"
        --top-p "0.95"
        --top-k "20"
        --min-p "0.0"
        --presence-penalty "0.0"
        --reasoning auto
        --reasoning-preserve
        --reasoning-budget 4096
        --gpu-layers-draft all
        --spec-type draft-mtp,ngram-map-k4v,ngram-mod
        --spec-draft-n-max 3
        --spec-draft-p-min 0.75
        --spec-ngram-mod-n-match 24
        --spec-ngram-mod-n-min 4
        --spec-ngram-mod-n-max 16
        --spec-ngram-map-k4v-size-n 8
        --spec-ngram-map-k4v-size-m 16
        --spec-ngram-map-k4v-min-hits 1
        --n-gpu-layers all
        --ctx-size 131072
        --repeat-penalty 1.0
        --jinja
        --metrics
        --model /data/llm/models/unsloth/Qwen3.8-27B-UD-IQ3_S.gguf
        --chat-template-file /data/llm/models/qwen3.6-chat-template.jinja
        --fit off
        --flash-attn on
        --cors-origins localhost
        --mmproj /data/llm/models/unsloth/Qwen3.8/mmproj-BF16.gguf
        --no-mmproj-offload
        --parallel 1
        --kv-unified
        --cache-type-k q4_0
        --cache-type-v q4_0
        --cache-type-k-draft q4_0
        --cache-type-v-draft q4_0

The chat template is froggeric's fixed qwen template, v22.5 as of today.

Because on the one hand, the prose and the presentation is painful (narrating irrelevant points, nonlinear X-axes, ambiguous chart labels, etc etc),

But on the other hand, the result that I'm assuming the author means to communicate ("on these evals, generation quality seems fairly good") sounds worthwhile to share?

Because I really struggle with this question at the moment. Am I allowed to draw an adverse inference that "if the writeup presents irrelevant text side by side with the data, then this may be a sign that the author does not understand the task that they are attempting to write up"?

So, if there are irrelevant remarks, these are mine. :)

Charts are vibe-coded - but it took quite a bit of hand-holding to get something decent. And the logarithmic scale for model size is my conscious choice (against Claude's initial ideas).

If you ever get to writing a blog post about kv cache quantisation, i'm interested in quantising K differently than V

The line is "Is this an interesting and accurate article that concisely makes it's case".

LLMs love to burn paragraphs writing about nothing which is why it's generally poor writing. Humans can do the same thing if they are trying to make very little information feel more substantial.

I say, stop trying to determine if an LLM was used and start judging based on your subjective measure that you'd have used before LLMs became widespread.

-----

Similarly, if what I said really is true, I would be implying that LLMs are charlatan/fraudster detectors (to some statistical level). And I refuse on principle to believe that that is actually the case.

In a week’s time I might remember the substance of your comment and some of its shape as a matter of course. Nothing LLM-written that I see today will stick, no matter how curated it was.

It is like a rollercoaster that doesn't have any jerk or jounce.

And if you don’t find any of the word choices a little bit perplexing, your brain can just gloss over it. There is no meaning here. It’s like smalltalk.

It really bugs me, this aspect of LLMs, and yet I find myself sounding like a lunatic when I describe it, until others chime in and say they have the same experience.

I use LLMs to generate starter/tutorial material. I may hate the way Claude writes but I absolutely don't hate the way Gemma 4 writes. But I have to continuously consult it in a way that I do not with human-written text, which gets its message across in a more persistent way I find less troubling.

(This non-memorability extends to AI images and video.)

You are right that there are some confounding factors in my life but while I was worried about middle age affecting recall, I find actually I am still remembering stuff humans wrote pretty much fine.

ETA: there is one thing that I have noticed that does affect recall that is specific to LLMs: watching text roll out word-by-word in LLM chat, I think, damages recall. It's engaging the wrong kind of memory and focus.

So I now let it generate and I'm trying to find decent ways to format it e.g. as PDF, to give it the best chance.

This could be generational; there were many studies twenty years ago that suggest that people a decade or so younger than me who grew up with full colour books and magazines and multimedia can read less-linear text layouts more comfortably, for example, so it's not out of the question that there's something generational going on here too.

Or they don't admit to reading fiction...

I use a combination of a Claude Max subscription and local inference, including qwen3.8-27b, 4bit. I have found qwen to be absolutely useless at anything but very specific, surgical code changes. In my experience, for anything even remotely nuanced, a frontier model is required.

Index methodologies here: https://artificialanalysis.ai/evaluations/artificial-analysi...

Also see some specific benchmarks here: https://huggingface.co/Qwen/Qwen3.8-27B e.g. qwen scores 61.7 on swe bench pro, while opus 4.6 scores 53.4.

If you want to argue with the benchmarks, go for it. Fwiw i am not saying qwen3.8-27b is better or as good as the frontier. But i am saying it has crossed the threshold and is now a useful tool for coding and debugging. From my experience, Qwen3.6-35b-a3b was what you describe - it could do surgical edits only.

I think you're referring to a 5060Ti 16GB, yes?

32k context is easily done there. 64k can work with a more aggressive quant, but you lose a bit of speed.

But I don’t quite follow you - how does a more aggressive quant slow it down? Less bits per token means faster inference not slower.

qwen3.8-27b 4bit has a following specifically for being exceptionally gifted for such a small model.

It does clear the useful enough to be worth it bar though.

Zero interest in remote models but local ones if they offer utility, sure.

Runs pretty well on a 7900XTX as well.

It's well-known that while quantization affects the sampling probability distribution (given the same context, which next token is the most probable), Qwen 3.8 27b seems to offset that by just thinking more and as a result eventually finishing the task (benchmark or otherwise).

So as long as the thinking (albeit longer) is sound, this leads to the same success rate (as shown in the article) but potentially at the cost of more tokens and hence more time.

I think it'll be further useful to chart each quantization's used tokens as well, in addition to the success rate.

Thanks for doing and sharing the research!

In the age of the rampocalypse, the peasants may not have a choice between the two.. time it is!

I have another personal benchmark problem (of a very different nature) that Qwen3.8-27B usually can’t solve at all, while Opus and GLM-5.3-Flash solve it consistently and very beautifully.

I have heard good things about Pi which supports auto-compaction, but I can't personally vouch for it since I use my own.

Especially because Qwen3.8 27B does not seem to be good enough to implement the simulator with all its intricacies and tiny subtleties correctly. So it's churning for hours with no real progress, where it just had to explore the initial problem space a tiny bit more.

And very interestingly, even with Q6, turning off reasoning entirely does not make it fall into that trap, or many others, and it more often solves the task in record time. Both because of what I just described, but also because the endless reasoning costs a lot of token, which significantly translates to time on a home rig...

For what it's worth, if you haven't already, you can also let it run overnight (if you have compaction enabled) to see if it ever gets out of that hole. The reason `xhigh` is the default is because 3.8 is trying to optimize for long horizon tasks where monitoring its every single thinking misstep might not be a good use of our time.

Like wanting to implement an entire emulator to figure out something it could easily glean otherwise (and does, in non-thinking mode).

This has been validated on Qwen3.6-35B-A3B and Qwen3.8-Flash-Next, but with the caveat that it was always on quants.

https://github.com/vagrillo/llama.cpp/blob/moe-expansion/ben...

But is there any model that actually works in a decent way at quantization of 1?

I use Qwen3.8 27B Q4_K_M for coding sometimes and therefore need a relatively long context. I settled on q8_0 because it is the only way to fit the model + 100k tokens into 24GB VRAM, but still wonder what am I loosing in quality, and what other options are there.

I also heard that KV cache quantization matters more with longer contexts. It may be interesting to benchmark this too: what the quality looks like on different combinations of model quantization × KV cache quantization × context size.

https://arxiv.org/html/2609.04098

I'm now running NVFP4 quantized both weight and cache on my RTX5090 and getting excellent results: 264k cache allocated for pool, 10k tok/s prompt processing, 200 tok/s generation for single stream, or 801 tok/s generation for 8 concurrent streams. Also have about 2Gb vram left for use of OS.

my coding agents regularly reach 200k context used without noticeable degradation.

P.S. I used setup from: https://github.com/seanyourhighness/vllm-sm12x-nvfp4-dflash2

For the kinds of things I use a local model for (legal document review), it’s just spectacular. It also has good vision support. I’ve been using 27B more and more over DS4.

To my own surprise, Q6_K_XL (from unsloth) comes up with a solution, anything Q5 doesn’t. To further surprise me, so far only the XL Q6 variant managed to solve it.

The problem, at least as stated, seems to be right on the edge of what the Q6 quantization can do.

Unfortunately even a successful run is rather long, so I don’t have a whole lot of data.

But the whole thing sure made me doubt the common idea that you wouldn’t perceive a difference until crossing past 4 bits quantization.

But there I literally did read “you don’t need anything better than 4 bpw” a bunch of times.

That sub is games, porn, and complaints about not having money. A waste of time.

Err… can someone explain to me what is meant here? Surely the model wouldn’t consistently “guess wrong” compared to randomly, as that would be better. I guess some things like general coherency (i.e. is it even readable or gibberish) factor into that score?

On the other hand, not sure where from 25% baseline for random answers come from. Since this is multiple-choice-out-of-4 test, random guessing should be correct in 1 in 15 cases, not 1 in 4.

General purpose agents can need up to 30k just to reply with "1+1=2" because their prompting is so overloaded. 60-70k is decently usable, still not great for anything complex. A long running task in a general purpose agent can easily hit 100k.

What the vast majority of people care about is performance around what desktop consumer GPUs can run. 8GB, 10, 12, and 16GB of VRAM. What do models that will run at full performance, do?

Also important to know is how Qwen3.8-27B stacks up against qwen3.6-35B-A3B, which due to being MoE, will run on a 16GB card with plenty of speed 90% of the time, at higher quant - so you get more parameters and better quant. But 3.8 is supposed to be "better", so...?

But I think there is a bit more to this:

> Measuring token prediction differences (KL-divergence, top-1 predictions) is easy, but it does not tell us whether the model gets worse at solving tasks.

A common issue is that it's rarely mentioned on which dataset KL-divergence is computed. It seems the most common dataset is wikitext - maybe because of tradition, to make numbers more comparable? It's not measuring how well the model follows bf16 on agentic tasks.

I've been trying to check KLD recently for some quants of Qwen 3.8 27B, and the numbers are dramatically different, depending on which dataset you use. KLD computed on agentic traces is much higher, and top-1 % is way lower than if you compute it on chats or wiki text.

You look at a published number, and see "oh, nice, top1 is 99% - quant is different just in 1 token out of 100", but chances are it's computed on wiki, and on agentic / coding it can be 10 tokens out of 100.

Common intuition is that on agentic tasks errors compound, and that's why it degrades more than metrics show - but maybe the metrics themselves are also wrong, too optimistic.

Still investigating it though :)

Thank you for calling this out. Using Wikipedia snippets for these is a terrible choice. I did a bunch of KL and other stats with the five Gemma 4 models, and the results were non-obvious. Anthropomorphizing:

Gemma 4 31B: "I guess we'll pretend I said this, but it's not me." (Baseline for stats)

Gemma 4 26B: "Dude, I'm certain I wouldn't have said this." (Bad KL)

Gemma 4 12B: "Umm, Me either!" (Similarly Bad KL)

Gemma 4 E4B: "I might say almost anything, this is fine." (Much better KL!!!)

Gemma 4 E2B: "I'm basically a toy. Let's play a game!" (Same KL as E4B)

Anyway, for comparing quantizations, it seems like the largest precision version should be given a one-shot prompt, and the result from that should be used as the corpus for the quantized versions.

Way better than wikitext-- but tells you nothing about errors tending to compound or cancel out.

Like say a test shows that only one token in a 10,000 token test would be different. Sounds very close, ship it!-- but what if trajectories with that single different token guarantees failure because it sets in motion a cascade of differences that ultimately result in a final distribution that doesn't include the solution?

On wikitext2 there was no difference. I concluded these have no long-term dependency and I should use Linux kernel. Still, the same.

So yes, KLD depends on the dataset. Still, it does not measure what any e2e test does.

As for KLD, have you tried it on something which is even closer to e2e task, like agentic traces from https://huggingface.co/datasets/nvidia/Nemotron-Cascade-2-SF... or similar datasets?

If you would like to do so, it is easy (and orders of magnitude cheaper) than running benchmarks.

It can't mess with your computer or files until you hook it to a harness, which interprets some of the model's output as commands to execute. So, model on bare metal, harness in a container or VM.

Due in large part to DeepSeek's MoE architecture. Generating 1 token through DS4's 13B active params requires roughly half the FLOPs to generate a token through the dense Qwen 27B.