Join the discussion
Write your take first — we'll ask for email only when you're ready to publish.
- Hacker News
- I am also seeing slower speeds, roughly the same ballpark, sometimes even lower - 10-11 tok/s on M5 Max. If there is that ONE version (GGUF or MLX) that runs roughly as fast as 3.6 used to run, please let me know.
What would be incredible is the 3.8 35B MoE version too, I can run 3.6 with 60 tok/s which is a really, really nice speed.
by iagooar - Apart from learning, I cant see the point of spending that kind of money and energy to get such awfull token generation speed. Assuming memory bandwith is the bottleneck, is it just a matter of time until we start to see hbm4 based chip able to run qwen3.8 for normal people running at more than 500 token / seconds? Is memory speed the only technological bottlenecks that prevent us from having fast local model?
- A slow AI tasks that can process my self hosted personal journal, my medical history in fasten, my diet and exercise in Mealie and Sparky Fitness to offer insights once a day or even once a week is better than nothing. Because I would never upload that data to any of the AI companies.by Larrikin
- I have 2x 3090s and I get 260 tokens/sec peak (110 avg) with Dflash2 and about 1500t/s prefill with a Q4 quant of Qwen 27b. I didn't buy an overpriced Apple product and it performs much better. It's extremely reliable for Agentic coding and I can run 2-3 simultaneous agents with a full ~260k context window.
I realize the price of NVIDIA has gone up but there are plenty of GPU options from others like AMD and Intel with reasonable performance.
by oceanplexian - The M3 Ultra was a lot cheaper for most of its lifetime, and really the main reason for buying it is if you want a lot of unified RAM, to run bigger models than 27B models.
The new M5 Ultra should deliver ~50% faster token generation (1.2TB/s mem bandwidth); and extrapolating from my M5 Max (since the M5 Ultra is literally just 2x Maxes), probably ~3x faster PP.
But I don't think it's fair to look at this only from monetary ROI vs API. With local models, you get privacy and ownership.
I do not trust _any_ API provider with my most personal information; such as for example, all my messages, emails, daily journals spanning a decade+, all my photos and videos, etc. So it unlocks new use cases that I simply don't feel comfortable with via API.
And a personal assistant with ALL my context and data, locally, has been incredibly useful for me :) Zero outages either, zero "overloaded", etc. Nearly-zero refusals too (I don't run abliterated models; thinking prefill has worked for anything I've wanted to do)
by dannyw - For me, a HUGE benefit to running local models is that my data stays mine, on my computers only.
Even if Google, OpenAI, Anthropic whatever promise not to use it, how sure can I be of that? Data is gold anyway. And we're in the middle of a massive gold rush. And they've already been caught scraping sites they had no business to, and pirating books. Clearly their promises and the law mean nothing to them. It's just something you pay off in a settlement if you get caught, a cost of doing business.
With local models besides the speed you also lose a lot of inference quality but it helps to mitigate that. For example making sure your RAG inputs are properly prepared and categorised so the model can find them easily without having to wade through a bunch of misdirected crap.
For example what I do with my bookmarks and chats (the latter are recorded per day), is before I enter them into a RAG corpus I run a small LLM over it to summarise what's being discussed or what the webpage is about. That really helped retrieval quality, and doing this is a batch task that can run asynchronously so speed is not very relevant. This way I get a lot closer to SOTA-model retrieval quality (like with Office Copilot 365 looking for conversations in Teams).
PS: I wouldn't be surprised if Microsoft runs something similar on their end :)
But yes the Mac Studio is outrageously priced especially for running such a small model as qwen 27b. For that price you can use something much much cheaper. It only shines for models that are much bigger, because there simply is not much hardware that can address fast 512GB banks.
by wolvoleo - I am running Ornith 1.5 Q4 on my work issued laptop which surprisingly has 40gb of RAM and 20GB of “unified” VRAM.
I am getting 50t/s of prefill and 5-6t/s of inference.
You can’t get crazy with it, but it’s pretty nice to just let it work away at simpler prompts overnight or during lunch break or while I am working on other things.
Being able to literally a 30gb blob and then being able to talk with it while running on your machine it’s a different kind of magic compared to a far away magic in the cloud.
by FacelessJim - These numbers are a lot lower than I expected from such pricey hardware.
I'll stick with my old AMD datacenter cards thanks. An Instinct MI50 32GB would cost around $550 or so. It has 1TB/s vram bandwidth thanks to its HBM2.
by wolvoleo - It's a software issue.
- any numbers to share? also, what inference engine do you use and with which api?by jdright
- I am also running some old AMD datacenter cards, 2x MI25 in my case. Getting around 30 tokens/second with short context.
Tensor parallel in llama.cpp using RCCL (disabled by default in llama.cpp for some reason). Surprisingly, for these cards HIP is actually faster than Vulkan, unlike the 9070 XT where Vulkan still wins.
ROCm nightlies do actually support these old cards, just not the ROCm stable releases.
by fotcorn - the fact that this author cannot get qwen3.8-27b run at the same speed as qwen3.6-27b, says the article is not worth reading. the author does not know anything about how to run local AI. 3.8 and 3.6 are the same model with different weight.
both tg and pp speed are so terrible on author's machine.
by ycui7 - The cheapest card that will run this model very well is a ln unlocked CMP 170HX. But you can run it on a 3090. I run it on an old spare A6000 Ampere. I think I wouldn’t use anything lower than 60 tok/s though, which you can get with MTP etc. I just use a full vllm stack but some people see a lot of speed with ninfer (there are non 5090 ports).
The large RAM Macs are unusable for inference of dense models as of now. Token generation is too slow.
by arjie - I am really surprised you’re seeing half the generation performance of 3.6 with 3.8 at the same parameter size and quantization (and same prefill performance to boot) - is there just an optimization in the stack somewhere for 3.6 that hasn’t landed yet for 3.8?
Curious if other folks are also surprised by this apparent discrepancy or have a ready explanation.
by Infernal - I suspect this might be due to MTP mispedictions. Either because the 3.8 quantized model weights do not include MTP heads, or as what happened with Ornith-1.5 recently, corrupted MTP heads, or due to a software issue in Ollama.
- Pretty impressive how the Mac ends up less expensive than the Strix Halo boxes, at least here in Ireland. A 128GB Mac Studio with an M5 Max (the Ultra can only have 96 or 256GB) still costs less than the "GMKtec EVO-X2" or the Nvidia DGX Spark with similar performance. Is it the same in the US?by rbanffy
- That would be such a sizeable disconnect that you might do some arbitrage.
Current prices in USD:
128 GB 1 TB M5 Mac Studio: $5399
128 GB 1 TB GMKtec EVO-X2: $3499
128 GB 1 TB Framework Desktop: $3748
by bronson - That's not right. Mac Studio M5 Max 128gb, in Apple Store Ireland, starts from 5,909e (~6,847 USD; tax included). On the other hand, GMKtec Evo-X2 128gb can be currently purchased for 3,299e (~3,822 USD; tax included) from GMKtec online store (free shipping from DE warehouse). It's quite the difference in price.by lowlat
- Yea Qwen3.8 wasn't fun to use on my 64GB M4 Max either (better than these numbers though), so my new daily driver is Ornith-1.5-35B-A3B-MLX-4bit. I recommend giving that a whirl if you're on similar hardware, it's definitely better than Qwen3.6 35b-a3b which was my go-to before.by pwython
- Can you tell more about your experience with Ornith? I've come across it, and the benchmarks on its landing page are unbeleavably good. But it isn't featured on more established benchmarks like ArtificialAnalysis and it's not on OpenRouter, so I wrote it off as scam.by TobTobXX
- I tried Ornith-1.5-35B-A3B-MLX-4bit on a 32G M2-Pro mac mini with pi and OpenCode. I didn’t get very good results coding in Python, Racket, and TypeScript. I have seen several positive comments like yours so I was probably doing something wrong. I amgetting the new 64G mac mini in 4 weeks, and I made a note to try Ornith-1.5-35B-A3B-MLX-4bit again.
- I'm encountering the same behavior. I've tried 4-8bit quants and get 14-17 tok/s with one run that achieved 19. I'm eagerly awaiting dflash2 support in Unsloth or LM Studio, as allegedly that should increase throughput to around 30tok/s, which is the baseline for what I consider at least somewhat interactive.
Jealous of the folks with 5090s running ninfer and getting >100tok/s. At those speeds it's a true frontier replacement IMO.
by Atreiden - dflash2 is atmost 10-20% above mtp, it won't get you to 30 tpsby nialv7
- I run similar workflows as the author on my 5090. It's really good and reasonably fast at ~90 tok/sec on LM Studio. I haven't tried ninfer yet. The only problem is having to be mindful about the context size. I am jealous of the folks with RTX 6000.by pllbnk
- That isn’t right. Use oMLX or something similar to serve the model with internal MTP enabled. You should get at least 40 tok/sec, but I’m not sure what hardware you are using. Dflash will mess up batching, not really worth it. You need to sift through hugging face for the right model though, and some of the MTP models are meant for rapid MLX and not oMLX.by seanmcdirmid
- Qwen 3.8 has the same architecture and the same parameter count as Qwen 3.6. Something is not right with the GGUF if it's 2 times slower. The post says "The hybrid attention architecture is new" and says the author's older Llama build from a "couple weeks ago" failed to run Qwen 3.8 because it did not support Qwen35 architecture, but both 3.6 and 3.8 are based on Qwen35 which was released in February 2026. The post doesn't make any sense.by kgeist
- I've been thinking about buying a system to run LLMs locally but the price for one that'll run Qwen3.8-27B well is quite offputting to say the least.
What I've been looking at instead is inference providers that use TEE and E2EE to provide cryptographic guarantees that my prompts and responses are only visible to me and the GPU itself.
Despite their docs and assurances of what their guarantees mean, I'm having trouble getting to a point where I'm actually comfortable trusting them with secrets though. Phala for example seems to be E2EE only to the gateway and will then forward prompts to (potentially third party) providers.
Has anyone been down this path and found a provider they feel safe with?
by Youden - Well, yes me. I was on the same path. Researching hardware and coming to the same conclusion. I found tinfoil.sh which looks promising?
My current solution is a private ChatGPT-like interface using OpenRouter’s API with Zero Data Retention enabled. Not perfect or verifiable but I think it’s acceptable for now.
by lzy - Not suggesting a provider but if you're willing to get a Chinamod GPU, go get RTX 3080 20G or RTX 2080 Ti 22G. Get a couple of them and you can probably run Qwen3.8-27B at a reasonable speed.
I've got a RTX 3080 20G for $450 like a year ago. With llama.cpp, bf16 kv cache, kv cache offloaded to RAM, Qwen3.8 27B UD-Q4_K_XL, single RTX 3080 20G, I got 10 tok/s initially and it dropped to 5 tok/s at 50k context. I'm thinking of getting another Chinamod GPU.
Building a dual Chinamod GPU machine would only cost like $1000~$1500. It's not too bad compared with the alternatives!
by jan_Sate