frontpage.
newsnewestaskshowjobs

Open Source @Github

fp.

Open in hackernews

Qwen3.8-Flash-Next: A New Architecture, Towards Ultimate Cost-Efficiency

https://qwen.ai/blog?id=qwen3.8-flash-next
113•tosh•58m ago

Comments

whwhyb•46m ago
looks like it's better than deepseek v4 flash
freakynit•42m ago
Those benchmarks look seriously impressive.. considering how small of a MoE model this is.
skarz•42m ago
do we really need breaking news about qwen posted every single day?
iAMkenough•40m ago
yes there’s no shortage of online real estate
tosh•37m ago
this is a new architecture (foreshadowing qwen 4)

> trained at just 1/9 the cost of Qwen3.7-Plus, while outperforming it across the board

https://x.com/Alibaba_Qwen/status/2092591393424515114

pseudony•31m ago
I and presumably quite a few others with AMD AI or Apple Mac platforms are very impacted by this.

:)

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.

KronisLV•28m ago
If there’s news, then yes. This is a pretty great new release for those still stuck on Qwen3.6 35B A3B if they have enough memory but don’t have super powerful compute.

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.

NitpickLawyer•24m ago
This particular release is interesting because it's a preview of qwen4 architecture. And, while benchmarks are iffy, this is a direct comparison, by the same team, with qwen3.8-27b that was pretty well received for a local model.

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.

dofm•21m ago
This actually is meaningful news, I think. Pretty wide audience appeal in the local LLM space too.
c16•20m ago
There are many topics, personalities and politicians we hear about daily who have no merit.

Qwen's advances do (currently) have merit.

christkv•37m ago
Looks like a good model for strix halo
Iolaum•26m ago
indeed, can't wait for it to be supported by llama.cpp (or other engines)?
tosh•36m ago
this is a new architecture (foreshadowing qwen 4)

> trained at just 1/9 the cost of Qwen3.7-Plus, while outperforming it across the board

https://x.com/Alibaba_Qwen/status/2092591393424515114

rohansood15•36m ago
Didn't expect it to beat 3.8 27B so cleanly.

Opus 4.6 Max self-hosted at 30 tok/s on a 5k Macbook in Aug 2026. The LLM timelines are crazy.

Squarex•29m ago
I don't like these comparisons. Sure it is impressive, but it does not have a world knowledge of larger models. It has most of theirs intelligence.
LaurensBER•20m ago
If/when we can get larger context this will mostly be mitigated by these smaller models being able to search the internet.

Self-learning/improving would be even better but that's still a long way to go.

rohansood15•16m ago
For world knowledge, you'd want it to find and reference the source material to be sure. At that point, it doesn't matter if the knowledge is embedded.
gruez•26m ago
>Opus 4.6 Max self-hosted at 30 tok/s on a 5k Macbook in Aug 2026. The LLM timelines are crazy.

How much memory does this translate to and what quantization (if any) were applied?

rohansood15•15m ago
128GB, 4-bit quantized.
lnenad•20m ago
Adding to my homelab stack, hopefully it doesn't overthink like the little model. Actually, hoping it thinks a bit less. Wait actually I'm really praying it reasons a bit more directly. But wait, I'm really sure that it must be a bit better.
grim_io•17m ago
That's low reasoning for a model, but max for a HN comment.
cyanydeez•7m ago
My stack is basically deer-flow with Qwen3.5-122B-A10B; this hopefully will be a speed and intelligence improvement. Running deer-flow overnight on any research topic or verify clear scoped programming issue is really neat.

Also, heating my home during the winter is nice.

Oh, also, I use llamacpp with --reasoning-budget; very simple way to move on.

martinald•20m ago
FYI: nothing seems to be able to run this (easily) yet. llama.cpp, vllm etc I couldn't get working because of no support in the mainline version.
kzrdude•4m ago
They are giving pointers to how to run it now using for example https://recipes.vllm.ai/Qwen/Qwen3.8-Flash-Next (and an especially provided vllm release).
amclennon•20m ago
It looks like this also undercuts the already absurdly inexpensive Deepseek Flash in pricing. Wild.
kzrdude•12m ago
Great. It's also smaller than DSV4 Flash, so it makes sense that way.
geooff_•10m ago
Where are you seeing that? At the bottom of this post from Qwen I see:

Qwen 3.8 flash: $0.16 / $0.47

Compared to

Deepseek 0723: $0.03 / $0.075

(units in USD/m tok)

pram•19m ago
It's in Unsloth Desktop already. Looks like it's 73GB, so 128GB Mac or Strix Halo etc will work. Exciting!
cwizou•14m ago
Download is available, but likely need to wait for an update, I get this which is understandable with the architectural change :

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

KolmogorovComp•15m ago
Will this be cheaper than DS4flash ?
armcat•10m ago
How is input token efficiency/verbosity on this model? Has anyone tried? GLM 5.2 was doing lot of turns and thinking piling up input tokens in the context (compared to Claude and GPT models). Then Qwen3.8-27B was 2x of that. Both delivered good output results but those cumulative input token costs were not cheap. Note this is on our specific business workloads. Genuinely interested in other people's experience (if you are able to try it out).
a_humean•5m ago
Waiting for llama.cpp support to land, but this might be a big deal for Strix Halo users.

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.

user43928•25m ago
For comparison with hosted models, GPT 5.6 Luna scores 67% on DeepSWE, compared to 59% here for Qwen.

Luna is $0.20 / $1.20 vs $0.16 / $0.47 with Qwen.

rohansood15•13m ago
This is a good counter argument. But you have to note that this is after OpenAI cut Luna costs by 80%. If you compare launch pricing, Qwen probably comes out ahead on a cost-performance basis.

Intent to Unship: XSLT

https://groups.google.com/a/mozilla.org/g/dev-platform/c/ZRyt5CzAGfA/m/ckDVFA-8EQAJ
1•tambourine_man•1m ago•0 comments

The Pay-to-Play Model Has Broken Youth Sports

https://nymag.com/intelligencer/article/youth-sports-private-equity-travel-leagues.html
1•sparklingmango•1m ago•1 comments

Indian workers seeking U.S. green cards could face 179-year wait, report finds

https://www.sfchronicle.com/bayarea/article/green-card-immigration-us-22403330.php
1•littlexsparkee•2m ago•0 comments

Access to Urban Woodlands Linked with Lower Use of Antidepressants

https://e360.yale.edu/digest/scotland-woodlands-antidepressants
1•speckx•3m ago•0 comments

Show HN: Otter: experimental multi-threaded JavaScript runtime written in Rust

https://github.com/gi-dellav/otter
1•gidellav•3m ago•0 comments

Why to Start a Startup in a Bad Economy (2008)

https://paulgraham.com/badeconomy.html
1•tosh•4m ago•0 comments

AI Is a Harsh Mistress

https://cacm.acm.org/opinion/ai-is-a-harsh-mistress-on-anima-machina-herd-acceptance-and-the-poli...
1•fscaramuzza•5m ago•0 comments

Wall Street Is Counting on Nvidia to Keep the AI Party Going

https://www.wsj.com/finance/stocks/wall-street-is-counting-on-nvidia-to-keep-the-ai-party-going-7...
1•tosh•6m ago•0 comments

Made two semgrep rulesets for auditing MCP servers

https://github.com/ShroudLabs-io/tlad-semgrep&&https://github.com/ShroudLabs-io/mcp-sinks.git
1•TDouley•6m ago•0 comments

How strict should your schema be?

https://typedb.com/blog/how-strict-should-your-schema-be
1•flyingsilverfin•7m ago•0 comments

Show HN: CueMap – deterministic-first memory retrieval for continuous recall

https://github.com/cuemap-dev/cuemap
1•kaandemirel•8m ago•0 comments

California says states, Meta agree to $17B settlement in child safety trial

https://www.npr.org/2026/08/26/nx-s1-5944781/meta-settlement-child-safety-lawsuit
1•rbanffy•8m ago•0 comments

DuckLabs to Join AWS, Projects to Remain Open Source – DuckDB

https://duckdb.org/2026/08/26/ducklabs-to-join-aws
1•xnx•8m ago•0 comments

"Police have arrested 3 of my mates because of our private WhatsApp group chat"

https://www.reddit.com/r/LegalAdviceUK/comments/1vxkcic/police_have_arrested_and_questioned_3_of_...
1•jackgavigan•10m ago•0 comments

SpaceX intends to invest up to $100B in Louisiana spaceport

https://arstechnica.com/space/2026/08/spacex-intends-to-invest-up-to-100-billion-in-massive-louis...
1•rbanffy•12m ago•0 comments

Meta agrees multi-billion settlement to claims that its platforms harm children

https://www.bbc.co.uk/news/live/cy5v0vrlp4xt
1•c-oreills•13m ago•0 comments

Can one hear the gauge of a string?

https://blog.veitheller.de/gauge.html
1•hellerve•13m ago•0 comments

Meta to Pay Up to $17.1B in Settlement over Social Media Addiction Claims

https://www.nytimes.com/2026/08/26/technology/meta-settlement-social-media-addiction-lawsuit.html
3•JumpCrisscross•15m ago•1 comments

We hand-finished 575,729 Urdu book pages, then trained a model on that work

https://ibteda.org/journey/
1•laamalif•16m ago•0 comments

A Man Who Saw Humanity from Two Billion Years Away

https://thereader.mitpress.mit.edu/the-man-who-saw-humanity-from-two-billion-years-away/
2•samizdis•16m ago•0 comments

Boston Scientific hit by cyberattack, global operations affected

https://www.reuters.com/legal/government/boston-scientific-hit-by-cyberattack-global-operations-a...
2•tartoran•17m ago•0 comments

The great American road trip is in peril

https://grist.org/culture/great-american-road-trip-national-park-climate-wildfire-drought/
1•speckx•17m ago•0 comments

The Wage Effects of Generative AI

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5842084
2•nilen•17m ago•0 comments

Show HN: Free economic nexus checker for online sellers (all 50 states)

https://nexussnapshot.com/
1•davyth•17m ago•0 comments

Petermann Glacier Ice Island

https://www.esa.int/ESA_Multimedia/Images/2026/08/Petermann_ice_tongue_April_August_2026
1•Sentralis•18m ago•0 comments

France reaches 94.9% fiber coverage in 2026

https://cartefibre.arcep.fr
2•nehalem501•18m ago•0 comments

One endpoint between your AI and all your connections, memory, skills

https://github.com/lanes-sh/link
4•s-xyz•19m ago•1 comments

Meta agrees to pay up to $16.7B to settle social media case

https://www.bbc.co.uk/news/articles/cd68q3wynnqo
3•didacusc•19m ago•0 comments

China's MiniMax sees revenue nearly quadruple in first half as AI demand surges

https://www.reuters.com/world/china/chinas-minimax-sees-revenue-nearly-quadruple-first-half-ai-de...
1•tartoran•20m ago•0 comments

Show HN: EBM Lens, searches biomedical papers, ranks evidence, grounds claims

https://github.com/mauroforlin/ebm-lens
1•mauroforlin•22m ago•0 comments