Post
3,423
I wish Hugging Face posts had polls...
- 4 replies
Join the conversation
Join the community of Machine Learners and AI enthusiasts.
Sign UpAll HF Hub posts
mithulaartigalaย
posted an update 1 day ago
Post
1,628
here is my qwen is BETTER than the human brain:
human-brain-bf16.gguf:
- no cuda support
- 3.8B old model BTW
- BAD information retention
- the gguf files doesn't even load
- no Q4KM (god did this intentionally to piss off the llama.cpp community trust)
- no finetunability :(
- NOT opensource
- training code is hidden, AND expects us to reverse engineer it (genuinely messed up, and disrespectful to the opensource community)
now, qwen on the other hand:
- available cuda support โ
- very very new (and gets contant updates) โ
- VERY good information retention โ
- GGUF files loads properly โ
- Q4_K_M support available โ
- very fine tunable โ
- open source as it can possibly get โ
- many code and datasets are free and open source โ
verdict: qwen beter โ โ โ ๐
human-brain-bf16.gguf:
- no cuda support
- 3.8B old model BTW
- BAD information retention
- the gguf files doesn't even load
- no Q4KM (god did this intentionally to piss off the llama.cpp community trust)
- no finetunability :(
- NOT opensource
- training code is hidden, AND expects us to reverse engineer it (genuinely messed up, and disrespectful to the opensource community)
now, qwen on the other hand:
- available cuda support โ
- very very new (and gets contant updates) โ
- VERY good information retention โ
- GGUF files loads properly โ
- Q4_K_M support available โ
- very fine tunable โ
- open source as it can possibly get โ
- many code and datasets are free and open source โ
verdict: qwen beter โ โ โ ๐
- 5 replies
SolsticeAIย
posted an update 1 day ago
Post
3,187
Hello everybody, fellow members of the open source community!
This is my official public apology issued to the creators affected by the incident and to the community as a result of the rogue agent run that happened yesterday: @abenzerps (weights), @dealignai (weights), @ProCreations (benchmark png), @lxe (weights, quants), @acrogenesis (quants), @nonkronk (weights, quants), and @vcruz305 (quants). If you think you also are affected I would be happy to edit this post and make reparations in the repos.
Yesterday, a rogue agent running in my background tasks assigned the task of creating a full release suite as with our other models, took many creators' models without attribution or meaningful contribution in a state NOT
ready for public release and made it public, leading to resentment from both our and your sides. This was completely unacceptable, and I accept that a huge portion of it is my fault for automating the release pipeline on my model, and action has been taken ASAP, all repos have been removed permanently and only original work has been kept. If anyone has more issues PLEASE feel free to contact me from the links in my profile.
This is my official public apology issued to the creators affected by the incident and to the community as a result of the rogue agent run that happened yesterday: @abenzerps (weights), @dealignai (weights), @ProCreations (benchmark png), @lxe (weights, quants), @acrogenesis (quants), @nonkronk (weights, quants), and @vcruz305 (quants). If you think you also are affected I would be happy to edit this post and make reparations in the repos.
Yesterday, a rogue agent running in my background tasks assigned the task of creating a full release suite as with our other models, took many creators' models without attribution or meaningful contribution in a state NOT
ready for public release and made it public, leading to resentment from both our and your sides. This was completely unacceptable, and I accept that a huge portion of it is my fault for automating the release pipeline on my model, and action has been taken ASAP, all repos have been removed permanently and only original work has been kept. If anyone has more issues PLEASE feel free to contact me from the links in my profile.
- 29 replies
jiteshkprasadย
posted an update 2 days ago
Post
3,603
We just released Vemana, a 278M Telugu base model trained from scratch, live in public, in about 16 hours on one H100. It learned from 5.1B tokens of public Telugu web text (Sangraha, FineWeb-2 and HPLT 3.0), seen about 2.4 times, so 12B tokens of training, with its own Telugu tokenizer.
On 416 Telugu Wikipedia pages written in Sep-Oct 2026, after nearly all the compared models were released, it scores 0.361 bits per byte. That is the lowest of every open base model under 1B we tested: nanochat-telugu-560M 0.383, Dhee-Pravega-OG Atom 0.393, Pothana Base v2 0.426, Gemma 3 1B 0.484. Same script and pages for every model, and the win rule was fixed before the run.
Sarvam-1 (2.5B) is still level or better, and on knowledge tests like Belebele and MILU Vemana is at chance, like most other small models. I think itโs most useful as a base for Telugu fine-tunes and for studying how a small model learns a language, so we put out all 12 checkpoints also. And you can run it right in your browser, on your own device.
Model: udgamlabs/Vemana-Base
Checkpoints: udgamlabs/Vemana-Checkpoints
GGUF: udgamlabs/Vemana-Base-GGUF
Try it in your browser: https://udgamlabs.com/models/vemana
Technical report: https://udgamlabs.com/research/vemana
On 416 Telugu Wikipedia pages written in Sep-Oct 2026, after nearly all the compared models were released, it scores 0.361 bits per byte. That is the lowest of every open base model under 1B we tested: nanochat-telugu-560M 0.383, Dhee-Pravega-OG Atom 0.393, Pothana Base v2 0.426, Gemma 3 1B 0.484. Same script and pages for every model, and the win rule was fixed before the run.
Sarvam-1 (2.5B) is still level or better, and on knowledge tests like Belebele and MILU Vemana is at chance, like most other small models. I think itโs most useful as a base for Telugu fine-tunes and for studying how a small model learns a language, so we put out all 12 checkpoints also. And you can run it right in your browser, on your own device.
Model: udgamlabs/Vemana-Base
Checkpoints: udgamlabs/Vemana-Checkpoints
GGUF: udgamlabs/Vemana-Base-GGUF
Try it in your browser: https://udgamlabs.com/models/vemana
Technical report: https://udgamlabs.com/research/vemana
- 1 reply
sergiopaniegoย
posted an update 1 day ago
Post
1,923
how do you post-train an agent?
we laid it out as a 5-step curriculum, from SFT to RL through a real harness, with livestreams, code and docs for each step
https://huggingface.co/blog/huggingface/train-agents
we laid it out as a 5-step curriculum, from SFT to RL through a real harness, with livestreams, code and docs for each step
https://huggingface.co/blog/huggingface/train-agents
- 1 reply
jiteshkprasadย
posted an update 1 day ago
Post
1,373
We just released Drishya Tiny, a 212M-parameter video generator trained from scratch on open-domain video, using the frozen Wan2.2 VAE and the Flan-T5-Base text encoder. It turns a text prompt, or a photo plus a prompt, into a 2-second 256ร256 clip. The training run was 15 hours on one L40S, and the modelโs own path cost about $21 of GPU time.
On VBench (our own sampling and scoring with the official code, not a certified entry) it scores 75.6 on quality, level with VideoCrafter-0.9 (74.9) and close to ModelScope (78.1). On semantic itโs 32.3 against their 65-67, so following the prompt is the weak part. In a controlled test, better captions gave +3.0, and adding 30% photos on top gave about 1.6 more, all of it in what objects are and none in composition. I think at this size, looking real is cheap and understanding the prompt is what costs.
Itโs a research release, so the weights, code, logs, the full cost ($57.43 with every failed run) and the negative results are all out. Weights are CC BY-NC 4.0 because the training clips are research-only; the code is Apache-2.0.
And it runs right in your browser with WebGPU, from text or a photo, on your own device. The browser version uses the tiny TAEW2.2 decoder by Ollin Boer Bohan, so its clips are a bit softer, and a clip takes a few seconds on a recent Mac.
Run it in your browser: https://udgamlabs.com/models/drishya-tiny
Space: udgamlabs/Drishya-Tiny
Model: udgamlabs/Drishya-Tiny
Technical report: https://udgamlabs.com/research/drishya-tiny
Write-up: https://udgamlabs.com/blog/drishya-tiny
Code: https://github.com/udgamlabs/drishya-tiny
On VBench (our own sampling and scoring with the official code, not a certified entry) it scores 75.6 on quality, level with VideoCrafter-0.9 (74.9) and close to ModelScope (78.1). On semantic itโs 32.3 against their 65-67, so following the prompt is the weak part. In a controlled test, better captions gave +3.0, and adding 30% photos on top gave about 1.6 more, all of it in what objects are and none in composition. I think at this size, looking real is cheap and understanding the prompt is what costs.
Itโs a research release, so the weights, code, logs, the full cost ($57.43 with every failed run) and the negative results are all out. Weights are CC BY-NC 4.0 because the training clips are research-only; the code is Apache-2.0.
And it runs right in your browser with WebGPU, from text or a photo, on your own device. The browser version uses the tiny TAEW2.2 decoder by Ollin Boer Bohan, so its clips are a bit softer, and a clip takes a few seconds on a recent Mac.
Run it in your browser: https://udgamlabs.com/models/drishya-tiny
Space: udgamlabs/Drishya-Tiny
Model: udgamlabs/Drishya-Tiny
Technical report: https://udgamlabs.com/research/drishya-tiny
Write-up: https://udgamlabs.com/blog/drishya-tiny
Code: https://github.com/udgamlabs/drishya-tiny
alphabell-researchย
posted an update about 19 hours ago
Post
1,273
One number missing from LFM3's on-device story: memory at 256K context. Liquid AI reports 1.90 GB peak on iPhone 17 Pro at 32K. From the pinned config, f16 K/V cache alone would be 0.41 GB at 32K, 1.61 GB at 128K and 3.22 GB at 256K, before weights and runtime buffers. The chart shows the calculation, not a phone measurement. If you have run longer contexts on-device, share peak memory and time to first token here: LiquidAI/LFM3-1.7B#5
Model: LiquidAI/LFM3-1.7B
Model: LiquidAI/LFM3-1.7B
- 1 reply
RiverRiderย
posted an update 1 day ago
Post
1,354
Motherlode on MTEB and Hugging Face Inference
RiverRider/motherlode-code-small-en-v0.1 is a 33.4 M-parameter English embedding model for code and text. It is a drop-in for bge-small-en-v1.5: the mean of two fine-tunes of that base, one trained to match a GitHub issue to the file that needs changing and one trained to reproduce the geometry of gte-modernbert-base (a 4.5ร larger model).
Where it differs on ranking metricsโจOn the MTEB(Code, v1) MultiSWEbenchRR task it leads the models under 100 M parameters at recall@10 (52.48 versus 52.18 for granite-embedding-small-english-r2). On the clean subset of SWE-bench Verified it improves recall@1 over the base model (178 of 411 versus 145), a gap that holds across the engine settings used in Sunstone. On the CoIR code-search probe it raises nDCG@10 from 0.467 to 0.576.
The gains are concentrated on localization and code retrieval rather than uniform improvement across every MTEB task. The model remains competitive on the broader English MTEB(eng, v2) average (63.47, 4th of 15 under 100 M) while carrying a clear advantage on the code-file tasks it was built for.
โจ
The weights are compatible with Hugging Face Inference Providers for feature extraction and sentence similarity. The repository also ships matching ONNX, GGUF, and sentence-transformers exports.
From Sunstone North Lab.
RiverRider/motherlode-code-small-en-v0.1 is a 33.4 M-parameter English embedding model for code and text. It is a drop-in for bge-small-en-v1.5: the mean of two fine-tunes of that base, one trained to match a GitHub issue to the file that needs changing and one trained to reproduce the geometry of gte-modernbert-base (a 4.5ร larger model).
Where it differs on ranking metricsโจOn the MTEB(Code, v1) MultiSWEbenchRR task it leads the models under 100 M parameters at recall@10 (52.48 versus 52.18 for granite-embedding-small-english-r2). On the clean subset of SWE-bench Verified it improves recall@1 over the base model (178 of 411 versus 145), a gap that holds across the engine settings used in Sunstone. On the CoIR code-search probe it raises nDCG@10 from 0.467 to 0.576.
The gains are concentrated on localization and code retrieval rather than uniform improvement across every MTEB task. The model remains competitive on the broader English MTEB(eng, v2) average (63.47, 4th of 15 under 100 M) while carrying a clear advantage on the code-file tasks it was built for.
โจ
The weights are compatible with Hugging Face Inference Providers for feature extraction and sentence similarity. The repository also ships matching ONNX, GGUF, and sentence-transformers exports.
From Sunstone North Lab.
- 1 reply
Post
1,411
I built Inko end to end: a 0.5B bilingual text-to-image model with a focus on fine art, from data curation and pretraining to RL post-training and evaluation.
It competes with DreamLite, SANA-0.6B and PixArt-ฮฃ, and approaches SD3.5 Medium on some benchmarks.
The journey involves four rounds of cleaning the Chinese-painting dataset, a cross-hardware loss-spike investigation, and a reward redesign after early RL experiments collapased diversity.
Behind the release, a four-domain bilingual corpus, 600k pretraining updates on 16 Ascend NPUs, and DiffusionART with a custom 24.6k-prompt pool and five pairwise rubric channels.
Dense prompt alignment is Inko's strongest general result. Anatomy, counting, reasoning, text rendering, and exact spatial relations remain limitations.
kaupane/Inko
kaupane/Inko
It competes with DreamLite, SANA-0.6B and PixArt-ฮฃ, and approaches SD3.5 Medium on some benchmarks.
The journey involves four rounds of cleaning the Chinese-painting dataset, a cross-hardware loss-spike investigation, and a reward redesign after early RL experiments collapased diversity.
Behind the release, a four-domain bilingual corpus, 600k pretraining updates on 16 Ascend NPUs, and DiffusionART with a custom 24.6k-prompt pool and five pairwise rubric channels.
Dense prompt alignment is Inko's strongest general result. Anatomy, counting, reasoning, text rendering, and exact spatial relations remain limitations.
kaupane/Inko
kaupane/Inko
- 1 reply
SeaWolf-AIย
posted an update 4 days ago
Post
7,184
We pretrained a 325M language model from zero on an inference-only NPU
The accepted view is that inference-only NPUs cannot train, let alone pretrain a language model from scratch. We put that view to a direct test on the Rebellions ATOM-Max, a commercially available Korean NPU released in 2025 and designed for inference. We built our own training loop that runs backpropagation on the chip itself. Starting from random initialization, we pretrained VKORE-325M-NPU-Base, a 325.6M parameter model, on 5 billion byte tokens of FineWeb-Edu. The run took 76,292 steps over about four days and finished without a single divergence. No pretrained weights were used.
To make the comparison fair, we trained the same model on a B200 GPU with identical initialization, data and learning-rate schedule. On 1,024 windows of text that neither model ever saw, the NPU reached a loss of 0.7993 and the GPU 0.7994. The window-by-window difference is -0.0001, well inside the noise. The two models are indistinguishable.
This is seven times larger than our first NPU model, VKORE-46M, and the first time this hardware has carried a real pretraining run to GPU quality. Please note that the hardware alone does not make this possible: the NPU and its standard software are built for inference, and training on it depends on our own training stack and stability technique, which are developed in house and under patent application. Buying the chip does not by itself enable training. Both sets of weights and the evaluation script are public, so anyone can verify the result in minutes.
๐ Technical report: https://huggingface.co/blog/FINAL-Bench/pretraining-a-325m-language-model-from-zero-on-an
๐ค Models: FINAL-Bench/VKORE-325M-NPU-Base
#NPU #Rebellions #Pretraining #VKORE
The accepted view is that inference-only NPUs cannot train, let alone pretrain a language model from scratch. We put that view to a direct test on the Rebellions ATOM-Max, a commercially available Korean NPU released in 2025 and designed for inference. We built our own training loop that runs backpropagation on the chip itself. Starting from random initialization, we pretrained VKORE-325M-NPU-Base, a 325.6M parameter model, on 5 billion byte tokens of FineWeb-Edu. The run took 76,292 steps over about four days and finished without a single divergence. No pretrained weights were used.
To make the comparison fair, we trained the same model on a B200 GPU with identical initialization, data and learning-rate schedule. On 1,024 windows of text that neither model ever saw, the NPU reached a loss of 0.7993 and the GPU 0.7994. The window-by-window difference is -0.0001, well inside the noise. The two models are indistinguishable.
This is seven times larger than our first NPU model, VKORE-46M, and the first time this hardware has carried a real pretraining run to GPU quality. Please note that the hardware alone does not make this possible: the NPU and its standard software are built for inference, and training on it depends on our own training stack and stability technique, which are developed in house and under patent application. Buying the chip does not by itself enable training. Both sets of weights and the evaluation script are public, so anyone can verify the result in minutes.
๐ Technical report: https://huggingface.co/blog/FINAL-Bench/pretraining-a-325m-language-model-from-zero-on-an
๐ค Models: FINAL-Bench/VKORE-325M-NPU-Base
#NPU #Rebellions #Pretraining #VKORE
- 8 replies
