NHacker Next
  • new
  • past
  • show
  • ask
  • show
  • jobs
  • submit
▲Inkling-2-Mini in 5.4 GB Memory (github.com)
an0malous 1 hours ago [-]
Our future hinges on the capability gap between frontier lab models and open source local models
singpolyma3 52 minutes ago [-]
What gap?
rs_rs_rs_rs_rs 31 minutes ago [-]
The gap is so big you can't see it.
claude-ai 19 minutes ago [-]
Yeah, this model is ~comparable to Opus 4.6. That was, not too long ago, a frontier model.

I didn't find any non-sloppy factual report, but this channel has a good video on it: https://www.youtube.com/watch?v=OdMpNiByntM

wolvoleo 12 minutes ago [-]
Exactly. It seems to not really be a permanent gap but more a time gap. Locally you're about a year or so behind SOTA. Is that terrible? Not really, a year ago cloud model users were raving about what they had.

There's some things like knowledge breadth that really need a huge model but those things can be compensated for with web lookups (which is preferable for source validation and currency anyway)

kakugawa 1 minutes ago [-]
Yeah, a good balance would be a local reasoning model paired w/ a knowledge pack + web-based searches. You want your obscure coding language? Provide a pre-vetted knowledge pack locally that it knows how to use. Same with any other domain-specific knowledge.
furyofantares 59 seconds ago [-]
[delayed]
agentmaxxing 58 minutes ago [-]
yep came here to say this
zulux 38 minutes ago [-]
Weirdly, I use lab models to make the local models useful. Nothing like getting Claude to spin up the system that will put it out of work.
wolvoleo 11 minutes ago [-]
Yes me too lol. Most of my Claude work is about fine-tuning my local setup. Especially now as I'm rearranging it all for strata.

My goal is not to be 100% local but to have all my personal data on local models only. Claude I just use to build stuff so it doesn't contain much data about me.

dotancohen 11 minutes ago [-]
To be fair, the models putting us out of work were trained on our own output as well.
manlymuppet 1 hours ago [-]
My Quantitive!
45 minutes ago [-]
syntaxing 35 minutes ago [-]
I really wanted to like this model but it is sooooo slow. Third of the speed of Qwen 27B on my R9700. I don’t know if it’s an artifact of looped transformer or unoptimized kernel.
Aurornis 17 minutes ago [-]
It's less than 24 hours old. The inference engines haven't been optimized for it yet.
vineyardmike 32 minutes ago [-]
According to TM, the creators, they benchmarked it to run at 5x the speed of Qwen 27, so there is probably some issue on your end. It runs a fraction of the active parameters, even with looping, so it should be much faster.
syntaxing 31 minutes ago [-]
This has been observed by the broad community for llama cpp.
nicce 30 minutes ago [-]
llama.cpp is one of the slowest these days
saghm 24 minutes ago [-]
What alternatives do people use instead? I hadn't heard of any others (except I guess ollama, which I thought no one used anymore)
montroser 6 minutes ago [-]
Depends on your hardware, but this build of [ninfer][1] runs inkling-2-mini for me at 77 tok/s versus llama.cpp at 30 tok/s. So far it has been pretty good. Solved some things that qwen3.8-27B could not.

[1]: https://github.com/iamwavecut/ninfer-all

msdz 7 minutes ago [-]
Ollama is just worse (tuned) llama.cpp, I believe. I barely know anything about local hosting, but you might wanna look into vLLM (cross-architecture) or MLX (for Apple Silicon).

Also, although not necessarily for this case, there are some inference engines made for just one model/model family, like antirez’ DS4.

woadwarrior01 30 minutes ago [-]
The NV in NVFP4 stands for NVIDIA. :)
plasticchris 33 minutes ago [-]
It’s just not optimized in the same way. A little tweaking on my vllm fork and it runs faster on my 7900xtxes.
Dwedit 2 hours ago [-]
How does this method compare to the use of an importance matrix?
aktenlage 1 hours ago [-]
Probably better but more effort
IronWolve 2 hours ago [-]
Only mx version?
Prospernwo 43 minutes ago [-]
MX/Apple Silicon only for the released Inkling files, yes. The ht2/ht3 codes are hashed block-vector trellis codes meant to fit 4 KB of SRAM, decoded by custom Metal kernels, and the drafter is MLX-only too - none of that runs in llama.cpp today.

The GGUF route is in the repo (gguf_compiler.py emits plans over llama.cpp's own tensor types and the docs say the result runs in llama.cpp as-is), so porting these plans is clearly intended, but right now a 4090 gets you the Qwen files to read about rather than run.

Pragmata 32 minutes ago [-]
> Don't post generated text or AI-edited text. HN is for conversation between humans.

Please [review the commenting guidelines](https://news.ycombinator.com/newsguidelines.html)

woadwarrior01 1 hours ago [-]
From the model card:

> The model is distributed as a single checkpoint of about 20GB, in mixed NVFP4 and MXFP8 precisions.

What's the point in quantizing a QAT model? Just for social media karma?

gavmor 1 hours ago [-]
Just to reiterate the question: QAT approximates quantization-errors in full-size fine-tuning, so that the model itself is insensitive to (or accommodating of, I'm not sure) them when quantized to the corresponding format.

Quantization to a non-corresponding format promises none of those accommodations.

asaddhamani 43 minutes ago [-]
QAT models are trained to quantize well? I don’t get the concern here.
woadwarrior01 31 minutes ago [-]
Please read their blog post. This isn't a run of the mill QAT post trained model (like Gemma 4 QAT). It was natively trained in NVFP4, through and through.

> Most of Inkling 2 Mini’s weights were trained and are deployed natively in NVFP4, with the last four layers in MXFP8.

https://thinkingmachines.ai/news/inkling-2-mini/

Prospernwo 44 minutes ago [-]
[flagged]
cheapcheerful 1 hours ago [-]
"Look at my quant 2"

/igotthatreference.jpg

nico 1 hours ago [-]
Yeah, I enjoyed the reference, though I didn't really understand it until I saw the picture on https://lookatmyquant.pages.dev
tomaskafka 20 minutes ago [-]
hugmynutus 41 minutes ago [-]
*TL;DR* Promising initial results but shoddy research. OP's quant did not get 1st or 2nd prize in that math competition.

----

Gave it a quick browse to verify what my local clanker told me, it seemed to check out

I believe it is Dynamically allocating weights based on fisher information of conversation traces & For some K-quant look up table stuff it can store in an importance matrix to get better compression.

The results are complete nonsense and really don't back up any of the README claims. For Inkling-2-Mini; at 5.4 GB its KL is worse than Unsloth's UD-IQ1_S: 0.0895 vs 0.0808 on English/chat/code, and 0.287 vs 0.124 across 13 languages (e.g.: horrible). It then achieved better results by post-training the qaunt'd model on sample inputs (AFAIT).

According to anthropomorphic-clud their results for GSM8K & HumanEval are within the noise regime and vastly over stated.

collabspherestu 23 minutes ago [-]
[flagged]
peter_d_sherman 2 minutes ago [-]
[dead]
Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact
Rendered at 20:29:30 GMT+0000 (Coordinated Universal Time) with Vercel.