2 things, 1st: Alibaba's official endpoint pricing. they don't want to undercut too much as there is profit to be made to be close to it but not too low

2nd, and maybe more importantly: KV is not as efficient (vram usage-wise) as something like deepseek v4 flash. for 256k, fp8 kv is 9.3gb (full precision ~17.3gb). deepseek v4 flash is ~2.5b for the same size at full precision (which is fp4/8, if you are interested in it, read the paper, its pretty cool).

Doing the math, hosting 27B at NVFP4 (~23gb) with 2.3M total ctx (9 agents) matches the vram usage of ds v4 flash for the same 2.3M ctx (2.3 agents). the break point is 1.5M (6 27B agents) if you use full precision 27B.

To be clear, the qwen3.5 architecture (what 3.8 uses) is still considered decent in terms of KV efficiency, its just that dsv4f's architecture is SOTA in that space, and with the lower active params, you get better max kv scaling and higher speed serving that, if you have a lot of gpus.

Thanks for the detailed response. I saw some details later in the thread that I had somehow overlooked that touched on this....but anti-procrastination settings prevented me from changing my question.

yeah, I was confused during the whole thing, i get 70 t/s on a 3090, which evens out around 50 t/s at 128k+ , have been running 3.6 and now 3.8 (both iq4_nl at 256k q4 kv) on the 3090 for months. I am confused as to what we 'discovered' here, it's a common config. and at less than 1/2 the price of the gpu (and double the bandwidth, though no fp4 cores to be fair).

A 3090 has 936 GB/s of bandwidth vs 432 GB/s on the 70W RTX PRO 4000 SFF, so 70 t/s there is not surprising. The interesting constraint here was fitting a workload-tuned 5.01 BPW quant + 256K + MTP into 24 GB while working with less than half the memory bandwidth

loading the model would be similar vram footprint, correct, however the size of KV is based on 'active' params, not total params. So while at 1k ctx both will be in the same ballpark vram footprint-wise, at 100k the story will be very different. 27B at q4 kv for 256k takes ~8gb, while 35B at q4 kv around ~3.5gb, so at full precision kv those would be ~32gb and ~14gb (all ballparks, if you want exact numbers, its not hard to test).

As for the "cost", here i think the interesting arguments are around speed vs accuracy/"getting the job done", not literal $ cost per token.

I am asking mostly for running on a 3090.

I think the tps difference between them (both fitting in vram) won't be more than 2x in practice.

I would happily take 20tps over 40tps, if the model gets 3x more correct answers.

Speed is (for the most part) active-parameter based, so a 30B-A3B model is roughly 10x the speed of a dense 30B (realistically closer to 8x) in the case when both fit. That's the proposition of MoE and why everyone is trying to make massive models with very few active params, so that they are still fast while having access to a lot of knowledge (at the cost of reasoning, as reasoning ability 'for the most part' comes from active params).

You can test this by running this nemo or 35B on the 3090. I have and its very fun (but sadly a worse model than 27B, so I usually keep 27B on my 3090)

I remember running both qwen 30b-a3b and 27b on my 3090, and on the initial test, the 27b was only like 2x slower.

Ran a quick test so that we both have accurate numbers, without MTP* at 10k ctx 27B hovers around 42 ts in llama.cpp, 35B around 135 ts. So not the 8x I assumed, just over 3x, but thats still a big difference.

For the sake of testing I turned MTP off, as that heavily depends on what the generated text is (structured text like code is very often a lot more predictable, therefore bigger boosts) and the quality of the quantization, as drafters learn how to "mimic" the full precision generated tokens, so when you layer the fact that MTP is a 'guesser' of the main model's next token, and quantization affecting what exact token is generated, it'd make comparisons like this needlessly noisy.

Thanks for sharing.

Yeah, that was my experience too (2x or 3x is indeed considerably faster), but not workflow-changing faster at 45tps baseline, especially for asynchronous tasks (which is my goal with a local 3090, to just let it do things non-stop, without my intervention).

What was the result with MTP?

Isn't MTP "losless"? The result should still be relevant when averaged across a fee queriers across different domaine I guess.

MTP is lossless in the sense that running a model with and without (at temp=0, meaning no randomness) will produce identical results. It's true that with enough samples across domains and runs with MTP it should even out around concrete numbers, but I don't have time currently for long tests. On a quick test (before I remembered MTP is on), 27B was around 60-70 ts and 35B around 180-200 ts, both going up and down but mostly in those ballparks, which is inline with the ~3x from not using MTP.

One somewhat related thing is that, without drafters (the models doing just generation) ts tends to slowly go down, while with drafters, as the model is "checking drafts" instead of pure generation, even when the avg ts has fallen to say 50 ts (27B, ~128k ctc for example), there still tend to be random spikes to up to 70 (mostly during coding), which is always funny to me but also makes models a bit more "resilient" to the slow degradation of ts.

Not by the looks of it, but it got me thinking, currently in the middle of Level1Techs coverage on the model and switchyard and he mentions "how easy it is to customize it". Fully admitting that I haven't yet read the docs, my issue with that is "we can train LORAs for 35B as well, why use this (according to benchmarks) worse model for customization instead of a slightly bigger better one?"

Assuming I eat my words after going through the docs and this is actually a more efficient model / loras adapt better, I don't see as much value in it as is, as a REAP of it (remove least-important experts, domain-locked tests show ~98% retained accuracy) to something like 20B-A3B (rouhgly matching gpt oss, which while a good model, is outdated knowledge-wise and not as good with tool in my xp).

Having a 20B-A3B model at q4 that has a lora to be your local orchestrator (delegating coding to server/cloud models) and ci/cd runner does start sounding like an appealing proposition to me, as that would fit in 16gb vram easily (fitting many consumer gpus and 24gb macs).