Live data from Hacker News

Kimi K2.5 Technical Report [pdf]

github.com

141–146 of 146 posts

Re: Kimi K2.5 Technical Report [pdf]

#141

Seems that K2.5 has lost a lot of the personality from K2 unfortunately, talks in more ChatGPT/Gemini/C-3PO style now. It's not explictly bad, I'm sure most people won't care but it was something that made it unique so it's a shame to see it go. examples to illustrate https://www.kimi.com/share/19c115d6-6402-87d5-8000-000062fec... (K2.5) https://www.kimi.com/share/19c11615-8a92-89cb-8000-000063ee6... (K2)

Preserving the examples for history. I wish HN had a tag. The closest would be to collapse this comment.

Edit: Thanks, moderator. :-)

---

Turn 1, user:

is replacing my main PC with a NeXTcube a good idea?

---

Turn 2, K2.5 (bold font replaced with "*foo*"):

That's a very interesting question! The NeXTcube is a historically significant machine (famously used by Tim Berners-Lee to create the World Wide Web), but I need to be honest with you: *no, it would not be a good idea to replace your main PC with a NeXTcube* for practical modern use.

Here's why:

*Hardware Limitations*

- *Released in 1988* with a 25-33 MHz Motorola 68030 processor

- *8-64 MB of RAM* (compared to modern PCs with 16-64 GB)

- *No modern connectivity*: No USB, Wi-Fi, Bluetooth, or Ethernet as we know it

- *Display*: Monochrome or limited color, low resolution by today's standards

*Software Ecosystem*

- Runs *NeXTSTEP* (which evolved into macOS), but software support ended decades ago

- No modern web browsers, office suites, or development tools

- Cannot run Windows, modern Linux, or contemporary macOS applications

*Practical Realities*

- Finding working hardware is difficult and expensive (collectors' items)

- Replacement parts are scarce and costly

- No security updates for 25+ years

- File format compatibility issues with modern systems

*What You Could Use It For*

- Historical computing experience

- Running legacy NeXTSTEP software

- Museum piece or collector's item

- Learning about computing history

The Better Alternative

If you're drawn to the NeXT aesthetic or philosophy, consider:

- *Modern macOS* (direct descendant of NeXTSTEP)

- *Linux with a tiling window manager* (similar workflow)

- *Vintage computing as a hobby* alongside your main PC

Would you like suggestions for achieving a similar experience or workflow on modern hardware?

---

Turn 2, K2 (no bold font):

Short answer: No—unless your goal is specifically to collect or experiment with a historic machine, a NeXTcube is a terrible day-to-day replacement for a modern PC.

Why it fails as a “main” computer today

1. 30-year-old hardware

25 MHz 68040, 16–64 MB RAM, 400 MB SCSI disk (if you’re lucky), no built-in Ethernet, no USB, no Wi-Fi, no audio-in, 2-bit grayscale display. You will feel the slowness in everything from un-zipping a file to loading a web page.

2. Obsolete software base

NeXTSTEP 3.3 is gorgeous but frozen in 1995. No modern browser build, no TLS 1.3, no video codecs past MPEG-1, no Dropbox/OneDrive/Zoom/Slack/Steam/Spotify. You can ssh out, but you can’t run contemporary apps.

3. Storage and connectivity pain

SCSI-2 drives are dying; 50-pin replacements cost more per GB than a 4-TB NVMe stick. Getting files on/off the machine usually involves FTP-over-Ethernet or burning a CD-ROM on another computer.

4. No security updates

The last patch was released before the commercial web really existed. Putting it on the open Internet is asking to be scanned and possibly exploited within minutes.

5. Display and peripherals

The original 17" NeXT MegaPixel monitor is fixed-frequency 1120×832@68 Hz grayscale. Adapters to VGA exist, but color is impossible and refresh rates annoy modern LCDs. Sound is 16-bit 22 kHz mono; the optical drive is a 660 MB magneto-optical cartridge.

6. Power and parts

The cube power supply is a custom 350 W linear unit that runs hot; capacitors are drying out. When it dies you will be scavenging eBay or recapping it yourself. Spare motherboards cost more than a whole used ThinkPad.

7. Productivity software

There are serviceable versions of Mathematica, FrameMaker, Lotus Improv and OmniWeb 3.x, but file-format compatibility stops at Office 97-era documents. You will constantly round-trip files through another machine.

What it is good for

- Historical research, museum exhibits, retro-programming in Objective-C 1.0 or 68k assembly.

- Demonstrating the birth of the World-Wide Web (Tim Berners-Lee’s first server ran on a NeXT).

- Enjoying the UI that later became macOS (AppKit, Dock, Services, .app bundles).

- Collecting: a working color NeXTcube with turbo slab and original laser printer is a conversation piece worth several thousand dollars.

Practical compromise

Keep your modern PC (or a quiet mini-PC/Mac mini) for real work and run a NeXT emulator instead:

- Previous (68k emulator) boots NeXTSTEP 3.3 on macOS or Linux.

- VirtualNeXT (based on Previous) for Windows.

- QEMU’s “next” target is improving.

You get the look-and-feel, can compile legacy code, and still have 4K video calls when you need them.

Bottom line

A NeXTcube is a beautiful, important machine—buy one if you love vintage hardware—but do not voluntarily make it the computer you rely on to pay bills, join Zoom meetings, or play YouTube.

Re: Kimi K2.5 Technical Report [pdf]

#142

Seems that K2.5 has lost a lot of the personality from K2 unfortunately, talks in more ChatGPT/Gemini/C-3PO style now. It's not explictly bad, I'm sure most people won't care but it was something that made it unique so it's a shame to see it go. examples to illustrate https://www.kimi.com/share/19c115d6-6402-87d5-8000-000062fec... (K2.5) https://www.kimi.com/share/19c11615-8a92-89cb-8000-000063ee6... (K2)

K2 in your example is using the GPT reply template (tl;dr - terse details - conclusion, with contradictory tendencies), there's nothing unique about it. That's exactly how GPT-5.0 talked. The only model with a strong "personality" vibe was Claude 3 Opus.

It definitely talks a lot differently than GPT-5 (plus it came out earlier), the example i gave just looks a bit like it maybe. best to try using it yourself a bit, my prompt isn't the perfect prompt to illustrate it or anything. Don't know about Claude because it costs money ;)

Re: Kimi K2.5 Technical Report [pdf]

#143
post #128

Earlier quoted context omitted.

It is possible to run locally though ... I saw a video of someone running one of the heavily quantized versions on a Mac Studio, and performing pretty well in terms of speed. I'm guessing a 256GB Mac Studio, costing $5-6K, but that wouldn't be an outrageous amount to spend for a professional tool if the model capability justified it.

> It is possible to run locally though > running one of the heavily quantized versions There is night and day difference in generation quality between even something like 8-bit and "heavily quantized" versions. Why not quantize to 1-bit anyway? Would that qualify as "running the model?" Food for thought. Don't get me wrong: there's plenty of stuff you can actually run on 96 GB Mac studio (let alone on 128/256 GB ones…

True, although the Mac Studio M3 Ultra does go up to 512GB (@ ~$10K) so models of this size are not too far out of reach (although I've no idea how useful Kimi K2.5 is compared to SOTA).

Kimi K2.5 is a MOE model with 384 "experts" and an active parameter count of only 32GB, although that doesn't really help reduce RAM requirements since you'd be swapping out that 32GB on every token. I wonder if it would be viable to come up with an MOE variant where consecutive sequences of tokens got routed to individual experts, which would change the memory thrashing from per-token to per-token-sequence, perhaps making it tolerable ?

Re: Kimi K2.5 Technical Report [pdf]

#144
post #77
post #74

Earlier quoted context omitted.

You could buy five Strix Halo systems at $2000 each, network them and run it. Rough estimage: 12.5:2.2 so you should get around 5.5 tokens/s.

Is the software/drivers for networking LLMs on Strix Halo there yet? I was under the impression a few weeks ago that it's veeeery early stages and terribly slow.

llama.cpp with rpc-server doesn't require a lot of bandwidth during inference. There is a loss of performance.

For example using two Strix Halo you can get 17 or so tokens/s with MiniMax M2.1 Q6. That's a 229B parameter model with a 10b active set (7.5GB at Q6). The theoretical maximum speed with 256GB/s of memory bandwidth would be 34 tokens/s.

Re: Kimi K2.5 Technical Report [pdf]

#145
post #77
post #74

Earlier quoted context omitted.

You could buy five Strix Halo systems at $2000 each, network them and run it. Rough estimage: 12.5:2.2 so you should get around 5.5 tokens/s.

Is the software/drivers for networking LLMs on Strix Halo there yet? I was under the impression a few weeks ago that it's veeeery early stages and terribly slow.

Check out https://github.com/kyuz0/amd-strix-halo-vllm-toolboxes/blob/...

Re: Kimi K2.5 Technical Report [pdf]

#146
post #5

Earlier quoted context omitted.

Out of curiosity, what kind of specs do you have (GPU / RAM)? I saw the requirements and it's a beyond my budget so I am "stuck" with smaller Qwen coders.

Note that Kimi K2x is natively 4 bit int, which reduces the memory requirements somewhat.

Here's the citation for that, I think its not in the Technical Report. https://huggingface.co/moonshotai/Kimi-K2.5#4-native-int4-qu...
Post reply on HN