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-Σ, 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.
The best text-only setup glm-5.2 + mm outperformed gemini-3.5-flash, the top vision model, by 5.4 points (6.5%). It was also:
• 1.5x faster (100s vs 150s mean per task) • the only setup with zero timeouts (46/46 completed; gemini timed out 4x on bulk-image and long-video tasks) • the only setup stable across runs (88.5 / 88.4) • top on video (100.0), image (91.7), and PDF (90.0) tasks
The trade-offs: the toolkit consumed 3.3x more tokens (4.25M vs 1.28M), and lost on audio (85.6 vs 71.3).
On completed tasks alone the two are nearly identical (91.0 vs 88.4): the toolkit's edge is efficient extraction that keeps long media tasks inside the time budget.
🌍 World Model Bench — does your world model actually think?
FID measures realism. FVD measures smoothness. But neither tells you whether the model understood the scene.
We just released WM Bench — the first benchmark for cognitive intelligence in world models. The core question: when a beast charges from 3 meters away, does the model know to sprint — not walk? Does it respond differently to a human vs an animal? Does it remember the left corridor was blocked two steps ago?
Those are cognitive questions. No existing benchmark asks them. So we built one.
- 👁 P1 Perception (25%) — Can it read the scene? - 🧠 P2 Cognition (45%) — Does it predict threats, escalate emotions, utilize memory? - 🔥 P3 Embodiment (30%) — Does the body respond with the right motion?
All evaluation is via simple JSON I/O — no 3D engine, no special hardware. Any model with an API can participate.
We also built PROMETHEUS as a live reference implementation — runs in your browser on a T4, no install needed. Combines FloodDiffusion motion generation with a LLM cognitive brain (Perceive → Predict → Decide → Act). Scored 726/1000 (Grade B) on Track C — the only directly verified model so far. Submissions from other teams very welcome.
🏆 Nacrith: a 135M model that out-compresses everything on natural language
What if a tiny LM could compress english text better than _every_ compressor out there — classical or neural, small or large?
Nacrith pairs SmolLM2-135M with an ensemble of online predictors and high-precision arithmetic coding.
What's inside
The standard LLM+arithmetic coding approach wastes ~75% of CDF precision on large vocabularies. Our CDF-24 fix alone recovers 0.5 bpb. On top: a token N-gram that skips the GPU on predictable tokens, an adaptive bias head, llama.cpp backend (7× faster than PyTorch), multi-GPU parallel compression, and a binary file format (NC06) — the first LLM-based binary compressor we know of.
Runs on a GTX 1050 Ti. ~500 MB weights, ~1.2 GB VRAM per worker.
Try it, break it, share your results — all feedback welcome. ⭐ on the repo appreciated!
Results across all systems we tested: - alice29.txt → 0.918 bpb (−44% vs CMIX, −20% vs ts_zip) — below the 2nd-order Shannon entropy bound - enwik8 (100 MB) → 0.9389 bpb (−8% vs FineZip/LLMZip's 8B model, −15% vs ts_zip) - Unseen text → 0.723 bpb on a doc published after training cutoff — no memorization, 26% better than FineZip/LLMZip on the same model
In 2017, my RNNs were babbling. Today, they are hallucinating beautifully.
10 years ago, getting an LSTM to output coherent English was a struggle. 10 years later, after a "cure" based on FineWeb-EDU and a custom synthetic mix for causal conversation, the results are fascinating.
We trained this on ~10B tokens on a single AMD GPU (ROCm). It is not a Transformer: Echo-DSRN (400M) is a novel recurrent architecture inspired by Hymba, RWKV, and xLSTM, designed to challenge the "Attention is All You Need" monopoly on the Edge.
The ambitious goal is to build a small instruct model with RAG and tool usage capabilities (ethicalabs/Kurtis-EON1)
📊 The Benchmarks (Size: 400M)
For a model this size (trained on <10B tokens), the specialized performance is surprising:
*SciQ*: 73.8% 🦄 (This rivals billion-parameter models in pure fact retrieval). *PIQA*: 62.3% (Solid physical intuition for a sub-1B model).
The Reality Check:
HellaSwag (29.3%) and Winogrande (50.2%) show the limits of 400M parameters and 10B tokens training.
We are hitting the "Reasoning Wall" which confirms we need to scale to (hopefully) unlock deeper common sense. As you can see in the visualization (to be released soon on HF), the FineWeb-EDU bias is strong. The model is convinced it is in a classroom ("In this course, we explore...").
The Instruct Model is not ready yet and we are currently using curriculum learning to test model plasticity.
Source code and weights will not be released yet. This is not a fork or a fine-tune: the base model is built in-house at https://www.ethicalabs.ai/, with novel components that do not exist in current open libraries.
🤝 Call for Collaboration: I am looking for Peer Reviewers interested in recurrent/hybrid architectures. If you want to explore what lies beyond Transformers, let’s connect!
The model: ✅ Scraped my GitHub repos automatically ✅ Pulled my experience from LinkedIn ✅ Designed an Aurora Glass theme ✅ Mapped every skill to projects ✅ Added animations I'd never code myself
bumped into one of the OG reads today!! handwriting generation & synthesis is still my favorite application of RNNs - supper amazed at how such a small model (3.6M params), trained overnight on cpu could reach such peak performance. Huge credit to the data (IAM-OnDB🔥) which was meticulously curated using an infra-red device to track pen position. Try demo here: https://www.calligrapher.ai/ Code: https://github.com/sjvasquez/handwriting-synthesis
All your GPUs you have suddenly vanish and any devices with a GPU turn to CPU only
Every time you use a HF space you randomly start dancing for 5 minutes OR Every HF model you’ve hearted / liked suddenly disappears and becomes inaccessible
RealHarm: A Collection of Real-World Language Model Application Failure
I'm David from Giskard, and we work on securing your Agents. Today, we are launching RealHarm: a dataset of real-world problematic interactions with AI agents, drawn from publicly reported incidents.
If you've followed the progress of robotics in the past 18 months, you've likely noticed how robotics is increasingly becoming the next frontier that AI will unlock.
At Hugging Face—in robotics and across all AI fields—we believe in a future where AI and robots are open-source, transparent, and affordable; community-built and safe; hackable and fun. We've had so much mutual understanding and passion working with the Pollen Robotics team over the past year that we decided to join forces!
You can already find our open-source humanoid robot platform Reachy 2 on the Pollen website and the Pollen community and people here on the hub at
I want to share my work of creating a reasoning mamba model
I used GRPO over Falcon3 Mamba Instruct to make this model. It generates blazing fast response while building good logic to answer challenging questions.
O1 Embedder: Transforming Retrieval Models with Reasoning Capabilities
Researchers from University of Science and Technology of China and Beijing Academy of Artificial Intelligence have developed a novel retrieval model that mimics the slow-thinking capabilities of reasoning-focused LLMs like OpenAI's O1 and DeepSeek's R1.
Unlike traditional embedding models that directly match queries with documents, O1 Embedder first generates thoughtful reflections about the query before performing retrieval. This two-step process significantly improves performance on complex retrieval tasks, especially those requiring intensive reasoning or zero-shot generalization to new domains.
The technical implementation is fascinating:
- The model integrates two essential functions: Thinking and Embedding - It uses an "Exploration-Refinement" data synthesis workflow where initial thoughts are generated by an LLM and refined by a retrieval committee - A multi-task training method fine-tunes a pre-trained LLM to generate retrieval thoughts via behavior cloning while simultaneously learning embedding capabilities through contrastive learning - Memory-efficient joint training enables both tasks to share encoding results, dramatically increasing batch size
The results are impressive - O1 Embedder outperforms existing methods across 12 datasets in both in-domain and out-of-domain scenarios. For example, it achieves a 3.9% improvement on Natural Questions and a 3.0% boost on HotPotQA compared to models without thinking capabilities.
This approach represents a significant paradigm shift in retrieval technology, bridging the gap between traditional dense retrieval and the reasoning capabilities of large language models.
What do you think about this approach? Could "thinking before retrieval" transform how we build search systems?
