Quazim0t0 (Dean Byrne)

reacted to satgeze's post with 🤗 about 18 hours ago

Post

1360

First GGUF quants of Tencent's Hy3 (299B MoE), built before official llama.cpp support exists.

Hy3 dropped ~30 hours ago with only MLX and MXFP4 quants, both datacenter-sized. So I converted it myself using a community llama.cpp fork that implements the hy_v3 architecture.

What's in the repo:

- IQ1_M (62GB, fits a 128GB MacBook), IQ2_M (90GB), Q2_K (101GB), all with 1M context baked in via YaRN
- IQ quants are importance-matrix: bootstrap style. The static Q2_K ran RAM-resident to compute the imatrix, then IQ1_M and IQ2_M were requantized from the archived f16 with it
- Fixed chat template (the stock one uses .format() calls llama.cpp's Jinja rejects)
- Build instructions for the fork, including the two gotchas that cost me three build attempts

Honesty section, because that is how these repos work: this is EXPERIMENTAL. Not needle-certified yet (1M is baked but unverified, certification ladder will be published either way). MTP layer exists in the checkpoint but no llama.cpp build can run hy_v3 MTP inference yet, so it is not included. Real gate outputs are on the card, misses and all, judge for yourself.

satgeze/Hy3-1M-GGUF

Full quant ladder (Q3 through Q8) is mirroring to ModelScope for bigger hardware.

9 replies

·

replied to their post about 19 hours ago

Really appreciate this, honestly. You took it somewhere I hadn't measured, first that the honest object is a channel mixer built from a growth process not the fungal story, then that fan-out against a matched dense MLP is the number that settles the efficiency side. That got me to actually go measure instead of speculate, which I wouldn't have done otherwise.

Thanks for taking it seriously enough to push. I'll get you the fixed-fan-out numbers from the clustering variant when my current run frees up, since I want to know the answer too.

replied to their post about 19 hours ago

Measured it: the readout weights all 96 tips basically uniformly and the tips spread across ~15/16 density regions, so there's no clustering to exploit yet, a top-k readback today would be a cliff and it'd need a learned gate. Makes sense in hindsight, negative autotropism is an anti-clustering force by design, so the thing that makes it a colony is what keeps it dense.

I'll build the second variant (flip in a clustering/positive-tropism term so the growth concentrates) and post the fixed-fan-out perplexity numbers. I've got another experiment running right now, so it'll be a bit, but I'll get to it when I can since I want to see this one too.

replied to their post about 20 hours ago

Appreciate it, but efficiency wasn't the point, and the post already framed it the way you did: a research LM whose channel-mixing block isn't an MLP, just a few differentiable steps of colony growth read back into the hidden state. The only thing I was testing is whether a weird dynamical system can carry the channel-mixing role and still learn language under a fixed pipeline. "Curiosity or cheaper mixer" is a great question, just a different experiment than the one I ran.

That said, I measured your number: active-channel fan-out (Jacobian participation ratio, /768) was ~469 for the growth block vs ~489 for a matched dense MLP, at 3.86M vs 3.54M params and ~1.07x MACs/token. So you're right, it densifies to roughly MLP cost. The growth math is cheap (O(N·F)), but it's bracketed by dense projections, so it's a hair more expensive, not less. Curiosity, plainly. The version that'd make your point true is a sparsified readback (each tip writing to a local slice of channels, or top-k tip-to-channel routing) so fan-out drops well under d_model. That one I'd like to run. Thanks for pushing on it.

reacted to PhysiQuanty's post with 🧠🚀🔥 1 day ago

Post

2393

🧠 Arithmetic-SLM : A 30M model that manages to compute simple arithmetic better than a 3B model 🚀
WhirlwindAI/Arithmetic-SLM
WhirlwindAI/arithmetic-slm

🏆 Leaderboard ArithMark-2 🏆
🥇 Qwen/Qwen2.5-Math-1.5B = 82.08%
🥈 WhirlwindAI/Arithmetic-SLM = 78.60% (31.7M Params)
🥉 Qwen/Qwen2.5-3B = 78.44%

Example WhirlwindAI/Arithmetic-SLM =
0.5 * 0.5 = 0.25 ✅
105 + 45 / 8 = 110 ✅
(132 / 12) + (46 - 15) = 42 ✅
(10 + 28) * 3 = 114 ✅
1 * (16 + 28) = 44 ✅
(21 + 27) * (14 - 7) = 336 ❌

leaderboard = """
|              Model               |    Params    |   Score   |
|----------------------------------|--------------|-----------|
|      Qwen/Qwen2.5-Math-1.5B      |     1.54B    |   82.08%  |
|    WhirlwindAI/Arithmetic-SLM    |    31.70M    |   78.60%  | <=
|         Qwen/Qwen2.5-3B          |     3.09B    |   78.44%  |
|        Qwen/Qwen2.5-1.5B         |     1.54B    |   77.72%  |
|    Qwen/Qwen2.5-Coder-1.5B       |     1.54B    |   74.88%  |
|   HuggingFaceTB/SmolLM2-1.7B     |     1.71B    |   66.12%  |
|        Qwen/Qwen2.5-0.5B         |      494M    |   63.04%  |
| facebook/MobileLLM-R1-140M-base  |      140M    |   53.88%  |
|     SupraLabs/Supra-50M-Base     |       52M    |   27.12%  |
"""

Bench =
AxiomicLabs/ArithMark-2.0
DataSet =
WhirlwindAI/Arithmetic
By Science AND FOR SCIENCE <3

2 replies

·

reacted to SeaWolf-AI's post with 👍👍👍👍 2 days ago

Post

2986

🚀 Adding a GPU without building one

AI is usually framed as "how smart is the model / how many GPUs did you buy." The real bottleneck is elsewhere — how efficiently you use the GPUs you already have.

Training happens once; inference runs the entire time users use your product. So a service's economics come down to cost per token. Inference acceleration uses software to pull several times more out of the same GPU — the effect of plugging in one more "virtual GPU."

VIDRAFT's VKAE, measured (B200, same-harness, no quality loss):

Qwen3.5-35B-A3B (MoE): 25.7 → 601 tok/s (23.4×)
Darwin-36B-Opus (in-house MoE): 25.0 → 280.8 (11.2×)
10,000+ tok/s peak aggregate under concurrency
The key: it's reproducible — model + serving shipped as one container.

docker pull vidraft/qwen35-vkae:601
Don't take our word for it — run it yourself. The mechanism will be released as a paper.

🏆 Leaderboard & demo 👉 VIDraft/vkae
Articles 👉 https://huggingface.co/blog/FINAL-Bench/vkae-leaderboard

3 replies

·

reacted to Bc-AI's post with 👍 2 days ago

Post

115

Hello everyone! We are Smilyai-labs, a small team of 4 high school students who love AI and find it very interesting! We are super excited to launch our first preview edition of our Nova-Standard model! Nova-1 is a small family of LLM's which we built from scratch! We designed a custom efficient architecture, featuring MoD layers and much more! We have added RoPE, making it perfect to use YaRN for context extrapolation! From internal tests, we have found that it can support 8K reasonably well and 16K with some performance decreases! Please note, that this is only a Beta version for testing, and we are actively working to fix it! We are only a small group of 4 students, so our progress might be a bit slow! Thank you for your patience! Thank you for checking out our work and a small follow would be appreciated! We are constantly updating, and we will roll out a gradio demo ASAP! Our hugging face org is at:

Smilyai-labs . Our model is at: https://huggingface.co/Smilyai-labs/Nova-1-standard-pretrained-preview . The Beta testers org is at:

Smilyai-labs-beta-testers .

reacted to stas's post with 🤗 2 days ago

Post

1635

I present to you a new experimental open book.

https://github.com/stas00/python-cookbook

I took my dense Python cheatsheet that I have been honing for many years and use a lot daily and turned it into a book of recipes.

Is this useful?

This is, of course, free, like other open books.

reacted to ProCreations's post with 🧠 2 days ago

Post

1328

want model think like grug?token efficient think! Grug dataset here. Grug model now ProCreations/grug-think

reacted to salma-remyx's post with 🔥 2 days ago

Post

1464

What's holding your code back?
Outrider finds, implements, and validates methods for your repo.

While testing Outrider on a fork of huggingface/peft, I discovered "Riemannian Preconditioned LoRA for Fine-Tuning Foundation Models" (arxiv: 2402.02347)

The work offers improved stability and faster convergence in LoRA finetuning by adjusting updates for curvature that LoRA optimizers typically ignore.

Not the most recent paper, so I was pleasantly surprised my action surfaced this method as a candidate before implementing a PR. Even more surprised this method had not already been merged upstream.

Turns out, the author did try contributing to peft a couple years ago, but people get busy and the PR was closed after going stale.

So I decided to revive it! I opened an issue and soon after the author engaged to help land the feature. Now huggingface/peft #3382 is open, a joint effort with the paper's author.

This whole episode has me thinking about the future of OSS maintenance with AI coding. The software projects which endure will be well-shaped to quickly land and help test new ideas.

Across 30 forks, I've seen several papers land as clean PRs for multiple repos, which offers a perspective on how methods impact applications. Recent methods matching multiple frameworks: STARE, Entity Binding, BINEVAL

Get Outrider: https://github.com/remyxai/outrider

reacted to aufklarer's post with 🔥 2 days ago

Post

1308

Voice cloning models measured across five languages: OmniVoice, Chatterbox, VoxCPM2, Fish Audio

I published a new Soniqo benchmark post for local voice cloning models across five languages:

https://www.soniqo.audio/blog/voice-cloning-benchmarks

Models:

- OmniVoice int8
- Chatterbox Multilingual fp16
- VoxCPM2 bf16
- Fish Audio S2 Pro fp16

Languages:

- English
- German
- Modern Standard Arabic
- Spanish
- Mandarin Chinese

The benchmark uses Google FLEURS test clips as dataset references. Each row includes the reference audio, generated audio, speaker similarity, WER/CER, generated audio length, and RTF.

Main result in this run: OmniVoice was the strongest all-around row set, with 0.707 mean speaker cosine across all five languages, 0.0% ASR error, and mean RTF 0.45. VoxCPM2 bf16 was especially strong on Arabic speaker match. Fish Audio S2 Pro showed strong German/Arabic similarity but slower RTF. Chatterbox Multilingual was competitive on Arabic and Spanish.

This is an engineering benchmark, not a human MOS study. The speaker-similarity values should be compared within this table because every row uses the same local speaker-embedding pipeline.

Try the stack locally with Speech Studio:

https://www.soniqo.audio/speech-studio
https://github.com/soniqo/speech-studio

Underlying Swift library/CLI:

https://github.com/soniqo/speech-swift

Soniqo models and exports:

soniqo @aufklarer

What model or language should I add next?

reacted to breitburg's post with 🔥 2 days ago

Post

1218

I've been experimenting with "pure" model alignment.

The core idea is to only train a verifiable version of a capacity until the model generalizes it to the non-verifiable version. For example, training the model on factual self-knowledge, like the model's scale, architecture, runtime situation, and being able to predict its own behavior, betting this generalizes to real introspection about states that do not.

The same principle applies to general instruction following -- no training on subjective judgement, only verifiable claims and inferences, betting the skill generalizes to instructions where correctness is a matter of judgment.

The primary alignment claim is that an identity and taste that will emerge this way will be much more robust and honest than hand-scripted ones (e.g.
"As an AI language model...").

During the training, we should never teach it to make any subjective claims or invent experiences that we assume it has, like "I don't have taste" or "I'm not self-aware in the way you think", as well as no narration of internal states like "I'm curious now".

The main threat, of course, is that we'll simply inherit the training distribution of all the things like "taste", and we'll get an average. However, with the recent research about the models' introspection abilities, it might be as well the case that we'll get something that's more honest than something that tries to adhere to a specific spec file.

I'm posting new experimental models trained that way in this collection: https://huggingface.co/collections/breitburg/neue

3 replies

·

reacted to SeaWolf-AI's post with 🔥 2 days ago

Post

2986

🚀 Adding a GPU without building one

AI is usually framed as "how smart is the model / how many GPUs did you buy." The real bottleneck is elsewhere — how efficiently you use the GPUs you already have.

Training happens once; inference runs the entire time users use your product. So a service's economics come down to cost per token. Inference acceleration uses software to pull several times more out of the same GPU — the effect of plugging in one more "virtual GPU."

VIDRAFT's VKAE, measured (B200, same-harness, no quality loss):

Qwen3.5-35B-A3B (MoE): 25.7 → 601 tok/s (23.4×)
Darwin-36B-Opus (in-house MoE): 25.0 → 280.8 (11.2×)
10,000+ tok/s peak aggregate under concurrency
The key: it's reproducible — model + serving shipped as one container.

docker pull vidraft/qwen35-vkae:601
Don't take our word for it — run it yourself. The mechanism will be released as a paper.

🏆 Leaderboard & demo 👉 VIDraft/vkae
Articles 👉 https://huggingface.co/blog/FINAL-Bench/vkae-leaderboard

3 replies

·

reacted to RDTvlokip's post with 🚀 2 days ago

Post

1982

I finally changed the architecture of my 15M French LLM. It worked. Then I almost fooled myself about how much and catching that was the real win.

After proving last time that architecture is a threshold, not a lever, I got stubborn: could I change how the model learns? Four honest attempts, Lion, a sharper AdamW β2, multi-token prediction, LayerScale. Four failures. The bottleneck wasn't the learning rule either.

So I changed the shape of the computation instead: loop the same transformer blocks 4×, deeper reasoning, zero added parameters. It beat the baseline on perplexity, the first thing in the whole project to move that number. Then I added my own twist: let each token decide how deep to think, halting on its own entropy.

My first evaluation was spectacular. Coherence up 65%. Hallucinated names down 62%.

It was noise.

Eight prompts, one seed. I re-ran on 50 prompts × 200 tokens and watched the gains shrink to "modest" and on out-of-domain prompts, recurrence actually made things worse. No universal winner. And none of it is new: it's Adaptive Computation Time (2016), the Universal Transformer (2018), and LoopViT (2026), recombined and measured honestly.

The real lesson:

A number from 8 prompts is a rumor. The eval harness that kills your own best result is worth more than the result it kills. Cite your lineage. Stay preliminary until multiple seeds say otherwise.

The three models are live. The write-up is honest about every caveat 👇

🔗 https://huggingface.co/blog/RDTvlokip/teaching-a-15m-french-llm-to-think-deeper

6 replies

·

reacted to fffiloni's post with 🚀 2 days ago

Post

1702

You can explore and try my Agentic Space Factory:
fffiloni/agentic-space-factory-UI-public

Be kind 🤗

Dean Byrne PRO

AI & ML interests

Recent Activity

Organizations

Dean Byrne PRO

AI & ML interests

Recent Activity

Organizations

Quazim0t0's activity