There are two flavors of Qwen 3.6:

- A 27B "dense" model

- A 35B "Mixture of Experts" model, which activates only 3B parameters for each token.

For your hardware, I strongly recommend `unsloth/Qwen3.6-35B-A3B-GGUF:Q4_K_M`. I have an M1 Max with 32GB VRAM from 2021 that can read at ~300-500 tokens/sec and write at ~30 tokens/sec with llama-cpp's default settings, which is plenty fast. The 27B model can read ~70tok/sec and write ~5tok/sec.

The 35B MoE model technically takes slightly more memory but is much faster because it's doing 1/9th the work. It's not quite as "smart", but it's comparable.

For coding tasks 27B is reported to be much more effective, altho you can probably only run 4b or 5b quants @ this memory.

Recommend https://www.reddit.com/r/LocalLLaMA/ as a great source for this type of discussion.

I played around with local LLMs on my M4 Max 64GB this weekend and this is exactly what I found. I put Opus 4.7 "head to head" on the same task as Qwen 3.6 and a few other local models. The 35B did not perform well IME - it needed a lot of handholding and even then the final result did not work until a few more tweaks, while Claude one shot the task. The 27B was much better and also one shot the task, but took about ~55min as opposed to about ~15min for Claude. The 27B is probably something that I could happily run for many use cases if I had some faster hardware... the main problem there seems to be that at larger context sizes, prompt decoding can take several minutes.

This matches my experience too. The little a3b model is quite capable for its size class, as is the 27B model, but it’s still an order of magnitude less effective than Claude on the “effectiveness / time” curve

Using omlx on the M1 max I get about 15tps from 27b

interesting! I might give omlx another chance, thank you

Thank you - I'll give that a go!

May I ask why the M instead of XL?

Obviously bigger != better but I don't know what the differences are.

These are dynamic quants, and they're basically just an indication of how far away from the desired quant it is allowed to go to achieve the goal. Generally, unsloth's toolchain moves quants up, rarely down.

* _0 and _1 do not use K quant and scales 32x32 blocks according to the original (B)F16 values; _0 scales the block using the original max and min values. _1 does this per row instead of per block.

* K quants do something similar, but now splits blocks into subblocks inside a superblock where the superblock has min/max scaling, but the subblocks also have scaling in the range of the superblock's scaling and are stored using less bits.

* K's M, L, XL are just how aggressively the subblocks and their scaling factors are chosen. Generally, it puts a max on how far you can deviate from the chosen quant to maintain the desired quality, but also gives them a bigger budget to perform that excursion in. XL most aggressively tries to preserve the intended quality, while S does the least.

* Dynamic quant on top of this scales entire layers, full of blocks, according to how much they effect various measurements (such as KLD and perplexity).

That said, there is no reason K_S is even produced by anyone, same with Q_0, Q_1, and I_NL. People should no longer be using those. M only is meaningful if you're trying to restrict the upper bounds: K_XL can reach BF16 for some weights, but rarely; people think this has a speed implication for hardware that has native 8bit in their tensor units (but it doesn't).

Unless you're specifically trying to cure a problem, stick with K_XL.

You seem to understand this stuff pretty well, any recommendations on resources (blogs, YouTube channels, whatever) for software engineers that want to keep up with this stuff on this kind of level?

A lot of the content about AI out there is kind of produced to the lowest common denominator. Basically a never ending scheme of get rich quick/passive income kinds of AI content.

Unsloth’s guides on getting various models running are great starting-off points for the “practicioner’s side” of things. Note that they include settings for llama-cpp, ollama, and other runtimes in addition to their own “unsloth studio” (their product seems like overkill imo)

If you’re curious about what a particular switch does, clone the llama-cpp repository to your computer and try asking your favorite pet rock prompts like “This is llama-cpp. Can you look at what the -ctk parameter does and explain to me?” Giving Claude/codex/whatever access to the actual code goes a long way, but it is just one opinion.

If you’d like to learn how transformer-based language modeling works in detail, I suggest starting with chapter 0 or 1 of https://arena-chapter0-fundamentals.streamlit.app/ depending on your skill level, then use that to work your way to reading research papers.

Graduate students who study these topics are generally as annoyed by the “get rich quick” style of advertising as you are, so the deeper you go toward academic research the quieter those voices tend to get, mercifully. That said, this is balanced by the unfortunate fact that top labs have strong posturing signals they try to send, so it can be hard to see which preprints actually have good ideas, which are trying to promote their group’s tech instead of doing science out of curiosity, and which have authors who’ve innocently deluded themselves into overfitting their own pet projects. Read widely but adversarially, test everything but hold fast to the good stuff, etc etc

Hey some of us are on hardware (gfx906 based Radeon MI50s with 32GB of stupidly fast VRAM and basically no compute) that inference significantly faster with Q_0 and Q_1 quants

Vega... unfortunately kinda sucks.

Its not amazing at compute (yet is a member of the GCN family, which I have been a fan of since its inception) and ended up being too expensive for perf/$ and perf/watt.

The only thing it did was make Nvidia rush Series 10 out the door and make it too good. Nvidia has been unable to live up to the gen-to-gen uplift Series 10 did, all because AMD made Nvidia blink.

Basically, you're 2 gens too early. CDNA2/gfx90a is the minimum you need to get any meaningful performance out of inference, or maybe CDNA1/gfx908 if you really don't need to quantize at all.

BTW, I did suggest this elsewhere in this HN story, but have you tried just disabling KV quant entirely? That is a huge speed uplift for compute-poor users.

Also, llama.cpp's support for gfx906 is probably never going to as good as it is for other cards, and good ROCm support for cards before they rebooted the driver/stack team is probably never going to materialize. I don't see the point in hanging onto them.

Like, if I was in your place, replacing it with even a 9060xt, with half the RAM, would be a step up. They go for $450. People have been building dedicated inference machines with these and they've been amazing, just throwing in 3 or 4 in, and scaling VRAM to meet needs.

I'd have to try the KV cache trick but folks get pretty competitive speeds with the current 31B/27B dense models e.g. https://www.reddit.com/r/LocalLLaMA/comments/1tc9j6u/mi50s_q...

If your hardware fits K_M but not K_XL, should you prefer going down to a lower quantization’s XL or sticking to the higher quant’s Q_M?

The correct answer should be "try it!"

But as models are starting to pack more information into less bits, some weights are just going to end up becoming super important and very sensitive to quant. So, I'd just move down a Q size, and continue with K_XL. Like, I'm betting Q3_K_XL will beat Q4_K_M on any given model in real world testing, even though its ~20% smaller, but perform worse on benchmaxxing.

The only exception I could think of is quantizing small models, like, my testing on Gemma E2B/E4B and Qwen 3.5 9B, quantizing at all was super noticeable... they can't spread the error across more weights.

Good news (at least for me), 24GB of VRAM is enough to store either of those in BF16 and then a ton of room for F16/F16 KV cache.

MTP recommended

on my M1 Max, MTP consistently lowers my performance! I’ve tried both llama-cpp’s recently landed MTP support (cloned and built Tuesday) as well as one of the other forks a few weeks ago. Suspect nobody’s done a comparison on hardware like mine.

I recommend sticking with the dense models for both Qwen and Gemma.

On testing I've done on same-quant apples to apples, with F16/F16 (ie, unquantized) kv cache, 35B-A3B underperforms against 27B on anything even remotely complex. But yes, 35B-A3B can be like 3-4x faster on my hardware.

By Qwen's own admission, on any meaningful benchmark (ie, ones that involve logic, math, or tool calling), 27B performs like 122B-10B and 397B-A17B, but 35B-A3B is somewhere between 27B dense and 9B dense.

Also, MTP recently got merged in, so I'd suggest downloading Qwen 3.6 MTP (I assume you get it from unsloth) and updating your copy of llama.cpp, and adding `--spec-type draft-mtp --spec-draft-n-max 2` to your arguments.

https://huggingface.co/unsloth/Qwen3.6-27B-MTP-GGUF/ https://huggingface.co/unsloth/Qwen3.6-35B-A3B-MTP-GGUF/

Also, I recommend not quantizing kv cache, and if you do, only quantize v. Lowering model quant while also lowering context size to fit F16/F16 or F16/Q8_0 massively improves model performance for thinking models. Also, quantizing cache, either k or v, decreases speed by a lot on some hardware.

I have a 24gb 7900xtx, so I can fit >32k F16/F16 context with Qwen3.6-27B, but use unsloth's Q3_K_XL. This performs better than Q(4,5,6)_K_XL with v quantized.

Edit: Oh, and since I mentioned Gemma 4, my testing mirrors my Qwen 3.5/3.6 experiences, 26B-A4B performs worse than 31B, but is also way faster. llama.cpp doesn't support Gemma 4's MTP style yet, so both could get even faster.

    I tried the qwen3.6-27b Q6_k GUFF in llama.cpp 
    and LM Studio on my M2 MacBook Pro 32GB machine 
    last week, and I barely get a token a second with either.

The fact that it was this slow makes me suspect it's a matter of insufficient free RAM. The entire model needs to fit into RAM (and stay there the entire time) for acceptable performance.

(not sure of exact diagnosis/fix, but definitely look in that direction if you're still having this issue when you give it another shot)

Also, there are two stages - prompt processing, and token generation. Prompt processing is notoriously slow on Apple Silicon unfortunately. If you have large context (which includes system prompts, lots of tools loaded by a harness like Claude Code, OpenCode, etc) it can take minutes for prompt processing before you see the first output token. On the bright side, the tokens are cached between turns, so subsequent turns won't be so bad.

You are using Q6 6 bit quantization; on my 32G MacMini I use Q4 and it is faster but when I use it with OpenCode, I set up a task and go outside to walk for ten minutes. Smart, capable, and slow. Still, I love using local models.

EDIT: I run with context wired at 64K

The 27B model is dense, so is relatively slow. The 35B-A3B model is marginally weaker but being MoE is much faster - like ~4-8x faster in basic benchmarks on my M1 Max.

For comparison, I just ran a couple of quick benchmarks (default settings) with llama-bench:

Qwen3.6-35B-A3B at Q6_K_XL gave 858 t/s pp512 (prompt processing) and 43 t/s tg128 (token generation).

Qwen3.6-27B at Q4_K_XL gave 103 t/s pp512 and 8 t/s tg128.

Have you tried enabling MTP? Those numbers are similar to what I was getting on my Strix Halo box, but configuring/enabling MTP doubled the TG speed of the 27B model (18-20 t/s now).

Thanks - I’m in the process. I’ve tried briefly, but so far it appears marginally slower. (Noting that llama-bench doesn’t support MTP yet so you’re reduced to running different prompts and eyeballing the log.)

So I’m assuming I’ve done something wrong along the way, but I’ve not had time yet to explore it.

Thanks for the info.

27B is the dense one. Try the Qwen3.6-35B-A3B variants for the MoE release. That's what I'm running on a Framework Desktop and I get ~50 tok/s plus or minus a few. The dense one is similarly slow for me -- not sure what to expect on your hardware from the MoE but it should probably be much faster.

Thanks!

Check out Unsloth Studio it provides MTP support now which 2x the token generation speed with no loss of accuracy: https://unsloth.ai/docs/models/qwen3.6#mtp-guide

I get 150t/s peak, 120t/s avg with Qwen3.6 27B Q4 with a 4090 on Linux. Now that MTP has landed into llama.cpp.

> qwen3.6-27b Q6_k

That's the dense model, you probably want a mixture-of-experts (MoE) one.

Here's what you probably want instead: https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF

Thanks!

My token throughput is much better using vLLM-mlx on my M2 ultra than llama.cpp. It might be worth a shot to give it a try.

you should be using dflash with that model, look it up