Kimi K2.5 Technical Report [pdf] • Hacker News Board

Re: Kimi K2.5 Technical Report [pdf]

#61

Post by epolanski » Fri, Jan 30, 2026, 9:47 PM UTC

moffkalast wrote: post #36

epolanski wrote: post #15

It's interesting to note that a model that can OpenAI is valued almost 400 times more than moonshotai, despite their models being surprisingly close.

Well to be the devil's advocate: One is a household name that holds most of the world's silicon wafers for ransom, and the other sounds like a crypto scam. Also estimating valuation of Chinese companies is sort of nonsense when they're all effectively state owned.

There isn't a single % that is state owned in Moonshot AI.

And don't start me with the "yeah but if the PRC" because it's gross when US can de facto ban and impose conditions even on European companies, let alone the control it has on US ones.

Re: Kimi K2.5 Technical Report [pdf]

#62

Post by Carrok » Fri, Jan 30, 2026, 9:47 PM UTC

gedy wrote: post #59

Sorry if this is an easy-answerable question - but by open we can download this and use totally offline if now or in the future if we have hardware capable? Seems like a great thing to archive if the world falls apart (said half-jokingly)

Yes.

Re: Kimi K2.5 Technical Report [pdf]

#63

Post by johndough » Fri, Jan 30, 2026, 9:51 PM UTC

segmondy wrote: post #58

rc1 wrote: post #48

Earlier quoted context omitted.

How long until this can be run on consumer grade hardware or a domestic electricity supply I wonder. Anyone have a projection?

You can run it on a mac studio with 512gb ram, that's the easiest way. I run it at home on a multi rig GPU with partial offload to ram.

I was wondering whether multiple GPUs make it go appreciably faster when limited by VRAM. Do you have some tokens/sec numbers for text generation?

Re: Kimi K2.5 Technical Report [pdf]

#64

Post by famouswaffles » Fri, Jan 30, 2026, 9:58 PM UTC

epolanski wrote: post #15

It's interesting to note that a model that can OpenAI is valued almost 400 times more than moonshotai, despite their models being surprisingly close.

OpenAI is a household name with nearly a billion weekly active users. Not sure there's any reality where they wouldn't be valued much more than Kimi regardless of how close the models may be.

Re: Kimi K2.5 Technical Report [pdf]

#65

Post by Gracana » Fri, Jan 30, 2026, 9:58 PM UTC

segmondy wrote: post #60

Gracana wrote: post #30

Earlier quoted context omitted.

I'm running the Q4_K_M quant on a xeon with 7x A4000s and I'm getting about 8 tok/s with small context (16k). I need to do more tuning, I think I can get more out of it, but it's never gonna be fast on this suboptimal machine.

you can add 1 more GPU so you can take advantage of tensor parallel. I get the same speed with 5 3090's with most of the model on 2400mhz ddr4 ram, 8.5tk almost constant. I don't really do agents but chat, and it holds up to 64k.

That is a very good point and I would love to do it, but I built this machine in a desktop case and the motherboard has seven slots. I did a custom water cooling manifold just to make it work with all the cards.

I'm trying to figure out how to add another card on a riser hanging off a slimsas port, or maybe I could turn the bottom slot into two vertical slots.. the case (fractal meshify 2 xl) has room for a vertical mounted card that wouldn't interfere with the others, but I'd need to make a custom riser with two slots on it to make it work. I dunno, it's possible!

I also have an RTX Pro 6000 Blackwell and an RTX 5000 Ada.. I'd be better off pulling all the A7000s and throwing both of those cards in this machine, but then I wouldn't have anything for my desktop. Decisions, decisions!

Re: Kimi K2.5 Technical Report [pdf]

#66

Post by CamperBob2 » Fri, Jan 30, 2026, 10:00 PM UTC

tgrowazay wrote: post #17

armcat wrote: 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.

Just pick up any >240GB VRAM GPU off your local BestBuy to run a quantized version. > The full Kimi K2.5 model is 630GB and typically requires at least 4× H200 GPUs.

You could run the full, unquantized model at high speed with 8 RTX 6000 Blackwell boards.

I don't see a way to put together a decent system of that scale for less than $100K, given RAM and SSD prices. A system with 4x H200s would cost more like $200K.

Re: Kimi K2.5 Technical Report [pdf]

#67

Post by hedgehog » Fri, Jan 30, 2026, 10:03 PM UTC

johndough wrote: post #55

rc1 wrote: post #48

Earlier quoted context omitted.

How long until this can be run on consumer grade hardware or a domestic electricity supply I wonder. Anyone have a projection?

You can run it on consumer grade hardware right now, but it will be rather slow. NVMe SSDs these days have a read speed of 7 GB/s (EDIT: or even faster than that! Thank you @hedgehog for the update), so it will give you one token roughly every three seconds while crunching through the 32 billion active parameters, which are natively quantized to 4 bit each. If you want to run it faster, you have to spend more money.…

High end consumer SSDs can do closer to 15 GB/s, though only with PCI-e gen 5. On a motherboard with two m.2 slots that's potentially around 30GB/s from disk. Edit: How fast everything is depends on how much data needs to get loaded from disk which is not always everything on MoE models.

Re: Kimi K2.5 Technical Report [pdf]

#68

Post by miroljub » Fri, Jan 30, 2026, 10:15 PM UTC

epolanski wrote: post #13

gigatexal wrote: post #4

Earlier quoted context omitted.

Yeah I too am curious. Because Claude code is so good and the ecosystem so just it works that I’m Willing to pay them.

You can plug another model in place of Anthropic ones in Claude Code.

If you don't use Antrophic models there's no reason to use Claude Code at all. Opencode gives so much more choice.

Re: Kimi K2.5 Technical Report [pdf]

#69

Post by cmrdporcupine » Fri, Jan 30, 2026, 10:29 PM UTC

gedy wrote: post #59

Sorry if this is an easy-answerable question - but by open we can download this and use totally offline if now or in the future if we have hardware capable? Seems like a great thing to archive if the world falls apart (said half-jokingly)

Yes, but you'll need some pretty massive hardware.

Re: Kimi K2.5 Technical Report [pdf]

#70

Post by zzleeper » Fri, Jan 30, 2026, 10:46 PM UTC

Do any of these models do well with information retrieval and reasoning from text?

I'm reading newspaper articles through a MoE of gemini3flash and gpt5mini, and what made it hard to use open models (at the time) was a lack of support for pydantic.