The model is downloading right now. It should take a couple of hours and it will be done
R PRO
juiceb0xc0de
AI & ML interests
Mechanistic interpretation
SAE weight training
Novel architecture exploration
Brain atlas builds
Recent Activity
updated a dataset 40 minutes ago
juiceb0xc0de/qwenpause published a dataset about 1 hour ago
juiceb0xc0de/qwenpause repliedto their post about 6 hours ago
HF community! Thank you so much the overwhelming flood of downloads I've received on my model atlases over the last couple of weeks. It makes me so happy to know other people out there are enjoying my work. With that being said, I am looking for an opportunity from the builders and trainers out there.
I've recently had the chance to do a mid pre training snapshot durin process and am looking for the opportunity to get some more of these images . This is a symbiotic trade. I am more than happy to do full reports on the status of your model and translate what your model may stand to gain from as you head towards the finish line. The atlas building process uses 8965 different prompt over a wide variety of behavioural features. I combines many different methods of interpreting what is going on inside the model between prompt and response that we are unable to see.
For anyone else interested in neural network imaging and mechanistic interpretation check out my library of atlases and my new work translating the numbers into a 3d visual format to accompany the sqlite database. If you've gpt a model that you wish to have atlased leave a commen and I'kk be sure to ge arounf to it as soon as possible.
https://huggingface.co/collections/juiceb0xc0de/cloud-atlasesOrganizations
replied to their post about 6 hours ago
replied to their post about 12 hours ago
Ya I got you. Give me a bit to set up up here and I'll get it started right away.
replied to danielhanchen's post 8 days ago
Hey just a heads up you split your unsl/oth link accidentally. ๐ฆฅ
reacted to danielhanchen's post with ๐ 8 days ago
Post
4734
Introducing Unsloth Desktop ๐ฆฅ
The first desktop app to run and train models locally.
โข Open-source. Runs on Mac, Windows and Linux
โข Supports MLX, diffusion image/video, audio, GGUF
โข Connect Claude Code and Codex to local LLMs
โข 50% more accurate, self-healing tool calls + sandboxed code exec
โข Works for CPU + multiGPU setups - NVIDIA, AMD, Intel, Mac
โข Train models 2ร faster with 70% less VRAM
โข Private web search, deep research, RAG, MCP and exports (NVFP4, GGUF)
โข Use Unslothโs OpenAI-compatible API and cloud models
โข Securely deploy LLMs remotely and access anywhere
Unsloth Desktop is now available on http://unsloth.ai
and GitHub.
GitHub: https://github.com/unslothai/unsloth
Blog and Guide: https://unsloth.ai/docs/desktop
The first desktop app to run and train models locally.
โข Open-source. Runs on Mac, Windows and Linux
โข Supports MLX, diffusion image/video, audio, GGUF
โข Connect Claude Code and Codex to local LLMs
โข 50% more accurate, self-healing tool calls + sandboxed code exec
โข Works for CPU + multiGPU setups - NVIDIA, AMD, Intel, Mac
โข Train models 2ร faster with 70% less VRAM
โข Private web search, deep research, RAG, MCP and exports (NVFP4, GGUF)
โข Use Unslothโs OpenAI-compatible API and cloud models
โข Securely deploy LLMs remotely and access anywhere
Unsloth Desktop is now available on http://unsloth.ai
and GitHub.
GitHub: https://github.com/unslothai/unsloth
Blog and Guide: https://unsloth.ai/docs/desktop
- 6 replies
reacted to sergiopaniego's post with ๐ 14 days ago
Post
2591
LFM2.5-2.6B just dropped!
and the @liquidai blog comes with some nice details about the training procedure, so let's analyze it.
basically, a full agent training pipeline but compressed into 2.6B
base model โ SFT โ specialized teachers per domain (SFT + RLVR) โ on-policy distillation back into one student โ agentic RL
the two most interesting stages
โ MOPD: the student generates, each prompt routes to its domain teacher for token-level feedback. teachers branch from the same SFT checkpoint, so their signal stays close to the student's distribution
โ agentic RL: multi-turn GRPO inside real harnesses (OpenClaw, Hermes Agent), one sandbox per rollout, a proxy captures token-level trajectories while the harness stays a black box
this makes a 2.6B that beats much larger models on instruction following and tool use
SFT, distillation, RL, RL envs: exactly what we're covering in our Training Agents livestream series (next one coming soon!)
โ model: LiquidAI/LFM2.5-2.6B
โ blog: https://www.liquid.ai/blog/lfm2-5-2-6b
โ live series: https://www.youtube.com/playlist?list=PLo2EIpI_JMQvQZm-kVlz4wY1vWF0LBcf5
and the @liquidai blog comes with some nice details about the training procedure, so let's analyze it.
basically, a full agent training pipeline but compressed into 2.6B
base model โ SFT โ specialized teachers per domain (SFT + RLVR) โ on-policy distillation back into one student โ agentic RL
the two most interesting stages
โ MOPD: the student generates, each prompt routes to its domain teacher for token-level feedback. teachers branch from the same SFT checkpoint, so their signal stays close to the student's distribution
โ agentic RL: multi-turn GRPO inside real harnesses (OpenClaw, Hermes Agent), one sandbox per rollout, a proxy captures token-level trajectories while the harness stays a black box
this makes a 2.6B that beats much larger models on instruction following and tool use
SFT, distillation, RL, RL envs: exactly what we're covering in our Training Agents livestream series (next one coming soon!)
โ model: LiquidAI/LFM2.5-2.6B
โ blog: https://www.liquid.ai/blog/lfm2-5-2-6b
โ live series: https://www.youtube.com/playlist?list=PLo2EIpI_JMQvQZm-kVlz4wY1vWF0LBcf5
- 3 replies
replied to sergiopaniego's post 14 days ago
Yo convolution layers... You caused me to pay attention hahaha get it. An AI model would laugh at this joke and say "That joke is doing more work than the 22 conv layers, this is a pun on the causal convolution layers." Ahhh dad jokes aside I'm actually kinda stoked to run a little atlas on your model. It's like 22 people sitting at a bar eavesdropping on a few people adjacent to the left of them and predicting based on that narrow view point with 8 patrons listening to the entire thing with clarity and scratching their heads. Ima see what saving KV cache is doing here. To hell with the benchmarks and arXiv papers! Let's weigh it measure it and properly fit it in a suit. I wanna see if it looks properly proportioned or if it's a saw toothed oddball with wide shoulders a narrow midriff big ol' booty tiny calves huge shins and stiletto sized feet.
reacted to Bc-AI's post with ๐ 18 days ago
Post
177
Hello Everyone! Bc-AI here from Smilyai-labs! Today we have done our latest update for CodVa-1-Small. It is very powerful for coding, and benchmark results will come soon. However, it is NOT good for other tasks, with high hallucination rates. We will perform RLHF and DPO very soon!
reacted to badaoui's post with ๐ 25 days ago
Post
2368
432 GB of ultra-fast HBM4 and up to 23.3 TB/s of memory bandwidth on a single GPU ๐คฏ.
Two weeks ago, we got early access to AMD's new Instinct MI455X, and our first goal was simple: make sure ๐ค Transformers works on day one.
Over the past few weeks, we worked closely with the AMD team to validate the platform, enable Flash Attention, add torchcodec support for multimodal models, and resolve issues uncovered during testing.
The result:
โ 99.5% success rate across our 24 core Transformers model architectures - already on par with our daily CI on previous AMD and NVIDIA platforms.
The hardware is just as exciting. With 432 GB of HBM per GPU, our early capacity experiments showed more than 3ร the concurrent long-context requests compared to MI300, thanks to the much larger KV cache capacity.
A huge thanks to the AMD team for the early access and the great collaboration!
Read the full blog ๐
https://huggingface.co/blog/badaoui/transformers-on-amd-mi455
Two weeks ago, we got early access to AMD's new Instinct MI455X, and our first goal was simple: make sure ๐ค Transformers works on day one.
Over the past few weeks, we worked closely with the AMD team to validate the platform, enable Flash Attention, add torchcodec support for multimodal models, and resolve issues uncovered during testing.
The result:
โ 99.5% success rate across our 24 core Transformers model architectures - already on par with our daily CI on previous AMD and NVIDIA platforms.
The hardware is just as exciting. With 432 GB of HBM per GPU, our early capacity experiments showed more than 3ร the concurrent long-context requests compared to MI300, thanks to the much larger KV cache capacity.
A huge thanks to the AMD team for the early access and the great collaboration!
Read the full blog ๐
https://huggingface.co/blog/badaoui/transformers-on-amd-mi455
- 1 reply
replied to kingabzpro's post about 1 month ago
Hey! If you're looking for a more in depth break down regarding pricing or any of the other features on Hugging Face the Docs section is a great resource as well. The Hugging Chat model is free to use for queries on a lot more than just the platform. You can find everything from optimizing inference speed to working with newer architectures like Mamba, or getting started with a Reachy Mini. Hope that helps!
reacted to danielhanchen's post with ๐ about 1 month ago
Post
6213
Gemma 4 is now faster and much more accurate! ๐
Google made huge improvements to tool-calling and chat accuracy, reliability + speed.
To get fixes, re-download our updated GGUF, MLX, NVFP4 quants!
Unsloth quants: https://huggingface.co/collections/unsloth/gemma-4
Gemma 4 Guide: https://unsloth.ai/docs/models/gemma-4
Google made huge improvements to tool-calling and chat accuracy, reliability + speed.
To get fixes, re-download our updated GGUF, MLX, NVFP4 quants!
Unsloth quants: https://huggingface.co/collections/unsloth/gemma-4
Gemma 4 Guide: https://unsloth.ai/docs/models/gemma-4
- 8 replies
reacted to Hari5115's post with ๐ about 1 month ago
Post
240
Fair warning space/photograhy lovers: this is weirdly addictive, Just one date becomes thirty !
There's a spot on the internet that shows you the exact image nasa shared with the world on any day since 1995. Type your birthday and see what the universe was up to.
(Born before 1995? Ah โ the cosmos wasn't posting yet. You're officially more vintage than the dataset. ๐ Try a big life date instead.) ๐
๐ญ See yours โ Hari5115/cosmic-moment
Or hit Surprise me and let the universe pick.
Post the image you got below ๐ โ let's see whose day space showed off for. ๐
Built on the open nasa apod dataset ยท public domain ยท not affiliated with nasa
Dataset: ๐ฆ Hari5115/nasa-apod
#space #photography #astro #cosmic
There's a spot on the internet that shows you the exact image nasa shared with the world on any day since 1995. Type your birthday and see what the universe was up to.
(Born before 1995? Ah โ the cosmos wasn't posting yet. You're officially more vintage than the dataset. ๐ Try a big life date instead.) ๐
๐ญ See yours โ Hari5115/cosmic-moment
Or hit Surprise me and let the universe pick.
Post the image you got below ๐ โ let's see whose day space showed off for. ๐
Built on the open nasa apod dataset ยท public domain ยท not affiliated with nasa
Dataset: ๐ฆ Hari5115/nasa-apod
#space #photography #astro #cosmic
reacted to danielhanchen's post with ๐ about 1 month ago
Post
4962
Weโre releasing Gemma 4 NVFP4 quants that run 1.5ร faster on your GPU.
Gemma-4-12B NVFP4 works on 11GB VRAM.
26B-A4B hits 13K tok/s (B200).
Unsloth NVFP4 enables faster, more accurate 4-bit Blackwell inference.
Blog: https://unsloth.ai/docs/basics/nvfp4
Gemma NVFP4: https://huggingface.co/collections/unsloth/nvfp4
Gemma-4-12B NVFP4 works on 11GB VRAM.
26B-A4B hits 13K tok/s (B200).
Unsloth NVFP4 enables faster, more accurate 4-bit Blackwell inference.
Blog: https://unsloth.ai/docs/basics/nvfp4
Gemma NVFP4: https://huggingface.co/collections/unsloth/nvfp4
- 3 replies
reacted to Quazim0t0's post with ๐คฏ about 1 month ago
Post
4285
I never really posted about my DaisyChain project because it's still work in progress. I decided to post a small bit about it and the demo.
DaisyChain Genomics: four small DNA/RNA specialists chained behind a learned router that behave like one big genomics model, at ~7ร less active compute.
I built a modular genomics model chasing a 500M-parameter foundation model, then caught myself measuring it wrong. Here's the honest version.
DaisyChain is a different bet: instead of one monolithic DNA model, it's four ~74M specialists (eukaryote, prokaryote, mRNA, splice) chained behind a learned router, each distilled per-domain from HuggingFaceBio's Carbon-500M. Every specialist reports how surprised it is (bits/base) and the router hands each sequence to the link most at home with it. In lineage it's a cluster Branch-Train-Merge mixture of experts, so you can chain on a new domain without retraining the others.
The pitch: ~295M total params (under Carbon-500M), but only one ~74M specialist runs per query, so ~7ร cheaper per token, routing at 100% held-out.
The mistake: Carbon works in 6-mers, and I'd been scoring likelihood as 6-mer cross-entropy. By that number I was +0.043 bits/base behind, splice even "beating" Carbon. But Carbon scores at the base-pair level, which is harder and more honest. Re-run their way:
Real gap: 1.862 vs 1.787 bits/base, +0.089 behind, not +0.043
No domain actually beats Carbon; the "splice win" was an artifact
Seq recovery: euk 31.5% vs 38.9%, bacteria 40.9% vs 54.1%
DaisyChain is still behind Carbon-500M (itself a draft model, not built to top benchmarks), but by a number I can defend, and the gap closes with every per-domain pass. ๐ผ
DaisyChainAI
DaisyChainAI/Daisychain-Genomics-Demo
DaisyChainAI/daisychain-genomics
DaisyChain Genomics: four small DNA/RNA specialists chained behind a learned router that behave like one big genomics model, at ~7ร less active compute.
I built a modular genomics model chasing a 500M-parameter foundation model, then caught myself measuring it wrong. Here's the honest version.
DaisyChain is a different bet: instead of one monolithic DNA model, it's four ~74M specialists (eukaryote, prokaryote, mRNA, splice) chained behind a learned router, each distilled per-domain from HuggingFaceBio's Carbon-500M. Every specialist reports how surprised it is (bits/base) and the router hands each sequence to the link most at home with it. In lineage it's a cluster Branch-Train-Merge mixture of experts, so you can chain on a new domain without retraining the others.
The pitch: ~295M total params (under Carbon-500M), but only one ~74M specialist runs per query, so ~7ร cheaper per token, routing at 100% held-out.
The mistake: Carbon works in 6-mers, and I'd been scoring likelihood as 6-mer cross-entropy. By that number I was +0.043 bits/base behind, splice even "beating" Carbon. But Carbon scores at the base-pair level, which is harder and more honest. Re-run their way:
Real gap: 1.862 vs 1.787 bits/base, +0.089 behind, not +0.043
No domain actually beats Carbon; the "splice win" was an artifact
Seq recovery: euk 31.5% vs 38.9%, bacteria 40.9% vs 54.1%
DaisyChain is still behind Carbon-500M (itself a draft model, not built to top benchmarks), but by a number I can defend, and the gap closes with every per-domain pass. ๐ผ
DaisyChainAI/Daisychain-Genomics-Demo
DaisyChainAI/daisychain-genomics
reacted to Quazim0t0's post with ๐ค about 1 month ago
Post
5427
Disabling Gated Access for some of my models today. I will update this post with the list as I go. I had to go back recently and make updates to a lot of models and bit off more than I could chew with managing all the releases. I didn't realize many of you asked for access and I apologize for not accepting your access to the models you were wanting to look at. I don't like to release something fully unless I feel I put what I could into it for the moment. Some models will remain on gated access, but I will now be accepting those who request to view the repo.
Disabled Gated Access:
Quazim0t0/Byrne-VLM-131M - v2 Updates + Training Instructions
Quazim0t0/Byrne-Speech - 12M Tiny Speech model
Quazim0t0/Byrne-ASR-English - 12M Tiny ASR Model
Quazim0t0/Byrne-VE - Byrne-VE โ Tiny Self-Distilled Vision Encoder (39M)
Quazim0t0/Positronic-144M - Research Artifact
Quazim0t0/SpikeWhale-SNN-216M - Research Artifact
Quazim0t0/Mycel-LM-79M - Research Artifact
Quazim0t0/Chimera-64M - Research Artifact
Accepting Gated Access Requests (7/9):
https://huggingface.co/Quazim0t0/Wheeler-63M
Also uploaded my Neural Photonic Project:
Three trained nets in series: light interferes through the MZI2.pt optical core (verified 256/256), is measured by the PD.pt neural photodetector (verified 1024/1024), and folded into a single OUTPUT byte by the real ADC8 neural-CPU adder. Every value below is computed end-to-end by the three loaded, verified nets โ no analytic formulas.
Demo: https://quazim0t0-neural-photonic-hybrid.hf.space/
Model Weights: Quazim0t0/neural-photonic
AND!
A work in progress:
Ashen Depths
https://quazim0t0-ashendepths.static.hf.space/index.html
Disabled Gated Access:
Quazim0t0/Byrne-VLM-131M - v2 Updates + Training Instructions
Quazim0t0/Byrne-Speech - 12M Tiny Speech model
Quazim0t0/Byrne-ASR-English - 12M Tiny ASR Model
Quazim0t0/Byrne-VE - Byrne-VE โ Tiny Self-Distilled Vision Encoder (39M)
Quazim0t0/Positronic-144M - Research Artifact
Quazim0t0/SpikeWhale-SNN-216M - Research Artifact
Quazim0t0/Mycel-LM-79M - Research Artifact
Quazim0t0/Chimera-64M - Research Artifact
Accepting Gated Access Requests (7/9):
https://huggingface.co/Quazim0t0/Wheeler-63M
Also uploaded my Neural Photonic Project:
Three trained nets in series: light interferes through the MZI2.pt optical core (verified 256/256), is measured by the PD.pt neural photodetector (verified 1024/1024), and folded into a single OUTPUT byte by the real ADC8 neural-CPU adder. Every value below is computed end-to-end by the three loaded, verified nets โ no analytic formulas.
Demo: https://quazim0t0-neural-photonic-hybrid.hf.space/
Model Weights: Quazim0t0/neural-photonic
AND!
A work in progress:
Ashen Depths
https://quazim0t0-ashendepths.static.hf.space/index.html
reacted to danielhanchen's post with ๐ about 1 month ago
Post
6430
DeepSeek-V4 can now run locally with Unsloth GGUFs! ๐ณ
Run lossless DeepSeek-V4-Flash on 168GB RAM or
3-bit works on 110GB Mac, RAM, VRAM setups.
Run via Unsloth Studio or llama.cpp.
GGUF: unsloth/DeepSeek-V4-Flash-GGUF
Guide: https://unsloth.ai/docs/models/deepseek-v4
Run lossless DeepSeek-V4-Flash on 168GB RAM or
3-bit works on 110GB Mac, RAM, VRAM setups.
Run via Unsloth Studio or llama.cpp.
GGUF: unsloth/DeepSeek-V4-Flash-GGUF
Guide: https://unsloth.ai/docs/models/deepseek-v4
replied to HannesVonEssen's post about 1 month ago
Sick update. I've been using your visualizer for awhile now. Its cool to see your adding community focused features like this.
reacted to davidmezzetti's post with ๐ about 1 month ago
Post
1066
CeleBERTy Small: Domain model for Pop Culture, Art, Music and Entertainment.
Build vector embeddings that perform better than larger models.
https://huggingface.co/blog/NeuML/celeberty-small
Build vector embeddings that perform better than larger models.
https://huggingface.co/blog/NeuML/celeberty-small
reacted to Quazim0t0's post with ๐ about 1 month ago
Post
2402
Created research language model whose channel-mixing block is not an MLP. It is a differentiable Neighbour-Sensing fungal-colony-growth model: each token is expanded into a colony of hyphal tips that grow in a bounded latent region, sense a shared density field, and steer their own growth โ the "MLP" is replaced by a few differentiable steps of colony growth, read back out into the hidden state.
Quazim0t0/Mycel-LM-79M
Also the original SpikeWhale project โ the one that sparked all the other SpikeWhale related projects. Every spiking primitive here is hand-written in plain PyTorch: the leaky integrate-and-fire (LIF) neuron dynamics, the fast-sigmoid surrogate gradient, and the backprop-through-time training loop. No snntorch, no spikingjelly, no norse, no bindsnet โ the network is a genuine from-scratch SNN.
Quazim0t0/SpikeWhale-SNN-216M
Quazim0t0/Mycel-LM-79M
Also the original SpikeWhale project โ the one that sparked all the other SpikeWhale related projects. Every spiking primitive here is hand-written in plain PyTorch: the leaky integrate-and-fire (LIF) neuron dynamics, the fast-sigmoid surrogate gradient, and the backprop-through-time training loop. No snntorch, no spikingjelly, no norse, no bindsnet โ the network is a genuine from-scratch SNN.
Quazim0t0/SpikeWhale-SNN-216M
- 5 replies
replied to albertvillanova's post about 1 month ago
Absolutely stoked! Huge milestone for the team. Enjoy a bit of peace now that its over. Or just keep on the grind it's what I always do!
reacted to albertvillanova's post with ๐ฅ about 1 month ago
Post
3765
๐ KTO is now part of the stable TRL API
As of Promote KTO to stable API, KTOTrainer and KTOConfig have graduated from trl.experimental to the stable trl API. https://github.com/huggingface/trl/pull/6175
This one closes out a long road. Over the past 6+ months, the "Align KTO with DPO" effort landed ~90 PRs methodically bringing KTO up to the standard we hold for stable trainers, one carefully-scoped change at a time:
- Feature parity with DPO: full VLM support (incl. multi-image), sync_ref_model, PEFT + Liger, ZeRO-3 + PEFT dtype fix, pad_to_multiple_of, activation offloading, IterableDataset and dict eval_dataset, remove_unused_columns, and reference-logprob precomputation at init.
- Consistency with DPO: aligned method order and signatures, tokenization, _prepare_dataset, PEFT handling, ref-model preparation for distributed training, and config layout โ plus a new DataCollatorForKTO and output format. Metrics moved into _compute_loss and simplified to direct averages via the shared _metrics attribute.
- Removing legacy baggage: dropped encoder-decoder support, BOS/EOS handling, null_ref_context, generate_during_eval, model_init, preprocess_logits_for_metrics, model/ref adapter names, and several dead config knobs.
- Coverage: a full test suite mirroring DPO, text collator tests, VLM tests, and slow tests.
- The promotion itself: the experimental โ stable move (#6175) and shim cleanup (#6287), handled so downstream users get a clean deprecation path.
Honestly, this has been one of the more complex tasks I've taken on since joining the team, not because any single change was hard, but because it demanded sustained consistency across a ~2,000-line trainer, with every branch, comment, and edge case kept in lockstep with DPO.
Huge thanks to everyone who reviewed along the way (especially @qgallouedec ), the incremental review cadence is exactly what kept this maintainable.
KTO now sits on equal footing with our other flagship trainers. ๐
As of Promote KTO to stable API, KTOTrainer and KTOConfig have graduated from trl.experimental to the stable trl API. https://github.com/huggingface/trl/pull/6175
This one closes out a long road. Over the past 6+ months, the "Align KTO with DPO" effort landed ~90 PRs methodically bringing KTO up to the standard we hold for stable trainers, one carefully-scoped change at a time:
- Feature parity with DPO: full VLM support (incl. multi-image), sync_ref_model, PEFT + Liger, ZeRO-3 + PEFT dtype fix, pad_to_multiple_of, activation offloading, IterableDataset and dict eval_dataset, remove_unused_columns, and reference-logprob precomputation at init.
- Consistency with DPO: aligned method order and signatures, tokenization, _prepare_dataset, PEFT handling, ref-model preparation for distributed training, and config layout โ plus a new DataCollatorForKTO and output format. Metrics moved into _compute_loss and simplified to direct averages via the shared _metrics attribute.
- Removing legacy baggage: dropped encoder-decoder support, BOS/EOS handling, null_ref_context, generate_during_eval, model_init, preprocess_logits_for_metrics, model/ref adapter names, and several dead config knobs.
- Coverage: a full test suite mirroring DPO, text collator tests, VLM tests, and slow tests.
- The promotion itself: the experimental โ stable move (#6175) and shim cleanup (#6287), handled so downstream users get a clean deprecation path.
Honestly, this has been one of the more complex tasks I've taken on since joining the team, not because any single change was hard, but because it demanded sustained consistency across a ~2,000-line trainer, with every branch, comment, and edge case kept in lockstep with DPO.
Huge thanks to everyone who reviewed along the way (especially @qgallouedec ), the incremental review cadence is exactly what kept this maintainable.
KTO now sits on equal footing with our other flagship trainers. ๐
- 2 replies