FR version is available. Content is displayed in original English for accuracy.
Advertisement
Advertisement
⚡ Community Insights
Discussion Sentiment
82% Positive
Analyzed from 3611 words in the discussion.
Trending Topics
#model#models#more#flash#qwen#https#memory#next#knowledge#bit
Discussion (124 Comments)Read Original on HackerNews
1. It overthinks — Just like the previous iteration. High confidence. 2. It doesn’t overthink — Improvement from the last model for your use case. Regression for others. 3. It sometimes overthinks — Best case all around. A feature, not an impairment.
One final thing worth mentioning: (I made myself irrationally angry writing this)
> [UGC styled humorously as LLMisms]
All joking aside, having interacted with Claude intensely for the last 8 months and about 30 hours/week in the last 3, I’ve started to notice how (for want of a better word) “readable” (“digestible” ? “comprehensible” ? “Predictable” is the wrong direction.) information chunked into LLM-shaped pieces are for me.
I can digest LLM-shaped pieces of data very easily probably because I’ve been spending too much time with Claude, sure.
But the other side of this is that the entire human species (using LLMs) is similarly being trained to digest interrelated pieces of information/data in these specific shapes, akin to how philosophical assertions can be formulated as a syllogism and, thus, become more readily understood because of familiar epistemological cadence and shape.
Many people reject such copy/prose/data because they detect AI-generated-so-not-worth-human-attention, but I do wonder if this is preparing many millions of loosely (and tightly) associated humans and their organizations to quickly exchange and digest information.
This is not to say current LLMisms are the end, only that such detectable patterns in information delivery will make comprehension and communication more efficient (as well as more limited precisely because of such structure).
/philosophical musings about the epistemological implications of LLM-shaped conversation tics
I find LLMisms very annoying to read, it’s almost like they are bullet points in the shape of a paragraph. It feels very “skippy” to me.
This. I don't know if the "honest answer" phrasing is part of the system prompt or alignment, but when people say "honestly" all the time I start wondering how honest they're being.
For example, even if you make thinking tokens literally just '....' (absolutely meaningless; zero information), you still see significant performance improvements: https://arxiv.org/abs/2404.15758 and https://arxiv.org/abs/2607.22925 for some starters.
Treat thinking more like a "loading screen message" that's been RL'd to somewhat resemble its actual internal state; which happens in its activations, not tokens.
Conversely I've found that it can be as succinct as Muse Glimmer when it has a clear path forward. This can be either through well defined requirements or through unambiguous steps to take based on its own reasoning. While I do think it's fair to call out how much smaller model overthinks especially on one-shot prompts, in practice it hasn't led to an overall increase in time to task completion at least for what I've been using it for.
As a result, qwen3.8 will churn over a prompt often for 5-10 minutes while gemma4 regularly finishes the same prompt in under 20 seconds, while giving a consistent and accurate response in my favorite test case. Qwen3.8, despite churning like that, often misses with an inaccurate answer.
Obviously, 'YMMV' depending on your use case... just sharing my two cents.
Also, heating my home during the winter is nice.
Oh, also, I use llamacpp with --reasoning-budget; very simple way to move on.
https://tools.simonwillison.net/markdown-svg-renderer#url=ht...
Surprised I didn't get one I liked as much as the Qwen 3.8 27B one https://simonwillison.net/2026/Aug/16/qwen-38-27b/#the-defau... , maybe because of quantization.
Opus 4.6 Max self-hosted at 30 tok/s on a 5k Macbook in Aug 2026. The LLM timelines are crazy.
Luna is $0.20 / $1.20 vs $0.16 / $0.47 with Qwen.
Why?
In Artifical Analysis's cost per task, Luna(max) costs $0.05 per task, and Qwen 3.8 27B costs $0.25 per task, a 5X increase. We'll see how 3.8-flash-next does.
You can't search what you don't even know exists.
Ex - nodejs natively supports a huge set of typescript with built-in type stripping these days. But ask most hosted models to build a typescript project and they default to a heavy compile step, or a tool like tsx, ts-node, etc.
Models with lots of "world knowledge" have a good chunk of that knowledge go stale, and there's no real way to refresh it without training a new model.
Another classic example of this back in the day was to ask who the president of the US was, and watch different models happily give different answers based on the date they were trained.
---
Personally, I'm really interested to see if we're headed towards a spot where the model is entirely distinct from the knowledge store.
We're vaguely there with the ability for models to go search the web, but I think the reliability of that path is going to continue declining (more and more spam content, less and less genuine value).
I kinda want a paradigm where I can pick and engine and a knowledge bank, and combine them as I please.
Ex - if I'm doing gardening, I can pick "gardening for models (version 32)" as my knowledge store.
If I'm doing auto-repair... "cars for dummies (version 3)". etc...
In a discussion on economic history, say, someone will opine that Alexander Hamilton had some particular opinion about tariff policy… based on their having a vague memory of a blog post where someone quoted a passage in support of some point. But wait - you can search the federalist papers, the text’s right there to be read, before you commit to saying online ‘Hamilton thought tariffs were a great idea’ you could take your internal ‘I seem to recall reading something about hamilton’s opinion on tariffs’ thought and turn it into a little RAG query where you pull up a source and check before you put another factoid out onto the internet.
And so I feel absolutely the same way about LLMs. I don’t care how much factual information was in the training data, when the LLM wants to rely on something it vaguely recalls having been trained on, it owes it to me to dig up a source and vet it.
There are limits to this, of course. I don’t want it to be thinking ‘but wait, maybe my memory of Python syntax is faulty. Is = used for assignment? <web search>…’.
But in general some caution about repeating vaguely recalled easily checked facts is warranted.
This is what I've been trying to focus on with local AI for now. I've been trying to build all new documentation so it's more AI friendly. It's been pretty interesting. Qwen-35BA3B with a small prompt does a good job of surfacing what I'd consider institutional knowledge.
I've been trying to silo the docs I write from the model with a prompt that tells it not to use general knowledge unless asked to. From the anecdotal testing I did, Qwen-35BA3B is great for it. It does a really good job of following the prompt and calling tools, so I've been able to play around a lot to see what seems to work best.
Ultimately, I think one of the most effective uses of AI will be having a distinct knowledge store combined with an opinionated agent (and sub-agent) setup along with different models for each task.
Who owns the knowledge store is going to be the big caveat. Right now I think the big online models are trying for generic, persistent memory and I'd be very hesitant to let that happen. Think of having someone with a perfect memory following you around forever, but someone else has the ability to make them disappear. That's not a good situation.
Self-learning/improving would be even better but that's still a long way to go.
Your mac is < $2K in Biden-era dollars. Presumably the economy will eventually recover; maybe in one Moore’s law doubling if the midterms go outrageously well. That’ll be two doublings since the halo launched. I’d expect this model to run on a sub $1K box by then. $2K ought to get you a 512b parameter model at that point. If we have to wait out the rest of the term, the cost cliff will be even more pronounced when it hits.
And that's even with assuming that we can continue to ignore the long-term problems like social security insolvency, the debt bomb, or climate change forever.
How much memory does this translate to and what quantization (if any) were applied?
Didn’t see this mentioned yet. I wonder what this means for the effective size. It’s evidently ~176B paramètres, but how does that get quantized. A 4-bit quant under 100GB seems unlikely, I’m suspecting this won’t run in 128GB unified memory
In principle I like the idea of trading more memory for compute though, even if there’s a memory shortage right now
Likely soon we'll see nvme offloading for ngrams as well. They're just an index, so that should be plenty fast for what it does. LLama.cpp support should come soon as well, and they might do some things with offloading first.
https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF https://unsloth.ai/docs/models/qwen3.8-next
> You will need at least 75 GB of RAM or unified memory to run the model. Its smallest 1-bit quantized version is larger than usual because of the model’s architecture so 1-bit isn't really 1-bit at all. However, this also means the quantization is less aggressive, allowing the model to retain more of its original accuracy than more heavily quantized models.
Lots of RAM required even for the 1-bit, which is already downloadable. Interested to see how well this one works compared to Ornith1.5-35B-A3B I've been running (and quite happy about).
Edit: but llama-cpp does not yet support it.
[0]: https://unsloth.ai/docs/models/qwen3.8-next#qwen3.8-flash-ne...
Original error: llama.cpp does not support this GGUF's model architecture ('qwen4exp')
Edit : Saw the pull request, should arrive soon enough https://github.com/ggml-org/llama.cpp/pull/27742
This will almost certainly require changes to llama.cpp or vllm to do it right.
6B active params helps around the memory bandwidth constraints, but a 128GB box can probably run the Q3/Q4 quants fairly easily with a decent context size. This might actually be better for strix users than 27B, which was already very good.
> trained at just 1/9 the cost of Qwen3.7-Plus, while outperforming it across the board
https://x.com/Alibaba_Qwen/status/2092591393424515114
It's new arch demo for future Qwen 4 family, but (as I understand) training recipe/data is same as any other 3.8 model.
Qwen 3.8 flash: $0.16 / $0.47
Compared to
Deepseek 0723: $0.03 / $0.075
(units in USD/m tok)
https://api-docs.deepseek.com/quick_start/pricing
8t/s though apparently and their cache hit rate is terrible so I don't think it's worth it over Relace.
DSV4 Flash 304B params, 167 GB download (at full size)
Qwen3.8 Flash Next 180B params, 360 GB download (at full size)
Same reason your phone has a few big CPU cores for real work, it's much better to "race to idle" than have an "efficient" core struggle. Shitty experience, shitty power efficiency.
Not to mention, they’re great for self-hosting and getting yourself to not be dependent on some API that can go down or be altered at any time.
Big models seem to mostly be good for pushing ahead the frontier - the smaller models tend to gain the frontier’s capabilities after only a handful of months anyway. Many are perfectly content remaining a few months behind the bleeding edge.
Aside from the pelican, I am sort of impressed by the fact that things are going the way in terms of really impressive small models.
Also I love how this uses N-gram embedding. I think that Longcat was the first one who used it (I submitted that submission on hackernews because I really just loved the idea of it that I understood), I am certainly more interested in local LLM models and its interesting how they are utilizing new architectures to do some really impressive optimizations!
(Do note that I created it using a free rate limited end-point that I found on the huggingface space section: https://victor-chat-with-qwen3-8-flash-next.hf.space)
Wasn't it introduced by Gemma?
I wonder if I could get this running through vLLM on 6x Nvidia L4 - the 3.6 worked great on 4 cards but sadly TP6 just isn’t a thing and I don’t have 8 cards available, maybe it’s gonna be okay with like TP2 and MTP. I have no idea at this time, probably need to test out what even might be possible.
This "next" release adds a new concept, first public release with n-grams, I think. And it's in a MoE size that is likely to be very fast and cheap to serve (faster than 27b for sure). It's also well suited for inference on alternative compute (i.e. sparks, macs, etc) so it's relevant to local users.
:)
It is very relevant and for a certain group of us, far more impactful to our work the next month(s) than any blog post could be.
> trained at just 1/9 the cost of Qwen3.7-Plus, while outperforming it across the board
https://x.com/Alibaba_Qwen/status/2092591393424515114
Qwen's advances do (currently) have merit.