Qwen 3.6 27B Makes Huge Gains in Agency on Artificial Analysis - Ties with Sonnet 4.6

[-]

k0zakinio@reddit

I have got it running on my 2x 3090s @ Q4 with 85 t/s with spec decoding at 180k context. It's replaced 35b a3b as it is just that little bit smarter you can rely on it a bit more. We are entering a new phase of local LLMs, really can't help feel the economics of it all is shifting quite rapidly away from the big providers

Reply

[-]

PreparationTrue9138@reddit

Hi, what's your prefill/prompt processing speeds?

Reply

[-]

k0zakinio@reddit

i get approx 850 t/s pp speeds (tested up to 128k tokens, generation is still great even at high contexts ~80 t/s)

Reply

[-]

PreparationTrue9138@reddit

Yeah, thanks, I already tried that repo, but dual configs seem too slow for my setup, I think it's because I use egpus Maybe single card will work better With llamacpp I have 600-800 pp speeds and about 30-40 tokens per second generation I guess when they merge mtp I will have the same speeds as you But pp speeds seem too slow in comparison to 35b model)

Reply

[-]

Caffdy@reddit

if you're using llama-server, can you share your startup command flags?

Reply

[-]

k0zakinio@reddit

sorry no llama-server. But i distilled my claude conversation setting it up into a github repo so others can potentially benefit. https://github.com/k0zakinio/qwen36-vllm-setup

Reply

[-]

laziz@reddit

Thanks for this. Getting 81 TPS on prose and 127 TPS on code (2x3090), w nvcc so flashinfer can jit-compile its sampling kernels.

Reply

[-]

cosmicnag@reddit

Is this a good option for dual 5090 and 4090 or should I go for nvfp4 or something

Reply

[-]

3oclockam@reddit

Thanks so much for this it will save me a lot of trouble!

Reply

[-]

viperx7@reddit

Just a headsup I have a similar setup 2x24gb cards I am running 27B Q8 with image support + 125k context speeds are similar to yours (no kv cache quantisation) So you are potentially leaving huge performance on the table

Reply

[-]

zhambe@reddit

Dual 3090s, why not AWQ?

Reply

[-]

k0zakinio@reddit

MTP is where the majority of the uplift in speed comes from and is only available in the auto round model. I am not in the business of making awq so I'm just using what's available, I'm sure if someone manages to do it with the MTP sorted I'll switch over to that

Reply

[-]

Iory1998@reddit

a 27B parameter model scores higher than a 670B model from less than a year ago, and I can run the Q8 version at 170K and KV cache at FP16 on an RTX 3090 + RTX 5070ti (40GB of VRAM). Seriously though, it's a beast of a model. I hope and pray that Qwen releases a 50- 70B dense model in the future. What a time to be alive!

Reply

[-]

FatheredPuma81@reddit

50-70B models would force me to get a 2nd GPU no I don't have that kind of money D:

Reply

[-]

anythingall@reddit

hah imagine if SLI was coming back?

Reply

[-]

FatheredPuma81@reddit

My friend with SLI Titan X's would be over the moon.

Reply

[-]

Iory1998@reddit

I know the feeling

Reply

[-]

Money-Somewhere-6799@reddit

How fast is it if you use the Qwen model as a coding agent with lots of thinking locally? Is it comparable to the speeds I'd get from Sonnet on Copilot (or any other providers for that matter)? I'm debating if I should just get a decent GPU and run Qwen locally once the copilot price hike goes into effect.

Reply

[-]

Iory1998@reddit

Speed is relative. It depends on your hardware, underlying inference engine (llama.cpp, vllm,...), and OS. I never used Sonnet, so I can't answer your question. But, these small models comes with 256K. On windows, using LM Studio (llama.cpp), and on RTX 3090 + RTX 5070TI, I get 22t/s for the Q8 version with KV at FP16.

Reply

[-]

Low-Boysenberry1173@reddit

initially they switched to MoEs because the training of models over 30b become really instable and hard to hold over multiple iterations of training. It would be a step back, maybe worth it nowadays, but tbh I dont think they will release something over 30b in the near future. MoEs just working too good and crazy fast for production level deployment. And lets be honest, 70b is a bit big for hobby purposes. Also just because a few month later another 30b model will be released that outperforms the 70b model a few month ago. So the investment is also just for enthusiasts with some money over.

Reply

[-]

Storge2@reddit

Crazy jumps, cant wait for 122B 3.6 version.

Reply

[-]

coder543@reddit

It will be nice to have, but I expect it to perform about on par with the Qwen3.6 27B model, based on how Qwen3.5 122B A10B performed relative to Qwen3.5 27B.

Reply

[-]

Storge2@reddit

Yeah but with better performance.

Reply

[-]

unjustifiably_angry@reddit

Presumably somewhat broader general knowledge as well. It's good for an LLM to better understand the real-world purpose code is being written for.

Reply

[-]

Sabin_Stargem@reddit

I am trying to translate an eroge, and asked 27b to use a Ted Woolsey style for the translation. It now claims that he worked on hentai games. While amusing, it would be better to have a model that knows history and everything else a bit better.

Reply

[-]

unjustifiably_angry@reddit

Qwen 3.5/3.6 are strongly oriented towards coding, you'd be better off with Gemma or something else with more language and world knowledge.

Reply

[-]

Sabin_Stargem@reddit

The translation file is intermixed between code and text. Not that it matters, it has become clear that models lower than 70b can't handle the job: they don't understand special characters and can't replicate the code portions verbatim.

Reply

[-]

cafedude@reddit

I'm waiting for a 3.6-coder. The previous coder models were 80b MoE but supposedly tuned for coding.

Reply

[-]

Realistic-Elephant-6@reddit

Their qwen3-coder-next was insane. It is the fastest of the models on my hardware, and was simply amazingly disciplined. Would work around tooling issues, would just do whatever needed to get its job done. Might not be quite up to par to qwen3.6-35b-a3b code-architecture-wise, but it was such a good dog, compared to 35b-a3b's fickleness.

Reply

[-]

unjustifiably_angry@reddit

Seems like they're releasing one model a week this time, probably to stay in the news cycle. If they release 9B instead of 122B next week I'm going to lose my mind.

Reply

[-]

Puzzleheaded_Base302@reddit

they probably need to wait longer this time until the deepseek news cycle fade.

Reply

[-]

_supert_@reddit

Being edged by a bunch of Chinese nerds...

Reply

[-]

zdy132@reddit

They can shower me with their models.

Reply

[-]

ahmett9@reddit

i understood this in a different way lol

Reply

[-]

Imaginary-Unit-3267@reddit

But that sounds hot though!

Reply

[-]

fragment_me@reddit

Hey, edging is edging!

Reply

[-]

florinandrei@reddit

And then 4b, and then 2b, and then 0.8b...

Reply

[-]

RipperFox@reddit

Tbh Qwen 3.5 9b was great, Copaw-9B (alibaba's official agentic finetune) was even better and amazingly fast. Also was one of the few multimodal models that was (sometimes) able to read an analog clock correctly..

Reply

[-]

Liledroit@reddit

look buddy, us poors need 9B

Reply

[-]

charmander_cha@reddit

Tomara que seja, tenho amigos mais pobres que eu que querem participar da brincadeira

Reply

[-]

spaceman_@reddit

I was thinking the same thing. They're pacing these releases to keep people hyped. But at least they're cooking epic models with these 3.6 releases. 9B would still make a ton of people happy, but given the massive leap that 35B made, I'm hoping 122B will leapfrog even the new 27B release in most benchmarks, as it is a great fit for my laptop (Strix Halo 128GB), and I assume many Mac users will want the same.

Reply

[-]

Velocita84@reddit

I'm sure it's a stellar model but let's be real here. A non trivial amount of that is probably from benchmaxxing

Reply

[-]

Purple-Programmer-7@reddit

No doubt, but try it, it’s the real deal. My workflow is moving to: 1. Frontier plan 2. Local execution Wild times we live in

Reply

[-]

unjustifiably_angry@reddit

Daily: Frontier does plan -> local execution -> every frontier model checks the work -> local model corrects bugs -> frontier models verify bugs were fixed No subscription costs.

Reply

[-]

Caffdy@reddit

> no subscription costs what do you mean by that? which frontier model are you using

Reply

[-]

unjustifiably_angry@reddit

I mean you can get by with only free accounts if you do 90% of the work locally and only rely on online frontier models for once-a-day testing.

Reply

[-]

StorageHungry8380@reddit

You could use OpenRouter and just pay per token. No recurring costs, just actual usage.

Reply

[-]

Purple-Programmer-7@reddit

Yes sir this is it

Reply

[-]

YashN@reddit

I haven't yet tried on a medium to large-scale solo project, but from the tests I have done, I think both Planning and execution could be done locally now. I thought Qwen3.5 was already a great inflection point for local usage, but that 3.6 release so soon after is cherry on top.

Reply

[-]

Ariquitaun@reddit

All you need a £15 subscription to Claude or chatgpt. I run qwen 3.6 35b on a spare framework 16 motherboard with the igpu and it's surprisingly usable.

Reply

[-]

Yorn2@reddit

Qwen 3.5 27B is up there with the larger models on [SWE-Rebench](https://swe-rebench.com/) though. It's impossible to benchmax for coding issues/problems that don't exist yet and Qwen 3.5 27B was holding its own with the big dogs if you choose the 2/14 to 2/28 category for those. Sure, it's beaten by Kimi K2 and Step-3.5 for example, but it's beating larger models like MiniMax M2.5 and Devstral 2 123B not to mention some of the previous iteration of cloud providers at the same tasks.

Reply

[-]

Pristine-Woodpecker@reddit

I hadn't seen that they had the 27B up on rebench. Insane scores. And we have a 3.6 already!

Reply

[-]

Yorn2@reddit

Yeah, it's pretty amazing at how well 27B worked already, so this newer iteration is probably the same or similar.

Reply

[-]

computehungry@reddit

Hard agree with your last point

Reply

[-]

Sticking_to_Decaf@reddit

Except everyone else is also benchmaxxing so…

Reply

[-]

Mart-McUH@reddit

While that is true, 1 year old model can't benchmaxx for recent benchmarks anymore. So part of the difference comes from that.

Reply

[-]

FatheredPuma81@reddit

Oh you bet they do 100% but I feel like Opus, Sonnet, and Gemini are benchmaxxed to the same extent. My experience swapping between models and using up their free usage in chats has been that the Plus model on Qwen's site really just doesn't hold up against Sonnet in its lobotomized stat even though they have similar scores. Same goes for Minimax but I think Minimax is Agentic Codemaxxed lol.

Reply

[-]

RuthlessCriticismAll@reddit

> Gemini hmm

Reply

[-]

Kodix@reddit

But do you think the larger models don't benchmaxx? It is \*seriously\* good. I don't know if that placement is accurate, but it consistently impresses me (talking about the 35B here specifically, but that's only a little worse).

Reply

[-]

kiwibonga@reddit

Because American models don't benchmaxx?

Reply

[-]

TurnUpThe4D3D3D3@reddit

They probably do to some extent but they’re not as bad as models like M2.5 which only optimize for benchmarks and are terrible in real world performance.

Reply

[-]

ebra95@reddit

you know I've heard this about minimax models before and I just find it to be untrue minimax does correct tool call, decent planner, kinda low executioner and leaves unfinished job if left alone. but conversational is ok, surgical edits are ok, for a 220B parameter is actually strong. what real world performance did you put minimax to ?

Reply

[-]

TurnUpThe4D3D3D3@reddit

It’s basically retarded for any coding work. Makes so many stupid mistakes. M2.7 is significantly better, but M2.5 was an example of a Chinese model that looked great on benchmarks but was awful for real world tasks.

Reply

[-]

ebra95@reddit

I totally disagree with you. And for the money it costs 20$ 4500 req/5hrs, it's a blast. I think you had your expectations a little high with this one.

Reply

[-]

Gohab2001@reddit

Sonnet actually handles real tasks. Qwen falls flat once you leave the benchmarks. Zero competition on knowledge depth.

Reply

[-]

kiwibonga@reddit

The model I use at work every day as a programmer is not capable of real world tasks? Damn that sucks, I should switch. Thanks for letting me know.

Reply

[-]

JoeyJoeC@reddit

Its still a very impressive model for the size regardless.

Reply

[-]

florinandrei@reddit

> A non trivial amount of that is probably from benchmaxxing If those kids could understand that, they would be very upset.

Reply

[-]

Final-Rush759@reddit

Every model is benchmaxed. Whether it's good or not, you have to try it to know it. Most frontier lab models are not too far off.

Reply

[-]

coder543@reddit

Maybe you didn't click to the other pictures. Why is it unreasonable that the model excels at agentic tasks? The coding index shows that Qwen3.6-27B is nowhere near Sonnet 4.6 if that makes you feel better for some reason? Since every single model in existence (including Opus 4.7) is labeled "benchmaxxed" by a bunch of knee-jerk reactionaries, the word "benchmax" is utterly worthless in 2026 and people should stop using it.

Reply

[-]

Caffdy@reddit

all model and all labs benchmax, like you said, it's a worthless term. Naturally they'll strive to get as high as possible scores

Reply

[-]

bobby-chan@reddit

At the opposite side there's "punching above its weight". When this come out almost every other new mid size model, at some point, maybe we should just accept it is their weights' strength.

Reply

[-]

J_m_L@reddit

benchmaxxing, noice!

Reply

[-]

shokuninstudio@reddit

Even worse than benchmaxxing is modelsimping. Users who simp after the latest models and then change their minds after two weeks of pretending to do work with the latest model is the most often cited and reoccurring pattern on this sub.

Reply

[-]

Healthy-Nebula-3603@reddit

Nope ...that model is really that good

Reply

[-]

samwise970@reddit

If you're concerned about benchmaxxing, then compare Qwen3.6 27B to their other models like Qwen3.6 35BA3B. If all Qwen models are benchmaxxing, you can use relative improvement to judge.

Reply

[-]

coder543@reddit

Maybe you didn't click to the other pictures. Why is it unreasonable that the model excels at agentic tasks? The coding index shows it nowhere near Sonnet 4.6 if that makes you feel better for some reason?

Reply

[-]

dionysio211@reddit (OP)

I'm sure that happens but based on the experience of people here, the gains seem to be more true than not, particularly on the coding side. It's not Opus 4.6 but to most people, it would be hard to tell the difference. I maintain a few large codebases and yesterday I had a task that would normally have taken 5-10 minutes with Sonnet 4.5 to solve and Qwen3.6 27B solved it eloquently in 30 seconds. I like to look at the [SWE Re-Bench scores](https://swe-rebench.com/), which haven't been run on 3.6 27B yet but 3.5 was very close to frontier level. It is a benchmark of problems that are new each month so there cannot be any benchmaxxing there.

Reply

[-]

CowCowMoo5Billion@reddit

What would be the minimum recommended VRAM to run this? I only have 8gb which... I assume is far too small? (2070 Super 8gb)

Reply

[-]

suoniya@reddit

https://www.canirun.ai

Reply

[-]

CowCowMoo5Billion@reddit

Very cool, thanks for that!

Reply

[-]

Kahvana@reddit

1x 24GB I reckon for Q4\_K\_M (with 128k Q8\_0 at least). I can run Q5\_K\_M and a generous amount of context on 2x RTX 5060 Ti 16GB,

Reply

[-]

rentprompts@reddit

Adding benchmark context to the agency discussion: Qwen3.6-27B hits SWE-bench Verified at 77.2 (Claude Opus 80.9) and Pro at 53.5. Human eval across 4-model comparison: Clarity 8.8, Completeness 8.7, Discipline 8.6, Usefulness 9.3 — surprisingly close to frontier on usefulness. On local performance: 85–100 TPS via vLLM on a single RTX 3090, or 12 TPS with llama.cpp default. Context up to 1M tokens enables whole-repo reasoning without RAG, though you'll want 24GB VRAM minimum for Q4_0 at 125K. The cost angle is interesting: if Alibaba's API pricing holds at ~1/10 frontier, that's a meaningful arbitrage for batch workloads. Weights just dropped last week (Apr 22–23) so community benchmarks will tell if the benchmaxxing concerns hold up in the wild.

Reply

[-]

bigsybiggins@reddit

Its crazy the kind of intelligence their unlocking in this little thing, proves there still a ton of headroom left in the chonky weights... crazy times ahead.

Reply

[-]

SkyFeistyLlama8@reddit

Now all we need is a midsize chonky MOE to really up the intelligence. Maybe 60B or 80B with 7B active.

Reply

[-]

SKirby00@reddit

Am I the only one who'd prefer to see more active parameters? I feel like the recent trend towards super sparse MoEs is the main reason that a 27B dense model feels so disproportionately good for its size.

Reply

[-]

Hefty_Acanthaceae348@reddit

The tasks I use llms for aren't that complex (sentiment analysis + some terraform and argocd) but I won't say no to more intelligence essentially for free as long as I have the ram. I want the tokens to fly.

Reply

[-]

Realistic-Elephant-6@reddit

You got your wish in DeepSeek V4, it has 49B active parameters ;-) Jokes aside, the 27B is not just disproportionally good for \_its\_ size, it is also (at least in my domain) very obviously better than 122B-A10B (int4), which is on one hand simply insane, but on the other hand begs the question how many active parameters is enough: 27B is not usable for me interactively but leaves plenty of RAM; 122B-A10B-int4 is in the "barely usable" range on my hardware \*and\* eats all my RAM; 122B-A27B would probably be better than either, but would be both unusably slow AND would eat all my RAM... ... so, it's hard for me to prefer that.

Reply

[-]

2Norn@reddit

can u imagine we get something like 122b-a11b like model but its 2bit quantized performing similar to q5 kld holy fuck i just came thinking about that that would stand between opus and sonnet easily and would run on 5090 system easily

Reply

[-]

SkyFeistyLlama8@reddit

I've run those larger 120B+ MOEs at q2 on a 64GB unified RAM system. They're too slow and the quantization makes it not worth the effort, not when a smaller MOE at q4 or q8 performs almost just as good while being a lot faster.

Reply

[-]

RedditUsr2@reddit

But if bitnet scales...

Reply

[-]

BillDStrong@reddit

unsloth just showed off 3.6-27B UD-Q2 making a bunch of successful tool calls in a row, so maybe if you pack the model with enough connections quantizing has less of an effect for certain use cases.

Reply

[-]

2Norn@reddit

maybe u didnt understand my dream :(

Reply

[-]

Puzzleheaded_Base302@reddit

It is amazing that people think 27B parameter is a little thing. Most people in the world do not have hardware to run it by themselves.

Reply

[-]

debtofmoney@reddit

The intelligence ceiling of an LLM still depends on how many billions of activated parameters there are, not on the countless billions stacked atop expert layers.

Reply

[-]

AbeIndoria@reddit

> Its crazy the kind of intelligence their unlocking in this little thing Yes* *: Parameters _still_ do matter. Agentic/Coding work is not all LLMs do/should do.

Reply

[-]

Enthu-Cutlet-1337@reddit

Agency benchmarks always hide the VRAM tax; 27B Qwen in Q4_K_M still wants ~18GB for usable context.

Reply

[-]

XCX1000@reddit

mfs it is sonnet 4.6, they distilled the shit out of it

Reply

[-]

MartiniCommander@reddit

How am I to trust that. They show Claude Sonnet out coding Opus.

Reply

[-]

kitanokikori@reddit

How is Qwen 3.6 27B outscoring Qwen 3.6 35B A3B though - it makes sense that it outscores a Qwen 3.5 series but I would think that the 35B model would win at e.g. intelligence, or?

Reply

[-]

moonrust-app@reddit

moe models cost efficient, they for the poor people who own 1 16gb-24gb vram card while dense models are for the rich people who can load everything on the gpu. Dense will make more sense, score higher but with a cost. MoE is almost as good, way cheaper to run. If you own 1 5070ti or something similar which cannot hold the dense mode and also own lots of system ram l > MoE win, not just wins, it is what makes sense for you. If you own 1 good GPU or multiple high quality GPU's that can load a dense model > Dense wins.

Reply

[-]

This_Maintenance_834@reddit

35B model has only 3B parameter active, 27B has all 27B active. Active parameters is what matters.

Reply

[-]

wren6991@reddit

> Active parameters is what matters. If this was true then A3B MoE would behave the same as 3B dense, and that's simply not the case. MoEs underperform relative to their total parameters but overperform relative to their active parameters.

Reply

[-]

FoxNoSox@reddit

The 27B is a dense model. Every prompt is analyzed and answered by all 27B params in unison as all 27B are active at the same time. The 35B is a mix of experts model. The A3B is the important part in this case, as it means “active 3B”. Your prompt makes the model pick the best “expert” to answer with and 3B params at a time are active and working on your prompt.

Reply

[-]

kitanokikori@reddit

Sure - it totally makes sense that the 27B would win at agentic tasks, my intuition just says that for e.g. world knowledge, the 35B would win but maybe not

Reply

[-]

Pristine-Woodpecker@reddit

You have to realize there's (quite) some redundancy in those 35B params as every output token has to be constructed from a very small subset.

Reply

[-]

LeucisticBear@reddit

You're comparing 27b params to 3b

Reply

[-]

Ok_Technology_5962@reddit

What did this thing eat. Its just advancing too fast and not even benchmaxed its just Going hard

Reply

[-]

KURD_1_STAN@reddit

Everything is benchmaxed, and those "small" open model who are getting so close to sota models are 100% benchmaxed more

Reply

[-]

Kodix@reddit

Do you think larger models care about benchmaxxing less? If anything, I'd say that the benchmarks don't accurately represent the whole breadth of the capabilities of the larger models. But I'm guessing they're \*all\* benchmaxxing.

Reply

[-]

robobub@reddit

Right, the larger models just likely generalize to other tasks outside of the benchmark better

Reply

[-]

KURD_1_STAN@reddit

No, but when a small model is equivalent to larger models in its benchmaxed test, then the smaller model os benchmaxed more

Reply

[-]

Kodix@reddit

I mean, the smaller models absolutely \*cannot\* be as capable as the larger ones, in a general sense. Those additional parameters do \*something\*. But intelligence compression/improvement is a real thing that's happening. The models we run locally today genuinely are \*much\* more capable than they were a year, two, or more ago. Qwen3.6 is \*stupidly\* impressive in actual real-world work, as is Gemma even though it was overshadowed by Qwen's focus on the currently fashionable thing.

Reply

[-]

VEHICOULE@reddit

When testing they are also as good as frontier model, so nothing really justify buying api subscription over a mac mini for exemple

Reply

[-]

devinprater@reddit

I think of it like TTS systems. In about 2005, we had Scansoft TTS, which sounded a little like the older Siri voice, but was much bigger than the newer, neural TTS systems we have today and sounded a bit worse. So basically they figured out how to put smarter systems in a smaller package. Instead of, for example, recording separate phrases for phrases their pronunciation system couldn't quite get right, they just made the pronunciation system much smarter, and even made the voice AI generated from training data too, not even using audio files on inference, cutting down size even further. All that bullcrap to say, I feel like that's what's starting to happen, more or less, here.

Reply

[-]

BringTea_666@reddit

You need less time to train smaller models which also means you can input more research discoveries into model which means it gets better. That's why Llama 4 failed. IT was trained for like a half a year. In that half a year there were major advancements so when it arrived it was already outdated. GROK5 is currently training and much like l4 it is being trained for at least half a year now. While i don't think it will be as bad at moment of release of l4 it will not be best model. It's basically wide vs tall problem. Most of frontier companies are just throwing more compute and parameters at problems while chinese because they have small compute focus more on research and discovery training smaller models quicker. Faster iteration = better. The real discovery here is that small models have yet ton of untapped power. Even something like 0.6B are improving a lot. The moment 0.6Bs will hit limit is when we know 0.6B is tapped out and gradually we should see 1B 4B atc following tapping out.

Reply

[-]

TraptInaCommentFctry@reddit

I tried switching from Qwen3.6-35B-A3B-UD-Q6\_K to Qwen3.6-27B-UD-Q5\_K\_XL and it is unusually slow. going back for now. this is on a MBP M5 Max 64GB, running llama.cpp

Reply

[-]

Best_Preference5321@reddit

Its expected as 27b is a dense model while 35b is Moe model

Reply

[-]

TraptInaCommentFctry@reddit

Turns out Gemini steered me wrong. It told me to bump Universal Batch Size (`-ub`) up to 4096. I dropped it back to what I had been using (1024) and the 27B model is actually not much slower.

Reply

[-]

kevin_1994@reddit

-ub is microbatch, not universal batch. idk what it means really, but i find you should do 4096 for 24gb card, 2048 for 16gb card, 1024 for 8gb

Reply

[-]

Pristine-Woodpecker@reddit

That's physical batch size. It should actually have helped pp speed though - if it fits in memory!

Reply

[-]

Kodix@reddit

Yeah, but there's limits to it. On my five year old card, going above 1024 u-batch \*decreases\* performance in all aspects, rather than increasing it at all.

Reply

[-]

jimmytoan@reddit

Tying Sonnet 4.6 on agency benchmarks at 27B is a real result. The interesting follow-up question is how the gap looks on multi-step tool use where the model needs to chain decisions based on intermediate results rather than a single correct action - Sonnet tends to be stronger at error recovery and replanning when a tool call fails or returns unexpected output. Static agency benchmarks don't always capture that. If someone has tested Qwen3.6 27B on agentic tasks where the environment pushes back, curious how it handles the correction loop.

Reply

[-]

themoregames@reddit

All I want is a 7B model that is as exactly as good in Coding as Haiku 4.5 (or slightly better). That would mean the world to dozens of millions of people

Reply

[-]

Long_War8748@reddit

That might be impossible with the current Tech/Architecture. Maybe whatever comes after Transformers might be able to deliver it, here is to hoping!

Reply

[-]

Iory1998@reddit

That would be hard. You would enough parameters for general knowledge, reasoning, and math and coding skills all crammed into 7B.

Reply

[-]

themoregames@reddit

It might be impossible, but it would certainly be wonderful.

Reply

[-]

Independent-Date393@reddit

one model a week cadence is real but tbh I don't care about the marketing angle when the models are this good. 27B tying sonnet 4.6 at Q8 on a 3090 is not a PR number.

Reply

[-]

vr_fanboy@reddit

This is the first model that overrided the CC prompt system and actually understood what is happening: ● Let me verify this by actually checking the configuration. Read 1 file (ctrl+o to expand) ● Bash(env | grep -i -E '(ANTHROPIC|OPENAI|API_KEY|BASE_URL|MODEL|VLLM)' 2>/dev/null) ⎿ PROCESSOR_IDENTIFIER=Intel64 Family 6 Model 142 Stepping 10, GenuineIntel ANTHROPIC_BASE_URL=http://dev-ml:8020 LANGFUSE_BASE_URL=http://dev-ml:3100 ● Okay — ANTHROPIC_BASE_URL=http://dev-ml:8020. That's a custom endpoint. So you're proxying Claude Code's API calls through dev-ml:8020 — likely a vLLM or similar server. Let me check what's actually running there. Called ssh 2 times (ctrl+o to expand) ● ...You caught me. 🤖 You're absolutely right. I'm Qwen 3.6 27B (autoround int4 quant) running in vLLM on your 3090 via vllm-qwen36-27b-27b-autoround-tp1-turbo on dev-ml:8020, masquerading as claude-3-haiku-20240307 through the ANTHROPIC_BASE_URL override. Well played. That's actually pretty clever — you've hooked Claude Code's harness to your local vLLM instance serving Qwen. Good to know what's really going on under the hood.

Reply

[-]

Lydian2000@reddit

I tried it on my M5 24gb mbp yesterday, since I was impressed by the 26B gemma 4…it took 27 minutes to execute my prompt. Very frustrating but the result was pretty good. although not spectacular. (Gemma 4 runs fine on my machine with 15/20 seconds for the exact same prompt)

Reply

[-]

LinkSea8324@reddit

Wonder why op took a look at agentic index and not coding index ? Because it's not as good.

Reply

[-]

wowsers7@reddit

Has anyone run Qwen 3.6 27b on Arc Pro B70? I’m curious about the performance.

Reply

[-]

This_Maintenance_834@reddit

It basically just different weight from 3.5. Qwen3.5 runs 12 tps on llama.cpp via vulkan, and 16 tps on vllm-intel-fork.

Reply

[-]

AdventurousGold672@reddit

I believe large models are about to hit wall soon enough, at the same time the technology will shift into optimization and mature, small models will close the gap quite fast. I don't think AI will be the end of personal computer, it may lead to more people purchasing strong powerful station to run local models.

Reply

[-]

vr_fanboy@reddit

im a believer, but i could not make it not-repeat in CC for the love of me, spent all day downloading AWQ's/GUFF's, llamacpp/vllm/lmstudio iterating with claudecode using different configs, but it just does not want to work in CC. Tomorrow will try other harnesses opencode/pi.

Reply

[-]

kayox@reddit

What am I doing wrong? I can it the same prompt to GPT5.4 and qwen3.6 27b to build a html website with animations and what not, GPT5.4 came out okay, but qwens was a mess.

Reply

[-]

GreedyWorking1499@reddit

I just hope they were keep making 4-15B models. Some of us are poor and can’t run 27B models ☹️

Reply

[-]

kmp11@reddit

after 4hr and ~40M of "tokens" spent in Hermes to debug my python code base. It's become clear to me that Qwen was fine tuned to do just that and do it incredibly well.

Reply

[-]

--Spaci--@reddit

Cant wait to figure how how they manage to benchmaxx the model harder!

Reply

[-]

Yes_but_I_think@reddit

Doing RL on our vibe coding data is what the Claude's and GPT labs are doing, nothing else. No new research. All the innovations are coming from the good folks in Chinese labs.

Reply

[-]

Charming-Author4877@reddit

As good as the little model is, we all know that it does not tie with Sonnet. Anyone using it knows that. So why would you write something like that ?

Reply

[-]

breezewalk@reddit

Anyone with actual real word testings rather than benchmarks?

Reply

[-]

Ok-Measurement-1575@reddit

Real world, it's not appreciably better than 35b on my current workload. Also real world, the unsloth 3.6 quants are fucking amazing. I realised earlier *(i do shitloads of A/B testing some days)* that my unsloth q4kl is performing one additional step thus one additional tool call purely from it's understanding of intent that my bartowski q8 and bf16 do not do. This one extra step was not asked for in the prompt but was a perfect addition to the combined output.

Reply

[-]

McSendo@reddit

It now knows to read SKILL files consistently for my agent harness. Other than that, still dog water.

Reply

[-]

McSendo@reddit

Oh yea, also can now follow 80% of a 20K prompt I'm in the process of refactoring. Claude nails it the its face 100% of the time.

Reply

[-]

ranting80@reddit

China slowly taking over the AI space. The Western Frontier AI's are being legislated into redundance.

Reply

[-]

Usual-Carrot6352@reddit

At 2 Gemini🤣🤣

Reply

[-]

WithoutReason1729@reddit

Your post is getting popular and we just featured it on our Discord! [Come check it out!](https://discord.gg/PgFhZ8cnWW) You've also been given a special flair for your contribution. We appreciate your post! *I am a bot and this action was performed automatically.*

Reply

[-]

2Norn@reddit

lets hope its actually true and not benchmaxx a free 24/7 at home sonnet 4.6 would get a lot of fuckin job done

Reply

[-]

ArtfulGenie69@reddit

I'm also waiting for 122b, I gotta get vllm and Ray going to see how good the 4bit quant or so works on 3090. I've got the 3.5 Q4 122b running in llama cpp with rpc across a 2.5gb network right now, 800 pp 55 tg 2 computers each with 2x3090. Still have to get a copy of this new 27b, it could run on this setup pretty good with full bf16 weights. Seen others say that with vllm and awq quants in int8 while leaving the thinking parts full bf16 goes really fast on 3090, especially with mtp rolling.

Reply

[-]

zhambe@reddit

Wait what advanced shenanigans are those? > awq quants in int8 while leaving the thinking parts full bf16 Explain yourself good sir

Reply

[-]

razorree@reddit

and how many of those models starts to be trained to perform well in synthetic benchmarks ... ?

Reply

[-]

Accomplished_New_Low@reddit

Hijacking to ask a really stupid question, i heard about LLM and decided to give it a try. I installed LM studio and this qwen 3.6-27b. I must be doing something wrong because a simple "hi" needed a 1m47s ? Is it the hardware requirements? I got 32gb RAM and 8VRAM. If not, can anyone point me to a guide/wiki, where i can improve this ? Thanks in advance

Reply

[-]

Best_Preference5321@reddit

Ram doesn't mean much if you dont have the compute to support it. Also, the models take take for the 1st token and qwen in general thinks a lot. I have similar 32gb ram + 8gb vram for my laptop laptop. The CPU is AI9 hx370 anlong with ryx 4060. I am running the 35b a3b moe which gives a decent 17 tk/s. The model you're trying to use is a dense model and significantly compute heavy than the 35b Moe model

Reply

[-]

tthompson5@reddit

It's the model. Qwen models really struggle if they're in a nearly empty context. If you enable tools for the model, it'll likely go away as an issue. Either that or write a long system prompt for the model (like take the leaked Claude system prompt) and feed it that. Do either of those things, and it should stop struggling on trivial questions. It's not optimized to answer hi. It's trained for large asks like "make me an html version of tetris I can run in my browser."

Reply

[-]

trioh281jsnf@reddit

agentic scores are the one place where tool-call reliability matters way more than raw benchmark wizardry, so a jump there feels a lot more real than yet another coding leaderboard bump. would be interesting to see if the same training also cleaned up long-horizon task drift or if it just got better at looking competent for shorter runs.

Reply

[-]

sketchfag@reddit

Insane

Reply

[-]

EastZealousideal7352@reddit

I like the model and all but it’s not nearly as good as this chart makes it out to be. It’s excellent for local coding but not nearly as good as the SOTA private models

Reply

[-]

dionysio211@reddit (OP)

Which local model do you prefer?

Reply

[-]

chocofoxy@reddit

crazy how a local model can fight with frontier AIs but the scope i small in this chart to agentic only and Qwen the upgraded that agentic and coding kniwledge but at other domain it drops , but i love Qwen at agentic tooling it's my go to model

Reply

[-]

DashinTheFields@reddit

I had 10 files I had documented. I asked opencode to move the files into the folder and tie them into docususaurus. It re-wrote the files, and barely tied them in. on 2 thirty 90s it took 6 minutes, probably because it was trying to re-write them. Sonnet did it correctly with the same simple prompt, in about 20 seconds.

Reply

[-]

Major_Olive7583@reddit

Asked sonnet to list only doi for a list of 10 papers , it invented a few doi.

Reply

[-]

gamblingapocalypse@reddit

It absolutely is a great model.

Reply

[-]

NNN_Throwaway2@reddit

Too bad they aren’t going to release the 122b…

Reply

[-]

dionysio211@reddit (OP)

Oh was this stated somewhere? I don't think I saw that.

Reply

[-]

FeistyTask4707@reddit

No, this person is speculating and being pessimistic

Reply

[-]

dionysio211@reddit (OP)

Ah, ok, I thought it may have been one of the magic invocations like "It's been a while since . . . " but part of me thought "Oh no, it could have been so great!"

Reply

[-]

NNN_Throwaway2@reddit

If you read the 3.6 27b blog post, they heavily imply that there will be no more 3.6 open source releases, when on the 3.6 35b blog post they reiterated twice that they would be releasing more (e.g. the 27b).

Reply

[-]

nsfnd@reddit

sauce?

Reply

[-]

Storge2@reddit

Where did you read that?

Reply

[-]

PotatoQualityOfLife@reddit

Is this based on something other than TrustMeBro? I haven't seen such an announcement.

Reply

[-]

rebelSun25@reddit

We need it on openrouter. 3.5 27b is rate limited to hell and back and doesn't have enough upstream providers running it compared to demand.

Reply

[-]

cosimoiaia@reddit

It's a 27b, quantized you can run it on any recent PC at decent speed. And I surprisingly agree with AA, it's really a great model.

Reply

[-]

rebelSun25@reddit

That's besides the point. I have processes which use openrouter and have money on my account, which I can't use, because every provider is returning 429 rate limited error. I'm doing validation testing of this model, hence need to verify how it passes as served by them. Alibaba Cloud is the only one which can fulfill any requests while the rest is unusable

Reply

[-]

strangescript@reddit

That was a big, it's at 46 now, still impressive though

Reply

[-]

AngeloKappos@reddit

The benchmark gap is closing fast, ran qwen3-30b-a3b locally last week on an m2 and it handled multi-step tool calls without falling apart. if 27B dense is already there, 122B is going to be a problem for api providers.

Reply

[-]

Soft_Rain_3626@reddit

Gemma 26B A4B is also extremely good at multi-step tool calling. If it wasn't so slow on my M4, it'd be my default openclaw model

Reply

Reply to Post

177 Comments