๐ Something big is coming
LH-Tech AI
LH-Tech-AI
AI & ML interests
Small AI and ML models at SupraLabs. Open source research. For you and the community.
Recent Activity
liked a dataset about 9 hours ago
openbmb/Ultra-FineWeb-L3 liked a dataset about 9 hours ago
allenai/dolma3_dolmino_mix-100B-1125 liked a dataset about 9 hours ago
common-pile/stackexchangeOrganizations
reacted to DedeProGames's post with ๐ฅ 2 days ago
reacted to appvoid's post with โค๏ธ 10 days ago
Post
7,610
New leaderboard coming to the small language model community!
The leaderboard measures cognitive skills, which is more challenging and beyond language modeling and instruction-following. This is a first step towards more everyday useful and general small models.
dotlabs/cognirank
The leaderboard measures cognitive skills, which is more challenging and beyond language modeling and instruction-following. This is a first step towards more everyday useful and general small models.
dotlabs/cognirank
reacted to AxionLab-official's post with โค๏ธ 14 days ago
Post
5,221
โ ๏ธ Community Notice
We would like to clarify that SupraLabs has no affiliation, partnership, or connection whatsoever with "SupraLarps" or its members.
Please avoid interacting with their organization, repositories, or Spaces under the assumption that they are associated with us.
We are currently aware of the situation and have already contacted the appropriate channels to address it.
Thank you to everyone who continues to support SupraLabs. โค๏ธ
We would like to clarify that SupraLabs has no affiliation, partnership, or connection whatsoever with "SupraLarps" or its members.
Please avoid interacting with their organization, repositories, or Spaces under the assumption that they are associated with us.
We are currently aware of the situation and have already contacted the appropriate channels to address it.
Thank you to everyone who continues to support SupraLabs. โค๏ธ
reacted to nightmedia's post with ๐ 15 days ago
Post
3,309
Qwen3.5 Performance Metrics
With the 3.5 architecture, a lot of the old quanting methods don't work as before. I noticed this when benchmarking Deckard(qx) quants and by mistake ran a q8 that was better. That only happens if the qx sucked--and it did--enhancing layers just because they look interesting doesn't work anymore, so until I get a clear understanding of the architecture, I will publish mxfp4 and mxfp8 of the 3.5 models, that seem very stable and high performant
I will start posting here the metrics I gather from the series, starting with the smallest. If I have numbers from previous or similar models, I will post them in comparison
Qwen3.5-0.8B
Detailed metrics by model
nightmedia/Qwen3.5-0.8B-mxfp8-mlx
nightmedia/Qwen3.5-2B-mxfp8-mlx
nightmedia/Qwen3.5-4B-mxfp8-mlx
nightmedia/Qwen3.5-9B-mxfp8-mlx
https://huggingface.co/nightmedia/Qwen3.5-27B-Text
nightmedia/Qwen3.5-122B-A10B-Text-mxfp4-mlx
More metrics coming soon.
I am running these on my Mac, an M4Max with 128GB RAM. Some performance numbers like tokens/second reflect the performance on my box.
This post will be updated with every model that gets tested. The larger models take hours, the 27B a couple days, so it will be a long process.
-G
With the 3.5 architecture, a lot of the old quanting methods don't work as before. I noticed this when benchmarking Deckard(qx) quants and by mistake ran a q8 that was better. That only happens if the qx sucked--and it did--enhancing layers just because they look interesting doesn't work anymore, so until I get a clear understanding of the architecture, I will publish mxfp4 and mxfp8 of the 3.5 models, that seem very stable and high performant
I will start posting here the metrics I gather from the series, starting with the smallest. If I have numbers from previous or similar models, I will post them in comparison
Qwen3.5-0.8B
quant arc arc/e boolq hswag obkqa piqa wino
mxfp8 0.351,0.501,0.733,0.462,0.348,0.682,0.573
mxfp4 0.339,0.489,0.738,0.433,0.330,0.672,0.553
Old model performance
Qwen3-0.6B
bf16 0.298,0.354,0.378,0.415,0.344,0.649,0.534
q8-hi 0.296,0.355,0.378,0.416,0.348,0.652,0.529
q8 0.299,0.354,0.378,0.414,0.346,0.650,0.535
q6-hi 0.301,0.356,0.378,0.415,0.350,0.651,0.541
q6 0.300,0.367,0.378,0.416,0.344,0.647,0.524
mxfp4 0.286,0.364,0.609,0.404,0.316,0.626,0.531
Quant Perplexity Peak memory
mxfp8 6.611 ยฑ 0.049 7.65 GB
mxfp4 7.455 ยฑ 0.057 6.33 GBDetailed metrics by model
nightmedia/Qwen3.5-0.8B-mxfp8-mlx
nightmedia/Qwen3.5-2B-mxfp8-mlx
nightmedia/Qwen3.5-4B-mxfp8-mlx
nightmedia/Qwen3.5-9B-mxfp8-mlx
https://huggingface.co/nightmedia/Qwen3.5-27B-Text
nightmedia/Qwen3.5-122B-A10B-Text-mxfp4-mlx
More metrics coming soon.
I am running these on my Mac, an M4Max with 128GB RAM. Some performance numbers like tokens/second reflect the performance on my box.
This post will be updated with every model that gets tested. The larger models take hours, the 27B a couple days, so it will be a long process.
-G
reacted to mithulaartigala's post with ๐ง 16 days ago
Post
3,133
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
reacted to mithulaartigala's post with ๐ค 18 days ago
Post
3,133
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
reacted to SeaWolf-AI's post with ๐๐ฅ 21 days ago
Post
7,252
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
reacted to Banaxi-Tech's post with ๐๐๐ 21 days ago
reacted to Banaxi-Tech's post with ๐ 22 days ago
reacted to mithulaartigala's post with โ๐ฅ 22 days ago
Post
3,281
We need to make software where people can start using decision models.
reacted to Banaxi-Tech's post with ๐ 22 days ago
reacted to mithulaartigala's post with ๐ฅ 26 days ago
replied to mithulaartigala's post 26 days ago
OCTOBER, 25TH, 2026.
replied to their post 29 days ago
Hey, yeah! Wanna cooperate with SupraLabs for that compute grant?
Message me on discord please:
lh_tech_ai
replied to their post 30 days ago
ok but i don't have time for a call/meeting now. can you DM me on discord pls?
replied to their post 30 days ago
I don't see anyone booking my calendar! Happy to talk and see if your the right fit for our program ๐
Please... can we chat on discord? lh_tech_ai is my name there