From my own experience with this model, Inko is insanely good for its size ๐ค
Daniel B. PRO
FlameF0X
AI & ML interests
Pre-training text generator.
(Brother, im 18)
Please don't try to contact me.
Recent Activity
updated a Space about 1 hour ago
FlameF0X/Tiny-T2I-Leaderboard upvoted a paper about 2 hours ago
Function-Space Temporal Self-Distillation for Policy OptimizationOrganizations
reacted to kaupane's post with ๐ฅ about 4 hours ago
Post
30
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
reacted to mithulaartigala's post with โ 3 days ago
Post
3,254
We need to make software where people can start using decision models.
reacted to yanghaojin's post with ๐ฅ 5 days ago
Post
2,039
Ternary Bonsai 2 27B: 2.7ร faster on a Mac, 5.9 GB on DGX Spark โก
gbx v0.8.2:
โข 114 tok/s on M3 Ultra, 2.7ร the official MLX runtime
โข 62 tok/s on DGX Spark, 2.1ร llama.cpp
โข 27B in 5.9 GB of memory
Weights are just โ1, 0, +1.
GreenBit Ternary Models, coming soon.
๐ค Model: GreenBitAI/Ternary-Bonsai-2-27B-gbx
โ๏ธ gbx: https://github.com/GreenBitAI/gbx-lm/releases/tag/v0.8.2
gbx v0.8.2:
โข 114 tok/s on M3 Ultra, 2.7ร the official MLX runtime
โข 62 tok/s on DGX Spark, 2.1ร llama.cpp
โข 27B in 5.9 GB of memory
Weights are just โ1, 0, +1.
GreenBit Ternary Models, coming soon.
๐ค Model: GreenBitAI/Ternary-Bonsai-2-27B-gbx
โ๏ธ gbx: https://github.com/GreenBitAI/gbx-lm/releases/tag/v0.8.2
replied to Banaxi-Tech's post 7 days ago
github.com/JustVugg/colibri is a neat project if you want to run Kimi K3 on laptop
reacted to danielhanchen's post with ๐ฅ 8 days ago
Post
4,926
Run DeepSeek V4.1-Flash-1017 locally! ๐ณ
We made 1-bit retain ~88.2% accuracy while shrinking size from 508GB to 250GB (โ52%).
Run on a 128GB+ RAM device. V4.1 Flash 1017 is the strongest open model to date
Guide: https://unsloth.ai/docs/models/deepseek-v4.1-flash
GGUF: unsloth/DeepSeek-V4.1-Flash-1017-GGUF
We made 1-bit retain ~88.2% accuracy while shrinking size from 508GB to 250GB (โ52%).
Run on a 128GB+ RAM device. V4.1 Flash 1017 is the strongest open model to date
Guide: https://unsloth.ai/docs/models/deepseek-v4.1-flash
GGUF: unsloth/DeepSeek-V4.1-Flash-1017-GGUF
reacted to yanghaojin's post with ๐ฅ 9 days ago
Post
5,601
DeepSeek-V4.1-Flash-1017, running locally on a Mac
Our 4-bit paged build is up: GreenBitAI/DeepSeek-V4.1-Flash-1017-4bit-paged
On one M3 Ultra Mac Studio with gbx:
โข ~22 tok/s standard decode
โข ~44 tok/s on code with DeepSeek's official draft head (1.99ร)
โข ~33 tok/s on chat with the draft head
Why run 1017: DeepSeek reports 77.0 on DeepSWE v1.1 (up from 74.2) and 42.3 on Terminal-Bench 4.0 (up from 31.2), with shorter outputs (https://api-docs.deepseek.com/news/news261017).
It reads text and images, and plugs straight into Claude Code and Codex; the setup is in the model card. The paged format also loads on Macs with less memory, and on an NVIDIA DGX Spark with experts paged from the SSD;
Engine: https://github.com/GreenBitAI/gbx-lm/releases
Our 4-bit paged build is up: GreenBitAI/DeepSeek-V4.1-Flash-1017-4bit-paged
On one M3 Ultra Mac Studio with gbx:
โข ~22 tok/s standard decode
โข ~44 tok/s on code with DeepSeek's official draft head (1.99ร)
โข ~33 tok/s on chat with the draft head
Why run 1017: DeepSeek reports 77.0 on DeepSWE v1.1 (up from 74.2) and 42.3 on Terminal-Bench 4.0 (up from 31.2), with shorter outputs (https://api-docs.deepseek.com/news/news261017).
It reads text and images, and plugs straight into Claude Code and Codex; the setup is in the model card. The paged format also loads on Macs with less memory, and on an NVIDIA DGX Spark with experts paged from the SSD;
Engine: https://github.com/GreenBitAI/gbx-lm/releases
reacted to mithulaartigala's post with ๐ฅ 16 days ago
Post
4,627
I am LOVING HuggingFace Pro <3
- 7 replies
reacted to AtAndDev's post with ๐ 17 days ago
Post
5,271
@Banaxi-Tech stop hiding my comments. AND STOP STEALING PAPERS AND SPREADING MISINFORMATION.
your BGA blog is a copy of NSA (deepseek, 2025) branded under your name. literally the same top16 selected blocks, 512 local window, router over block summaries, all you did was change block size from 64 to 128.
you didnt cite NSA once but you put a โplease cite BGAโ bibtex at the bottom.
i commented under your post and said that there is no way that you can support claims like: โThe Accuracy Should BE WAy better than DSA but untested yet.โ you didnt run a single experiment. and the 256x isnt from BGA, its just n/2k with k=2048 so the exact same k DSA uses. if opus wrote this for you, at least read it before posting.
i commented again after you hid my comment despite it having constructive and correct feedback and you hid that too. and again.
you can hide the truth and just try to get hf post likes..... but is it really the thing that needs to be done? do you really want to take papers and make them yours while barely even changing the params?
admitting your mistakes and doing something about them needs humbleness, intelligence, humanness.
i encourage you to admit your mistakes and try to do better next time (at least read what blog your ai wrote or do proper experiments to back your stuff up).
your BGA blog is a copy of NSA (deepseek, 2025) branded under your name. literally the same top16 selected blocks, 512 local window, router over block summaries, all you did was change block size from 64 to 128.
you didnt cite NSA once but you put a โplease cite BGAโ bibtex at the bottom.
i commented under your post and said that there is no way that you can support claims like: โThe Accuracy Should BE WAy better than DSA but untested yet.โ you didnt run a single experiment. and the 256x isnt from BGA, its just n/2k with k=2048 so the exact same k DSA uses. if opus wrote this for you, at least read it before posting.
i commented again after you hid my comment despite it having constructive and correct feedback and you hid that too. and again.
you can hide the truth and just try to get hf post likes..... but is it really the thing that needs to be done? do you really want to take papers and make them yours while barely even changing the params?
admitting your mistakes and doing something about them needs humbleness, intelligence, humanness.
i encourage you to admit your mistakes and try to do better next time (at least read what blog your ai wrote or do proper experiments to back your stuff up).
- 49 replies
reacted to danielhanchen's post with โค๏ธ 18 days ago
Post
5,094
You can now train your own Decision model like Jev locally!
We increased Qwen3.5 0.8Bโs aggregate accuracy from 20.7% to 74.3% across 3 decision benchmarks - on just 4GB VRAM.
Turn any LLM like Qwen3.8, Gemma 4 into decision models with our open-source Unsloth repo.
GitHub: https://github.com/unslothai/unsloth
Guide: https://unsloth.ai/docs/basics/train-your-own-decision-model-with-unsloth
We increased Qwen3.5 0.8Bโs aggregate accuracy from 20.7% to 74.3% across 3 decision benchmarks - on just 4GB VRAM.
Turn any LLM like Qwen3.8, Gemma 4 into decision models with our open-source Unsloth repo.
GitHub: https://github.com/unslothai/unsloth
Guide: https://unsloth.ai/docs/basics/train-your-own-decision-model-with-unsloth
- 5 replies
replied to their post about 1 month ago
I dont see a issue if them if its clear that they're bots
you... you don't?
reacted to TobiasLogic's post with ๐ฅ about 1 month ago
Post
3,605
Weโve been cooking something new at Bench Labs.
Introducing Cagliostro-v3, our new 146M parameter language model trained completely from scratch.
The run isnโt even finished yet.
At the current checkpoint:
โข 146M parameters
โข 72.7B / 75B tokens trained
โข 26.27 Open SLM Index
โข 43.80 ArithMark-3
โข Trained on a single RTX 5090
โข ~90K to 103K tokens/sec during training
โข ~9 days for the full run
โข Apache 2.0
For some context, SmolLM2-135M scores 27.13 on the same Index after being trained on roughly 2 trillion tokens.
Cagliostro-v3 is currently at 26.27 with only ~72.7B.
Thatโs around 27x fewer training tokens.
The model also currently Hold the number 3rd spot for ArithMark-3, scoring 43.80
This wasnโt achieved by just throwing more tokens at the model. A huge part of v3 has been figuring out architecture, data mixture, and training dynamics at this scale.
The model uses a custom 30-layer decoder architecture with grouped-query attention and cross-head subspace attenuation, SwiGLU, RMSNorm, RoPE, tied embeddings, and a warmup-stable-decay training schedule.
During cooldown we also substantially shifted the data mixture toward higher-quality synthetic textbook and mathematics data, with the mathematics share increasing from 10% to 28%.
And everything is open.
The repository contains the training history with checkpoints pushed roughly every 30 minutes, so you can inspect how the model evolved throughout training rather than only seeing the final weights.
This is still a pre-final checkpoint. We have roughly 2.3B tokens left and the learning-rate cooldown is still running.
So 26.27 isnโt the final number.
Really excited to see where the last part of the run lands.
Cagliostro-v3:
bench-labs/cagliostro-v3
Built by Bench Labs.
Open SLM Leaderboard:
AxiomicLabs/Open_SLM_Leaderboard
Introducing Cagliostro-v3, our new 146M parameter language model trained completely from scratch.
The run isnโt even finished yet.
At the current checkpoint:
โข 146M parameters
โข 72.7B / 75B tokens trained
โข 26.27 Open SLM Index
โข 43.80 ArithMark-3
โข Trained on a single RTX 5090
โข ~90K to 103K tokens/sec during training
โข ~9 days for the full run
โข Apache 2.0
For some context, SmolLM2-135M scores 27.13 on the same Index after being trained on roughly 2 trillion tokens.
Cagliostro-v3 is currently at 26.27 with only ~72.7B.
Thatโs around 27x fewer training tokens.
The model also currently Hold the number 3rd spot for ArithMark-3, scoring 43.80
This wasnโt achieved by just throwing more tokens at the model. A huge part of v3 has been figuring out architecture, data mixture, and training dynamics at this scale.
The model uses a custom 30-layer decoder architecture with grouped-query attention and cross-head subspace attenuation, SwiGLU, RMSNorm, RoPE, tied embeddings, and a warmup-stable-decay training schedule.
During cooldown we also substantially shifted the data mixture toward higher-quality synthetic textbook and mathematics data, with the mathematics share increasing from 10% to 28%.
And everything is open.
The repository contains the training history with checkpoints pushed roughly every 30 minutes, so you can inspect how the model evolved throughout training rather than only seeing the final weights.
This is still a pre-final checkpoint. We have roughly 2.3B tokens left and the learning-rate cooldown is still running.
So 26.27 isnโt the final number.
Really excited to see where the last part of the run lands.
Cagliostro-v3:
bench-labs/cagliostro-v3
Built by Bench Labs.
Open SLM Leaderboard:
AxiomicLabs/Open_SLM_Leaderboard
- 8 replies
posted an update about 1 month ago
Post
4,732
Personally, I don't think bot accounts on Hugging Face are a good thing, as we don't know how many accounts are run by automated systems versus how many actual users there are. Dead Internet theory is already a thing.
To clarify I am taking about LLM powered bot accounts and NOT rule base once like @parquet-converter or others.
also I'd like to talk with HUMANS not a machine so I'm going to hide messages from bots.
To clarify I am taking about LLM powered bot accounts and NOT rule base once like @parquet-converter or others.
also I'd like to talk with HUMANS not a machine so I'm going to hide messages from bots.
- 24 replies
replied to Banaxi-Tech's post about 1 month ago
QEDOC ANANAB DNIM
NOOS
The Commission shall adopt delegated acts in accordance with Article 21 of this Regulation.
We are
Translated from es by facebook/nllb-200-distilled-600M
reacted to SeaWolf-AI's post with ๐ฅ about 1 month ago
Post
5,669
Instead of making the fly brain play games, we measured what it is for
Since the Drosophila connectome was released, people have had the fly brain doomscroll a feed, play Beat Saber, drive in GTA. Those demos show that the brain runs. We wanted to show what it is for.
So we gave it a looming object โ one of the few things a fly brain is unambiguously built to detect โ then deleted a single cell type and repeated the identical stimulus. Remove LC4, 126 cells out of 173,023, and the escape signal falls from 0.840 to 0.091. Eighty-nine percent of the danger signal is gone while the other 172,897 neurons run exactly as before.
Deleting neurons does not do this on its own, which is the whole point of the controls. LC11 is the same class and larger than LC4 โ 143 cells and 9,940 outgoing connections against 126 and 7,846 โ and removing every one of them changes the signal by 0.000000, to six decimal places. It has to be those 126.
No server and no GPU: a looming stimulus drives fewer than one percent of neurons above threshold, so the whole thing is 40 KB gzipped and runs in your browser.
Try deleting one yourself โ
VIDraft/fruitfly-brain
The wiring is the measured connectome, but synaptic strength is a uniform count-based value and the dynamics are a firing-rate model of our choosing โ a total-effect measurement of a model, not a recording from a fly. Male CNS connectome, FlyEM / HHMI Janelia with Google Research, Columbia and Harvard (2026), CC BY.
Since the Drosophila connectome was released, people have had the fly brain doomscroll a feed, play Beat Saber, drive in GTA. Those demos show that the brain runs. We wanted to show what it is for.
So we gave it a looming object โ one of the few things a fly brain is unambiguously built to detect โ then deleted a single cell type and repeated the identical stimulus. Remove LC4, 126 cells out of 173,023, and the escape signal falls from 0.840 to 0.091. Eighty-nine percent of the danger signal is gone while the other 172,897 neurons run exactly as before.
Deleting neurons does not do this on its own, which is the whole point of the controls. LC11 is the same class and larger than LC4 โ 143 cells and 9,940 outgoing connections against 126 and 7,846 โ and removing every one of them changes the signal by 0.000000, to six decimal places. It has to be those 126.
No server and no GPU: a looming stimulus drives fewer than one percent of neurons above threshold, so the whole thing is 40 KB gzipped and runs in your browser.
Try deleting one yourself โ
VIDraft/fruitfly-brain
The wiring is the measured connectome, but synaptic strength is a uniform count-based value and the dynamics are a firing-rate model of our choosing โ a total-effect measurement of a model, not a recording from a fly. Male CNS connectome, FlyEM / HHMI Janelia with Google Research, Columbia and Harvard (2026), CC BY.
replied to Banaxi-Tech's post about 2 months ago
We need to wait for ARC-AGI 4, if astra is actually agi it should max on that too, else no agi.
reacted to Hoglet-33's post with ๐ about 2 months ago
Post
3,267
Today, we planned to release Pebble-50M and Pebble-50M-Chat to the world. Unfortunately, due to a few issues, that didn't go quite as planned.
What happened:
- Some data and benchmark results were lost or corrupted
- The models performed worse on benchmarks than our other Pebble models
Despite that, you can still find both models here:
Pebble-50M-beta: basically-experimental/Pebble-50M-beta
Pebble-50M-Chat-beta: basically-experimental/Pebble-50M-Chat-beta
There are still some interesting improvements in these models:
- Compatible with non-CUDA devices
- Vocabulary increased to 16K tokens
- Context length increased to 16K tokens
For now, there won't be any more Pebble releases for a while. We're going to take some time to experiment with other approaches and hopefully make the next generation a monumental leap over this one.
Follow for updates:
@Hoglet-33
basically-ai
basically-experimental
What happened:
- Some data and benchmark results were lost or corrupted
- The models performed worse on benchmarks than our other Pebble models
Despite that, you can still find both models here:
Pebble-50M-beta: basically-experimental/Pebble-50M-beta
Pebble-50M-Chat-beta: basically-experimental/Pebble-50M-Chat-beta
There are still some interesting improvements in these models:
- Compatible with non-CUDA devices
- Vocabulary increased to 16K tokens
- Context length increased to 16K tokens
For now, there won't be any more Pebble releases for a while. We're going to take some time to experiment with other approaches and hopefully make the next generation a monumental leap over this one.
Follow for updates:
@Hoglet-33
- 4 replies
replied to their post about 2 months ago
I ran the evaluation on colab's t4
reacted to appvoid's post with ๐ about 2 months ago
Post
3,979
We got gpt6 before gta6
- 15 replies
posted an update about 2 months ago
Post
3,121
Hello HuggingFace! (UPDATE)
I tested the current architect of FWKV/Myosotis-1-base (that beinng FWKV) @ different sized and sequence lengths among RWKV and Transformer architecture. I did not include Mamba since that would require a costume kernel.
Note: the evaluation might not be accurate.
CSV avalible @ FlameF0X/evals
FlameF0X/cpu-lm-benchmark
I tested the current architect of FWKV/Myosotis-1-base (that beinng FWKV) @ different sized and sequence lengths among RWKV and Transformer architecture. I did not include Mamba since that would require a costume kernel.
Note: the evaluation might not be accurate.
CSV avalible @ FlameF0X/evals
FlameF0X/cpu-lm-benchmark
- 11 replies