basically just the street price of each component added up. not a good package deal tbh.
if you have no components whatsoever right now, and just want a fair amount of ram + vram...maybe? i would lean more towards AMD for pre-built rigs that have value but tbh i havent done a head to head of AMD and Intel model libraries (just nvidia vs either).
Wow this is big news for those looking to get the most out of some lower power hardware.
Beyond getting quality outputs for less hardware, I could see this helping with MVPs or V1s of early bootstrapped start ups feeling the AI hardware crunch. Should help them validate faster, and maybe even spur investment, find co-founders, etc.
With the phone potential, I could also see MLX getting more love from moves like this. Good stuff.
Gonna fire up my old 2019 alienware rig (w/ radeon 5700 xt) to see how it performs.
Today, we’re announcing Bonsai 27B: the first 27B-class model to run on a phone.
Bonsai 27B is the new multimodal flagship of the Bonsai family. Based on Qwen3.6 27B, it brings a new capability tier to local AI: multi-step reasoning, structured tool use, long-context workflows, and coherent agentic loops.
Until now, models in this class have been impractical to deploy locally. A 27B model occupies roughly 54 GB in 16-bit precision, and even a strong 4-bit build is around 18GB - too large for a phone and for most laptops.
Bonsai 27B changes that.
It comes in two variants:
• Ternary Bonsai 27B: 5.9 GB, 1.71 effective bits per weight, optimized for laptop-class quality.
• 1-bit Bonsai 27B: 3.9 GB, 1.125 effective bits per weight, optimized for phone-class footprint.
Everything is open-sourced today under the Apache 2.0 license.
@ZackKorman “Oh hey how do I use AI to help my lil company out? I got some invoices and maybe it can find some patterns to help me and my team grow!”
**sees veteran graveyard**
“Jesus Christ”
@regulatorynerd personally I'm a bigger fan of Zoho CRM 2: Electric Boogalo. better soundtrack, ways to elevate KPI measurement with senior stakeholders, and you dont miss much plot contuinity jumping stright into it.
@1337hero@MLWithPhil got it, thx guys. and yea, im more building proof of concepts not so much daily drivers. 2x speed prob far more useful day to day than ecc.
@1337hero yea was looking into the tradeoffs of speed vs some stability on longer runs for 5090 vs 4500. I do plan to do a fair amount of overnight loads on large docs/datasets and felt the ECC would help out some (and help a bit with thermal/lower power loads, etc).
True story: when I was a kid, my neighbor across the street wasn’t allowed to go past the level with Boos (donut plains?) on Super Mario World since boos were ghosts and ghosts were anti-Christian. He only played the first several levels up until that point, and didn’t think anything of it when he had to reset or turn it off.
I would say that the 'no' 30% + ~half of that 'yes limited' 50% are still in procurement for AI dev tools. either the limited tokens are a pilot run or were put in place after rising costs. And let's go with 1/3 of the 'yes unlimited' share will have leadership scrambling once they see end-costs or get sticker shock.
That rough culminate % tracks with what I heard from the ~100 or so enterprises + mid-sized companies I briefly chatted to at our big conference a few months ago.
In other words - we're still early for agentic dev tool opportunities (at least if you're targeting the bigger players).
@VladTheInflator I get what you’re saying but there were some generational opportunities recently with the late 00’s early 10’s post-recession, and covid era valuations. Both ZIRP phenomenons.
About 2 weeks. There’s ways to get Mongolia in a week and still cover good ground. Rented a aussie spec Ford Ranger. No cell service for 99% of the trip, but planned ahead with waypoints for gas stations and potential camps but also found stuff along the way. Had an awesome time at a small village bar and drank copious amounts of vodka with a Mongolian wrestler who then tried to drunkenly force his way into my truck. He then chased me in his car with his friends (he was wicked drunk), marking my first hot pursuit. Got human fleas for a day in this Soviet era building hotel thing in pissed stain mattress. Forged a new road in a barren desert after going in this gap between the gobi dunes and almost got stuck in sand since my offline map app said there was a road (saved about 3 days not backtracking though, worth it). This trip made me a man and I hope to bring my daughter (who is about to come into this world in a few months) on the same or similar journey.
That, or devs more than likely didn’t have limits on usage. Or if they did, they weren’t strong such as frequency caps, team/user, etc. Even for our platform, we evangelize cost control and optimizations to our harness design in our GTMs, but had some random employee added to our instance that blew through $1k of tokens in a week without anything to show for it. We started to put more policies in place to enable and govern.
That’s just for marketing. Could only imagine runaway spend with actual builders.
@ralphbakshi loved it. can i do a remake, ralph? i know that's in studio hands but even if so, i would also want your blessing. i could probably get gia coppola to direct.