正在加载视频...

视频加载失败

加载此视频时出现问题。这可能是由于临时网络问题，或视频可能不可用。

Bonsai 2 dropped hours ago. It's already doing real work in Osaurus. Read two CSVs. Wrote the Python. Drew the chart. Then found the one channel losing money and told me why. 27B. Ternary. Nothing left the Mac.

Osaurus

10,748 subscribers

34,387 次观看 • 16 天前 •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

11 条评论

Osaurus 的头像

Osaurus16 天前

PrismML said the big Bonsai 2 gains were agentic coding and tool use. This is what that looks like on a Mac. The model wrote the code, ran it in an isolated sandbox, and handed back a real PNG. Not a code block. A file. Weights, ready for Osaurus:

Osaurus 的头像

Osaurus16 天前

JANG conversion by @dealignai. Model by @PrismML. Osaurus is free, open source, and runs this on your Mac today.

ClassicMain 的头像

ClassicMain16 天前

Runs extremely well on rx 7900xtx as well

Haseeb Mir 的头像

Haseeb Mir16 天前

Can i run it ony Macbook M2 Pro 8GB RAM?

Osaurus 的头像

Osaurus16 天前

Not sure about this model, but we do have a smaller model we just released that should work for 8GB

Gregor 的头像

Gregor16 天前

The 'told me why' is what I'd stress-test first. Models at 27B nail the mechanics, but confident attribution is exactly where hallucination hides. Did the channel diagnosis hold up when you verified it manually?

Muhammad 的头像

Muhammad16 天前

Which mac? (Hardware)

Osaurus 的头像

Osaurus16 天前

M5 128GB

Muhammad 的头像

Muhammad16 天前

Ultra or max?

Osaurus 的头像

Osaurus16 天前

Max

Muhammad 的头像

Muhammad16 天前

Wanted give it a try on my m1pro, guess as always i stick to deepseek and qwen

相关视频

A 27B parameter model used to need a server room. Now it runs on: • iPhone • Android • Mac PrismML's Bonsai makes it possible. • 1-bit weights • 27B params in just 3.9GB • ~90% of full precision quality (PrismML evals) • Better than the 2-bit version at less than half the size The entire Bonsai family is now live in RunAnywhere. • 1.7B to 27B models • 1-bit + 2-bit ternary • Thinking mode On iOS and Mac, Bonsai runs through ggml's llama.cpp and Apple's MLX. On Android, Bonsai runs through ggml's llama.cpp and directly on the Qualcomm Hexagon NPU through QHexRT, our proprietary inference runtime. We built custom silicon kernels to make 1-bit inference on an NPU possible for the first time. Two years ago this needed a data center. Now it thinks in airplane mode. Now available in the RunAnywhere app on the App Store and Google Play.

A 27B parameter model used to need a server room. Now it runs on: • iPhone • Android • Mac PrismML's Bonsai makes it possible. • 1-bit weights • 27B params in just 3.9GB • ~90% of full precision quality (PrismML evals) • Better than the 2-bit version at less than half the size The entire Bonsai family is now live in RunAnywhere. • 1.7B to 27B models • 1-bit + 2-bit ternary • Thinking mode On iOS and Mac, Bonsai runs through ggml's llama.cpp and Apple's MLX. On Android, Bonsai runs through ggml's llama.cpp and directly on the Qualcomm Hexagon NPU through QHexRT, our proprietary inference runtime. We built custom silicon kernels to make 1-bit inference on an NPU possible for the first time. Two years ago this needed a data center. Now it thinks in airplane mode. Now available in the RunAnywhere app on the App Store and Google Play.

RunAnywhere (YC W26)

17,587 次观看 • 2 个月前

Splash is here! A month ago, DFlash 2 ran Qwen3.8-27B at 70 tok/s on an M5 Max. Today, Splash runs it at 144 tok/s on the same Mac ⚡ Getting that speed used to take work. Splash is built around the model, with the DFlash 2 draft already in it. Two commands and you're running.

Splash is here! A month ago, DFlash 2 ran Qwen3.8-27B at 70 tok/s on an M5 Max. Today, Splash runs it at 144 tok/s on the same Mac ⚡ Getting that speed used to take work. Splash is built around the model, with the DFlash 2 draft already in it. Two commands and you're running.

Zhijian Liu

92,333 次观看 • 16 天前

Darkbloom is the first model provider to support Ternary Bonsai 2 27B -- concentrated intelligence that fits on your phone. Try it now: First 250 users get 100 million free tokens; PrismML's new flagship model: - a ternary compression of Qwen3.8 27B at 2bits per parameter. - 8.5 GB total, 5x smaller than original - keeps 98.2% of the Qwen's FP16 benchmark performance. - 75% cheaper than Qwen 27B. Qwen3.8 27B already operates comparably with Opus 4.6 and 5.6 Luna on certain tasks. This one does it in the memory of a phone. 262K context, image input, Apache 2.0. From our first run on the network, on a single M5 Max with no caching: - 35 tok/s decode at 1K context, - 31 tok/s at 10K, - 19 tok/s at 50K. But we expect more performance gain coming in a few weeks! That's a full 27B reasoning model running comfortably on any Mac. 1,000+ Macs are serving on Darkbloom right now. Go try it out!! Thank you to Babak Hassibi Sahin Lale Omead Pooladzandi Reza Sadri Tushar Bansal Evan Walters and the whole PrismML team. This is exactly the kind of model Darkbloom was built for. You can read the essay by Bonsai on Why Local AI Matters:

Darkbloom is the first model provider to support Ternary Bonsai 2 27B -- concentrated intelligence that fits on your phone. Try it now: First 250 users get 100 million free tokens; PrismML's new flagship model: - a ternary compression of Qwen3.8 27B at 2bits per parameter. - 8.5 GB total, 5x smaller than original - keeps 98.2% of the Qwen's FP16 benchmark performance. - 75% cheaper than Qwen 27B. Qwen3.8 27B already operates comparably with Opus 4.6 and 5.6 Luna on certain tasks. This one does it in the memory of a phone. 262K context, image input, Apache 2.0. From our first run on the network, on a single M5 Max with no caching: - 35 tok/s decode at 1K context, - 31 tok/s at 10K, - 19 tok/s at 50K. But we expect more performance gain coming in a few weeks! That's a full 27B reasoning model running comfortably on any Mac. 1,000+ Macs are serving on Darkbloom right now. Go try it out!! Thank you to Babak Hassibi Sahin Lale Omead Pooladzandi Reza Sadri Tushar Bansal Evan Walters and the whole PrismML team. This is exactly the kind of model Darkbloom was built for. You can read the essay by Bonsai on Why Local AI Matters:

Gajesh

48,051 次观看 • 16 天前

running PrismML's Bonsai 2 27B on a single RTX 3060. 12GB VRAM. (config below) 220K context, ~35 tok/s decode, ~550 tok/s prefill. a 27B at 1.72 bits/weight, real ternary. two packings: PQ2_0 (7.21 GB) gets 220K and the numbers above, PTQ1_0 (5.95 GB) is smaller so it takes the full 262K but runs slower -- 26 tok/s decode, 260 prefill. max context go PTQ1_0, speed go PQ2_0. needs PrismML's llama.cpp fork, kernels aren't upstream yet. stock llama.cpp rejects these files. day one config, expect it to move. config: llama-server -m Ternary-Bonsai-2-27B-PQ2_0.gguf -ngl 999 -c 220000 -fa on --jinja -np 1 --cache-type-k q4_0 --cache-type-v q4_0 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --presence-penalty 0.0 --repeat-penalty 1.0 --reasoning on

running PrismML's Bonsai 2 27B on a single RTX 3060. 12GB VRAM. (config below) 220K context, ~35 tok/s decode, ~550 tok/s prefill. a 27B at 1.72 bits/weight, real ternary. two packings: PQ2_0 (7.21 GB) gets 220K and the numbers above, PTQ1_0 (5.95 GB) is smaller so it takes the full 262K but runs slower -- 26 tok/s decode, 260 prefill. max context go PTQ1_0, speed go PQ2_0. needs PrismML's llama.cpp fork, kernels aren't upstream yet. stock llama.cpp rejects these files. day one config, expect it to move. config: llama-server -m Ternary-Bonsai-2-27B-PQ2_0.gguf -ngl 999 -c 220000 -fa on --jinja -np 1 --cache-type-k q4_0 --cache-type-v q4_0 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --presence-penalty 0.0 --repeat-penalty 1.0 --reasoning on

Doğukan

161,755 次观看 • 17 天前

Keith Lee really doing God’s work. He posted this 2 hours ago and the man already got 100 calls and a bunch of cashapps. This really be warming my heart fr

Keith Lee really doing God’s work. He posted this 2 hours ago and the man already got 100 calls and a bunch of cashapps. This really be warming my heart fr

Somebody Thick Ass Mama🤎

3,334,320 次观看 • 3 年前

15 LOSING Traders vs One Of The BEST In The WORLD🚨 15 losing traders sit down with a $30M trader for brutal live breakdowns on overtrading, revenge trading, gambling, and the real reasons they keep losing money. Umar Ashraf holds nothing back. Full episode👇

15 LOSING Traders vs One Of The BEST In The WORLD🚨 15 losing traders sit down with a $30M trader for brutal live breakdowns on overtrading, revenge trading, gambling, and the real reasons they keep losing money. Umar Ashraf holds nothing back. Full episode👇

Riz Iqbal

163,355 次观看 • 4 个月前

SNL's Trump: "People are asking why we had to strike in the early hours of Saturday. One, it's after the stock market closes for the weekend. And two, it's to cause immeasurable fear, rage, and chaos in the SNL writers room.

SNL's Trump: "People are asking why we had to strike in the early hours of Saturday. One, it's after the stock market closes for the weekend. And two, it's to cause immeasurable fear, rage, and chaos in the SNL writers room.

LateNighter

1,776,434 次观看 • 7 个月前

I found out how to speed glitch out of LeBron stepback consistently… First 2 clips is me doing it to the left then right Last clip is me mixing the Kemba,Dbook, and Lebron speed glitch all into one 🔥 #NBA2K26 #Godfather

I found out how to speed glitch out of LeBron stepback consistently… First 2 clips is me doing it to the left then right Last clip is me mixing the Kemba,Dbook, and Lebron speed glitch all into one 🔥 #NBA2K26 #Godfather

Chapuh

61,197 次观看 • 8 个月前

Nassim Taleb explains in under 3 minutes why the order of your trades matters more than your win rate Making $10K then losing $10K is not the same as losing $10K then making $10K - the second one can kill your account This is why 3 Polymarket bots ran 48,000 trades and didn't blow up once Watch the lecture. Then read the full math behind $1.3M in 30 days ↓

Nassim Taleb explains in under 3 minutes why the order of your trades matters more than your win rate Making $10K then losing $10K is not the same as losing $10K then making $10K - the second one can kill your account This is why 3 Polymarket bots ran 48,000 trades and didn't blow up once Watch the lecture. Then read the full math behind $1.3M in 30 days ↓

0xDipper

318,382 次观看 • 5 个月前

Goldman Sachs hired this guy at 20 regardless of his grades JP Morgan later headhunted him and gave him $100M on day one After he left Lehman Brothers, the bank collapsed with $600B in debt He explains exactly why Goldman makes money 90% of trading days It's the same mathematical foundation I documented in the article below Watch the video first Then read the blueprint

Goldman Sachs hired this guy at 20 regardless of his grades JP Morgan later headhunted him and gave him $100M on day one After he left Lehman Brothers, the bank collapsed with $600B in debt He explains exactly why Goldman makes money 90% of trading days It's the same mathematical foundation I documented in the article below Watch the video first Then read the blueprint

Skynet Insights

114,995 次观看 • 4 个月前

Nothing lays bare the failures of one party rule in California like seeing drug addicts and the mentally insane dominate our streets. It's why I'm running for State Assembly. 36 hours left to vote- your community depends on it 🫡🗳️

Nothing lays bare the failures of one party rule in California like seeing drug addicts and the mentally insane dominate our streets. It's why I'm running for State Assembly. 36 hours left to vote- your community depends on it 🫡🗳️

Dick Lucas 🇺🇸

18,250 次观看 • 4 个月前

“i’m not a cat, i don’t say meow” 🐮 8 years ago today, doja cat’s “moooo!” verified episode dropped 🎶 the rapper wrote and recorded the viral track in just a six hours inside her bedroom. 🤯 head to our youtube channel to watch the full #verified episode.

“i’m not a cat, i don’t say meow” 🐮 8 years ago today, doja cat’s “moooo!” verified episode dropped 🎶 the rapper wrote and recorded the viral track in just a six hours inside her bedroom. 🤯 head to our youtube channel to watch the full #verified episode.

Genius

55,557 次观看 • 1 个月前

A Mac Mini M4 ran 3 hours 50 minutes on a pocket power bank. It idles at 1.5 watts. Light work pulls 3 to 5 watts. The same machine that once needed a wall outlet now sits on a desk as a full home server, drawing less power than a night light. An iPad becomes the wireless display. Touch works. Apple Pencil works. Then the Vision Pro takes over for a bigger screen. The tiny bank never dropped below critical. No wall plug. No UPS. Just the Mac Mini doing real work for almost 4 hours. That efficiency is why the M4 Mini is the new default for always-on home servers. Quiet. Tiny. Cheap to keep alive. The power bank still had charge left.

A Mac Mini M4 ran 3 hours 50 minutes on a pocket power bank. It idles at 1.5 watts. Light work pulls 3 to 5 watts. The same machine that once needed a wall outlet now sits on a desk as a full home server, drawing less power than a night light. An iPad becomes the wireless display. Touch works. Apple Pencil works. Then the Vision Pro takes over for a bigger screen. The tiny bank never dropped below critical. No wall plug. No UPS. Just the Mac Mini doing real work for almost 4 hours. That efficiency is why the M4 Mini is the new default for always-on home servers. Quiet. Tiny. Cheap to keep alive. The power bank still had charge left.

HodlReaper

12,887 次观看 • 1 个月前

Brett Meiselas: "Why does Congress as a whole work less than the average American. You would think that this body, people who are running the country, that they should actually be working and be in the office. Some people work in one job for 12 plus hours. Some people work in two jobs, three jobs. And then you see these people being paid with taxpayer money, just taking off all the time that they could get."

Brett Meiselas: "Why does Congress as a whole work less than the average American. You would think that this body, people who are running the country, that they should actually be working and be in the office. Some people work in one job for 12 plus hours. Some people work in two jobs, three jobs. And then you see these people being paid with taxpayer money, just taking off all the time that they could get."

MeidasClips

15,567 次观看 • 4 个月前

Same bug: - Codex: Two hours, nothing. - Opus 4.6: One prompt, fixed. We live in the best time in the history of programming. If you're a DataFast user, the real-time map is coming soon to the iOS and Android apps!

Same bug: - Codex: Two hours, nothing. - Opus 4.6: One prompt, fixed. We live in the best time in the history of programming. If you're a DataFast user, the real-time map is coming soon to the iOS and Android apps!

Marc Lou

108,848 次观看 • 8 个月前

🇺🇸 Trump returned to the White House 2 hours after this major event just to brief the press, called the room "totally unified," then reminded everyone why he thinks that unity means nothing. "It's not so much the press. It's the press plus the Democrats because they're almost one and the same." 4 years of war with the media, and the message hasn't changed by a single word.

🇺🇸 Trump returned to the White House 2 hours after this major event just to brief the press, called the room "totally unified," then reminded everyone why he thinks that unity means nothing. "It's not so much the press. It's the press plus the Democrats because they're almost one and the same." 4 years of war with the media, and the message hasn't changed by a single word.

Mario Nawfal

49,546 次观看 • 5 个月前

Ox Alpha, Qwen max/27b and DS4 Flash were told rebuild the Warcraft III menu live in the browser with OMP. Each model ran as an autonomous coding agent it analyzed my capture of the real menu (video and images), then built it over 5 turns. Which one got the closest?

Ox Alpha, Qwen max/27b and DS4 Flash were told rebuild the Warcraft III menu live in the browser with OMP. Each model ran as an autonomous coding agent it analyzed my capture of the real menu (video and images), then built it over 5 turns. Which one got the closest?

Loktar 🇺🇸

24,196 次观看 • 1 个月前