I've created a 2.54BPW quant that fit on my hardware with 128k context, 20 tps tg and 200tps pp, while maintaining high scores on many benchmarks: https://huggingface.co/tarruda/Qwen3.5-397B-A17B-GGUF/discus...
If by ROI you mean saving more money than using paid APIs, then I don't think it is worth it. All you gain is full sovereignty over your AI usage.
Running w/ Cursor and doing some "nights and weekends" type coding / conversations, I was hitting $100-200 of usage within a few weeks. I know there's probably better ways to manage costs, but I was getting enough value out of it to keep bumping my spend limit from $20 => $40 => $80 => $120 (and then I stopped spending! :-)
Messing around with local-llm, I've settled on `omlx` and `gemma` for "conversational", and I think it's `qwen-120b-a3b-6bit` or something for the "heavy hitter". Gemma "gets it" a lot more, whereas that particular `qwen` tends to fall into the "MuSt WrItE CoOooDeee!" behaviour in a lot of cases instead of holding a conversation, and does an awesome job of randomly spitting out ascii-art diagrams or including full-blown bash shell scripts to illustrate different cases.
My POV is: "Local for slightly slower/casual usage", the ~1% of battery usage per minute of LLM is shockingly accurate (eg: 30 minutes == 30% drop!). "Gemma for discussion and emitting DESIGN-... docs", and "Qwen for converting DESIGN-... to PLAN-...", (as well as implementation, but generally from a fresh context loading the relevant PLAN-... or supporting docs)
...then supplement that with direct Cursor usage in case I screw up some setting on being able to get the local LLM working, or if I need to include literal web-research or really having access to some SOTA model. Using the pi-coder harness locally, web pages are kindof a difficult conundrum as they can be kindof gigantic and are really worthy of special casing, some sort of sub-harness, etc... but the more "stuff" you put into the agent, the less context window (and memory!) you have available, so it's a real balancing act.
The other biggest problem is that you're limited (locally) to ~20-80tps and in some cases you have to chew on or "swallow" the whole prompt up to that point if you end up with some sort of cache miss (TTFT). The `omlx` server does a pretty good job (after you tweak some settings and stuff) of allowing MANY prompt continuations to nearly immediately start generated tokens, but sometimes if I have two agents going (eg: Gemma talking shit about Qwen's output or vice versa) in a longer context window, then you'll take that hit.
"Other people's compute" is definitely more freeing, but even looking at $200/mo usage that's $2400 vs. the ~$6k for a maxed out MBP. Call it $2500 vs. $7500 and you'd say that "local AI gives you a 3-year amortization window for a slower, worse experience" ... but if you're strategic about your usage, the ability to "talk for free" and occasionally "burst" to an online provider or having some hugging-face tokens to try out different models that you can't quite run locally is really nice. Talking to the AI (locally) to even just do non-coding planning without worrying about data leakage or privacy issues is phenomenal, and you end up owning a really nice laptop!
In some ways, seeing the "advantage" of having the local 128gb capacity for LLM, I'm semi-wishing I'd have gotten a mac mini instead, but then I can't quite do the 100% offline stuff (eg: coffee-shop) that the maxed out laptop allows.
If it were a mini running locally, I'd feel more comfortable calling it the always-on "AI brain" to process my emails, run crontab summaries, whatever kindof "open-claw-ish" stuff that you could do w/o relying on having to "keep the laptop lid open all the time". I'm sure there's ways to repurpose things, but longer-term, call it even 3-5 years from now... any sort of 128gb machine will be more than capable where you'd want to have one "doing stuff" locally within your home network (IMHO).
>"...if you're strategic about your usage, the ability to "talk for free" and occasionally "burst" to an online provider or having some hugging-face tokens to try out different models that you can't quite run locally is really nice. Talking to the AI (locally) to even just do non-coding planning without worrying about data leakage or privacy issues is phenomenal, and you end up owning a really nice laptop!"
^ this resonates, loudly.
But I was not super impressed with deepseek 4 flash using it from the official API either, so it doesn't seem quantization fault. It is a good model, but nothing out of the ordinary in the few benchmarks I ran on it (with full awareness that benchmarks are biased).
Oddly enough, though, Qwen 3.6 35B A3B and Gemma got some really good reviews, despite being way smaller than any of these ones.
Qwen 3.5, 122B A10B: https://huggingface.co/unsloth/Qwen3.5-122B-A10B-GGUF
Qwen Coder Next, 80B A3B: https://huggingface.co/unsloth/Qwen3-Coder-Next-GGUF
It's kinda weird that DeepSeek V4 Flash is supposed to be 284B A13B, but shows up as 158B in HuggingFace, probably some weird bug: https://huggingface.co/unsloth/DeepSeek-V4-Flash and that's not even just Unsloth but like the official source too https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash (so also doesn't fit the category unless you get a heavily quantized version to run, but cool regardless)
Mistral Medium 3.5 is interesting because it's 128B but dense, so probably too slow for most folks: https://huggingface.co/unsloth/Mistral-Medium-3.5-128B-GGUF
GPT-OSS, 120B A5B: https://huggingface.co/unsloth/gpt-oss-120b-GGUF
I’m on an M1 Max with 32GB VRAM, so I’m looking forward to the 27B or 35B-A3B models. Is dropping $5k for an RTX 6000 or a DGX Spark really the best option?
- Your RTX 6000 is closer to $10k now
- Sparks are creeping into the $4-5k range
- AMD Strix are ~3.5k
- Apple depends on chipset and memory. Sweet spot would be 128gb M3 Ultra, probably $6-8k but admittedly haven't been tracking closely. New M5 might come in the fall. You can get a new 128gb M5 Max laptop for ~5-6k today.
- a 4x3090 rig would take $5-6k
Every platform has tradeoffs, but it's mostly ecosystem, memory bandwidth, and power consumption. They're all slow. The best option is likely to rent hardware on Runpod. The RIO on self-hosting is very low unless you have a specific need or you're ok treating it as a hobby.
>The best option is likely to rent hardware on Runpod.
Vast.ai is much cheaper, but the broader point here is contestable. The only dimension in which cloud GPU rentals win is cost. You lose the confidentiality, integrity, and availability benefits of local deployments.
Not that I'd encourage anyone to throw large amounts of money to have access to LLMs, but you're definately going to be better off buying something that you can amortize over multiple years with a multi year warranty.
This whole thing is really starting to remind me of the crypto hype phases of 2016-2018 when everyone thought their investment in GPUs was going to make them rich.
Yes, LLMs are sloppy, and local models usually more so (but things change fast).
But the local ones have one big advantage: they are private. So you can safely feed them the collection of your private documents and things you wouldn't trust people like sama with. The fact that some people do not care is one of the failures of our educational system.
In October/2024 I got my Mac studio M1 ultra with 128G, IIRC it was ~$2500. With recent prices explosion, it has certainly gotten more expensive. https://frame.work/ is selling 128G strix halo mainboard for $2700, but you have to add storage and case.
Non-NVIDIA backends tend to get less support and new features land slower, or features that are expected to improve performance wind up hurting it instead. That sort of thing.
For basic “token in/token out” workloads without fine tuning, it’s probably fine ??
Unfortunately, the prices rose on these a lot, but unevenly. Beelink GTR 9 Pro is $4400, Framework Desktop is ~$3500, for what is basically the exact same mainboard as a Bosgame M5 for $2800.
Apple's M5 Max is another attractive option. Apple silicon traditionally had great MBW and was good at TG, but struggled with PP, but the new neural engines in those GPU cores have made a big difference in a good way here.
Gorgon Halo is rumored for June announcement with Q4'26 release with basically +100 MHz clocks on Strix Halo, LPDDR5X-8533 instead of LPDDR5X-8000, but more importantly, 192 GB max instead of 128 GB.
I'd say it's better to wait for Gorgon Halo than to grab Strix Halo now. However, Medusa Halo, rumored for H2'27, is slated to have up to 26c Zen 6 (heterogeneous cores - kinds funny that AMD is heading towards these as Intel retreats from them), 48 CU of RDNA 5 instead of 40 CU RDNA 3.5, and a 384 bit bus w/ LPDDR6, which should make 256 GB at more like ~490-600 GB/s MBW, which will really make Strix and Gorgon Halo obsolete.
Also worth keeping an eye out for Serpent Lake (intel CPU + nvidia iGPU on a single board with unified memory, rumored for 2028-2029 iirc), and on the 160 GB Crescent Island Intel dGPU.