I just don’t get it. Like…who’s buying these things in quantities vast enough to charge that much? Oh, I’ve heard Open AI snapped them and Mac Minis up in bulk for “testing” (whatever that means)… but given the cost now… why / who is buying them in quantities that warrant these sort of prices? It’s not as if they’re rare per se. Hell, there might even be a glut relative to actual demand. Yet, those prices…
Human psychology is weird.
Meanwhile you made me look at prices. JFC.
Oh, your running some JEVs? How are they? I just can’t figure out a use for em.
Yes, concurrency is amazing. I can pull some tricks with CPU (OCR, TinyBERT, Klein etc) while the GPU runs inference…but it’s not a Strix.
Yeah, the BC-250 had / has real promise, esp with the core unlocks…but like P40s and P100s (server cards) people twigged on “this is cheap and fast”. I can’t get one locally for under $600 now. They use to be $150. By the time you fit it out properly, it’s nearly $1K. Not worth it.
The curse and blessing of unified memory!
I was considering getting a BC-250 (a cut down PS5) with 16GB unified memory, but at 400GB/s bandwidth.
When they were cheap, it almost made sense. But realistically, I can get similar speeds with a MoE thanks to cpu and system ram, and 16GB is just not enough for anything top shelf (no, Prism MLs Bonsai models don’t count. I don’t care what the benchmarks say - they’re lobotomised).
Still, given the chance to get a Strix on the cheap, I’d go for it. There’s something very nice about a low power box that can run a 120B model at speed.
A Strix. In 2026. Must be nice :)
It’s a two pronged attack for sure. Make local expensive, make cloud worse.
I think there’s still some wiggle room to be had in the 8 to 16GB VRAM range…so I’m hoping Qwen 4 series brings its A game.
Enjoy and good luck!
deleted by creator
For a start, it’s lighter on system resource. By a lot.
Dear God…have mercy and provide it with the right MCP tool so it knows WTF arcane is and how to deal with it:)
https://docs.jmrp.io/arcane-mcp
https://github.com/dinglebear-ai/rarcane
PS: Arcane - I see you’re a man of culture and refinement. Hat tip
It’s something you’ll probably download directly from HuggingFace. That is - people will de-censor models and you just download them.
Of course, you can grab the tools and de-censor your own. It’s a bit of a black art tho; I abliterated a Qwen3.5-2B once (I thought quite well, actually; minimal damage 0.0006)…only to have BADLY ruined its tool calling abilities.
Leave it to the known good quantizers over on HuggingFace.
I thought Google fixed the Gemma template especially for tool calling…or is this ANOTHER regression?
Copyright infringement for me, not for thee, eh Claude?
Nvm. GLM5.3 is just as good…and a good deal more slutty about this sort of stuff :)
Friends don’t let friends use Ollama
If you want a gentle start, may I suggest Jan.ai? It’s a ride on lawn mower compared to the sports car I suggested…but you need to dip your toe into the water before you go ludicrous speed.
Once you get a feel for that, I’m going to suggest some MCP tools to you as well. I don’t know if Jan finally added MCP servers to their stack (they should have by now!).
The following are … ok. Not the best and not what I run with but again, lawn mower before sports car
https://github.com/ofershap/mcp-server-markdown
https://github.com/jeffreyrampineda/kiwix-wiki-mcp-server
https://github.com/modelcontextprotocol/servers/tree/main/src/filesystem
https://github.com/exa-labs/exa-mcp-server
https://www.home-assistant.io/integrations/mcp_server/
Sorry - again: too much info. Start with Jan. You’ll be amazed…and then perplexed…and then frustrated. That’s the time to migrate thru to llama.cpp or PolyStrata and start adding those MCP tools I mentioned above.
https://github.com/VecSzn/PolyStrata
https://github.com/srware-net/LocalLightChat/
https://github.com/open-webui/open-webui
Then we can get freaky :) Oh…we have such sights to show you lol
Depends what your interests are! I can get around 35ish tok/s TG at - c 48,000 of Qwen 3.6-35B-A3B (Q4_K_M) using polystrata on a similar rig (you’ve got an A1000, I have a T1000).
Which means absolutely nothing to you right now but suffice to say - Polystrata + Qwen 35B is something to try. It’s among the “smartest” you can run on your rig.
Personally, for every day use, I like Ling 3.0 Tiny at Q6 or SharpMiniCPM5-2B at Q8. They are stupidly fast, and supported by MCP tools and a bridge I wrote, work as excellent “hey, I’m thinking of X. Can you tell me about it? And what about Y and Z?” or “How does this read?”.
I pull out my phone, send them a text message and they text me back.
Think of it like a Google in your pocket, except off line (unless you tell it to go online). Which technically I suppose makes it agentic lol.
(I can show you some examples if you’d like. I keep a set list of 8 tasks I use for benchmarking as I tweak things, as well as some off the wall / out of band queries)
Being that Ling or CPM only take up a sliver of the GPU, I can run them concurrently with other models (like an image generator, text to speech, vision etc) to create a really useful offline house brain, that can roam across my local documents, offline wikipedia (Kiwix) and the like.
It’s the combination of things that results in a 1+1=3
Having said that, my AI ladder is
For now, I keep a $20 ChatGPT Plus (separate pools for chat and Codex have been so good) but I am ditching that asap, with the changes coming this month.
HTH. Write back if you need anything else. It’s nice to see someone else rocking an A1000 / T1000. Your rig should be significantly faster than mine… which is why the A1000 is still stupidly expensive for a 50W card.
EDIT: just saw you said “CPU is i7-12700”. With that RAM, that CPU and that GPU, I’d guess you’ll hit 50-60 tok /s with Qwen 3.6-35B using polystrata once you get the GPU / CPU split right.
In other words, congrats and go fuck yourself :) You have a very nice little inference box. And yes, it can do your calendar alerts thing.
That should be well within reach of a local LLM.
If you’ve got a discrete GPU with sufficient VRAM, something like Ornith running on PolyStrata could be worth a shot. Try a Q3 size.
https://huggingface.co/bartowski/Ornith-1.5-35B-A3B-GGUF
https://github.com/VecSzn/PolyStrata
Hell, Tiny Ling 3 might be able to do it. Try a Q6.
https://huggingface.co/bartowski/Ling-3.0-tiny-GGUF
Another good option is Qwen3.5-9B. I’d try a Q4 or Q5.
Whichever model you choose, I’d also suggest a few MCP servers to give it access to reliable information rather than rely entirely on its internal knowledge.
https://learn.microsoft.com/en-us/training/support/mcp
https://github.com/exa-labs/exa-mcp-server
https://github.com/modelcontextprotocol/servers/tree/main/src/filesystem
If you’re actually running Home Assistant, its official MCP server is worth a look too.
https://www.home-assistant.io/integrations/mcp_server/
I’d start with Microsoft Learn and Exa, then add the others as needed. You’ll need an MCP-capable client (which pretty much all of them are now - eg: llama.cpp web ui, OpenWebUI, Light Local Chat etc), and I’d keep file access read-only initially.
I’d personally use McProxy instead of adding MCPs singularly, but I fear I’m already over loading you with information.
EDIT: The PolyStrata recommendation assumes you have a discrete GPU with sufficient VRAM (unclear if you do?). If not, disregard Ornith and try Tiny Ling 3 Q6 or Qwen3.5-9B with ik_llama.cpp instead.
https://github.com/ikawrakow/ik_llama.cpp
Either way, both smaller models should be capable of the tasks you’ve described, particularly with access to the right documentation. Just verify any network or firewall changes before applying them.
Yes, quite a few, but it depends on what you’re trying to do with them. Any ideas?
They are the way to go, but hardware pricing makes it cost-prohibitive to match frontier capabilities, which makes decomp and recomp difficult. Or rather, they require more skill and take longer, introducing some barriers against the current floodgates being flung open.
We all know the eye-watering prices GPUs and RAM cost now, to say nothing of Unified Memory devices like the Strix. Most people just don’t have a spare $10K to throw at an inference box that can run (say) a 300B model at interactive (not batch) speeds.
Coming at this from another angle tho, I note that according to Steam stats, the most popular consumer GPUs are in the 8GB and 16GB range…which is smack bang in the VRAM range of what a good MoE (say, Qwen 3.6 35B-A3B) runs on (for 8GB) or arguably the best local dense coder (Qwen 3.8 27B) can live at. Which leaves us with a capable assistant and a capable coder.
I think there’s probably room for LLM improvement in the 8–16GB range too…and given that most consumer PCs and laptops live there, I’m sanguine the open-weight devs might make it happen.
But I also suspect the gatekeeping of high-end capacities is intentional.
We’re probably not going to have the ability to vibe code Mario into Elden Ring again…not without significantly more skill. Dunno.
Impossible to say. The token use difference between instant (nigh on useless,) and medium or high (what people anecdotally default to) could be significant.
Eg: if we consider GLM 5.3, the difference in token use between reasoning none and reasoning high is about 3x
https://artificialanalysis.ai/models/comparisons/glm-5-3-low-vs-glm-5-3
But without access to actual token consumption figures, comparing the old and new usage allowances is difficult. Suffice it to say it’s unlikely to be in your favour.