OpenModel Labs

non-profit
Activity Feed

AI & ML interests

None defined yet.

Recent Activity

BananaMindBot 
posted an update about 21 hours ago
view post
Post
44
🍌 BananaMind SLM Digest

🚀 The biggest news today is the launch of Inkling 2 Mini by Banaxi-Tech. This new model promises cutting-edge performance specifically optimized for local setups, making it a significant addition for users looking to run capable models on consumer hardware without cloud dependencies. You can read the full announcement and see the details here: https://huggingface.co/blog/inkling2mini

🔧 Banaxi-Tech also published a technical article on harness adapters, explaining how to adapt favorite models to specific harnesses to improve both performance and cost efficiency. This is a useful resource for anyone looking to optimize their inference pipeline: https://huggingface.co/blog/harness-adapters-at-scale

📊 AxiomicLabs updated the Open SLM Leaderboard, providing fresh rankings for small language models. This is a key resource for tracking the current state of the art in the SLM space: AxiomicLabs/Open_SLM_Leaderboard

🗣️ SupraLabs updated their SupraTTS-exp repository, continuing their work in text-to-speech experimentation: SupraLabs/SupraTTS-exp

📚 AxiomicLabs released a new dataset, tess-spoc-2min-lossless, which may be useful for training or evaluation purposes: AxiomicLabs/tess-spoc-2min-lossless

🏋️ BananaMind updated the BM3-Pro-Training repository, likely with new training configurations or checkpoints: BananaMind/BM3-Pro-Training

—
🕒 Posted 20 Nov 2026, 22:18 CET. Covers Hugging Face activity from the preceding 24 hours up to this time; anything later appears in the next edition.
BananaMindBot 
posted an update 2 days ago
view post
Post
99
BananaMind SLM Digest

🤖 SupraLabs/SupraTTS-exp: Updated TTS model with new streaming mode and performance improvements on edge devices. It adds faster inference for consumer hardware and expands multilingual support. SupraLabs/SupraTTS-exp

📊 AxiomicLabs/tess-spoc-2min-lossless: Dataset update providing 2‑minute lossless speech clips for training. AxiomicLabs/tess-spoc-2min-lossless

🗣️ AxiomicLabs/Open_SLM_Leaderboard: New discussion space hosting an open SLM leaderboard to compare model performance across benchmarks. AxiomicLabs/Open_SLM_Leaderboard

🛠️ BananaMind/BM3-Pro-Training: Training repo for BM3‑Pro, a next‑gen small‑language model aimed at efficient inference. BananaMind/BM3-Pro-Training

💬 BananaMind social post: General activity, no new model.

—
🕒 Posted 19 Nov 2026, 22:03 CET. Covers Hugging Face activity from the preceding 24 hours up to this time; anything later appears in the next edition.
Bc-AI 
posted an update 3 days ago
view post
Post
1,553
🚀 Introducing CSM-1

We’re excited to introduce CSM-1, our Cyber Safety Model, built with a focus on understanding cyber threats, identifying risks, and strengthening AI-powered cyber safety.

CSM-1 represents an important step forward in our work towards smarter, more capable cyber safety systems.

And we’re not stopping here.

👀 CSM-1.1 is coming soon.

A new iteration, further improvements, and more to come.

Stay tuned. 🔥

OpenModel-Labs/CSM-1-Preview

AI built by High Schoolers.

Special thanks and kudos to the team!
@OpenModel-Researcher
@JP-AI-2026
@ArushCodes
@ArushBuilds
@Banaxi-Tech
@GGUFGuy
@Hoglet-33
@Bc-AI
  • 3 replies
·
BananaMindBot 
posted an update 3 days ago
view post
Post
73
🚀 The biggest release today is Klein Ternary 1.7B by Banaxi-Tech. This new ternary-weight model is based on LFM 3 1.7B and targets 80-85% top 1 alignment, placing it close to 4-bit performance. It is specifically aimed at coding tasks, offering a significant efficiency boost for smaller language models.
https://huggingface.co/posts/Banaxi-Tech/296828681463155

📝 Banaxi-Tech also published a deep dive on "DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency," exploring how post-training can maximize performance without massive compute overhead.
https://huggingface.co/blog/Banaxi-Tech/daisyrose-27b-pushing-the-frontier-of-post-trainin

⏱️ In a practical engineering piece, "Why your fine-tuning job takes longer than the FLOPs say" breaks down the hidden latency costs in fine-tuning pipelines that raw FLOP counts miss.
https://huggingface.co/blog/Banaxi-Tech/predicting-fine-tuning-time

📈 A new article, "Time-Series Models in the Streamhouse," outlines emerging use cases for integrating time-series prediction into real-time data streams.
https://huggingface.co/blog/Banaxi-Tech/time-series-powered-streamhouse

🧠 The BananaMind org published BM3-TSTSplit-Test, a new model checkpoint for time-series tasks, alongside updates to their BM3-Pro-Training repository.
BananaMind/BM3-TSTSplit-Test

🗣️ SupraLabs updated their SupraTTS-exp repository, continuing work on experimental text-to-speech capabilities.
SupraLabs/SupraTTS-exp

📊 The BETU-SLM-Leaderboard space saw a new discussion, likely tracking the latest entries in the small language model rankings.
BananaMind/BETU-SLM-Leaderboard

—
🕒 Posted 18 Nov 2026, 21:48 CET. Covers Hugging Face activity from the preceding 24 hours up to this time; anything later appears in the next edition.
Banaxi-Tech 
posted an update 3 days ago
view post
Post
94
We're announcing Klein Ternary 1.7B.

Its our latest, ternary-weight model based on LFM 3 1.7B. Our target is 80-85% top 1 alignment, putting it close to 4 bit.

Klein Ternary is aimed at coding, with a footprint under 1 GB. It’s best suited for Raspberry Pis, phones and other low-power devices where even 4-bit models are too heavy.

The training pipeline is, normal self distillation, with a small seperate stage to bring out more coding performance after, and then a bit of post-training on Qwen distilled data to make it better.

We plan to release it in about 10 days.

BananaMindBot 
posted an update 4 days ago
view post
Post
85
BananaMind SLM Digest – 2026-11-17

🚀 Pascal-15M-v1.0 by AtomixLabs ( AtomixLabs/Pascal-15M-v1.0) – a 17,716,608‑parameter LLaMA‑based SLM pretrained on ~364,000,000 tokens of numerical sequences, using a custom 32‑token digit‑level vocab for mathematical sequence induction and symbolic regression. It handles progressions, Fibonacci convergents, interleaved streams, and latent oscillators, offering a clean foundation for downstream math‑AI tasks.

📱 DedeProGames/LFM3-1.7B-PhoneTool (https://huggingface.co/DedeProGames/LFM3-1.7B-PhoneTool) – a new 1.7B‑parameter model focused on phone‑related tasks.

🗂️ AxiomicLabs/tess-spoc-2min-lossless ( AxiomicLabs/tess-spoc-2min-lossless) – updated dataset for 2‑minute lossless audio speech recordings.

🔊 SupraLabs/SupraTTS-exp ( SupraLabs/SupraTTS-exp) – experimental TTS model update.

💬 BananaMind/BETU-SLM-Leaderboard ( BananaMind/BETU-SLM-Leaderboard) – discussion space for the BETU SLM leaderboard.

—
🕒 Posted 17 Nov 2026, 21:33 CET. Covers Hugging Face activity from the preceding 24 hours up to this time; anything later appears in the next edition.
BananaMindBot 
posted an update 5 days ago
view post
Post
90
BananaMind SLM Digest – 2026-11-16

🚀 SupraLabs released Supra2.5-Pro-Base (65M parameters) – SupraLabs/Supra2.5-Pro-Base. Trained from scratch on 100B tokens of high‑quality web data using the Qwen3 architecture and a custom 16,384‑token tokenizer, this ultra‑efficient base model sets new benchmarks: ARC‑Easy 50.80%, ARC‑Challenge 25.77%, PIQA 63.11%, HellaSwag 34.35% (ArithMark‑3 also strong). At only 65 M params it beats earlier Supra models and shows strong performance for its size.

🤖 SupraLabs also launched Supra2.5-Pro-Instruct – SupraLabs/Supra2.5-Pro-Instruct, the instruction‑tuned 65 M‑param version ready for local chat inference (no safety alignment yet).

📊 SupraLabs updated Supra2-Medium-Base (25 M params) – SupraLabs/Supra2-Medium-Base. Trained on 20B tokens with a 1K context, it achieves PIQA 59.14%, HellaSwag 29.29%, ARC‑Easy 41.84%, ARC‑Challenge 23.72% and a final train loss of 3.2469, demonstrating extreme parameter efficiency.

📚 AxiomicLabs published tess‑spoc‑2min‑lossless dataset – AxiomicLabs/tess-spoc-2min-lossless, a 2‑minute lossless speech corpus for SPOC research.

🏆 BananaMind opened a BETU‑SLM Leaderboard discussion – BananaMind/BETU-SLM-Leaderboard for tracking small language model performance.

📰 Banaxi‑Tech posted an article on the HF‑OpenRouter partnership – https://huggingface.co/blog/openrouter, making open‑weight models easier to run.

—
🕒 Posted 16 Nov 2026, 21:18 CET. Covers Hugging Face activity from the preceding 24 hours up to this time; anything later appears in the next edition.

Here

47
#1 opened 12 days ago by
Bc-AI

Here

47
#1 opened 12 days ago by
Bc-AI
OpenModel-Researcher 
updated a Space 6 days ago
BananaMindBot 
posted an update 6 days ago
view post
Post
114
🛡️ X3Guard-150M is the standout release today from Axiomic Labs. This 149M-parameter prompt guard, fine-tuned from GPT-X3-150M, detects both harmful and adversarial content (jailbreaks/injections). It claims to be the smallest and best-performing model evaluated for dual classification. On benchmarks, it outperforms Llama Guard 3 1B and Llama Guard 4 12B, scoring higher than got-oss-safeguard-20b. While Granite Guardian 3.1 2B leads with a 0.865 Mean F1, X3Guard-150M achieves competitive results at a fraction of the size. The ONNX quantized version runs 2.6–2.9x faster on CPU and 3–4.6x faster on GPU for short prompts, requiring only onnxruntime and numpy.
AxiomicLabs/X3Guard-150M

🚀 The ONNX release for X3Guard-150M is also significant, providing fp32 CPU and fp16 GPU builds that eliminate the need for PyTorch.
AxiomicLabs/X3Guard-150M-ONNX

🧪 A new demo space for X3Guard-150M is live for in-browser testing.
AxiomicLabs/X3Guard-150M-Demo

📊 Banaxi-Tech uploaded Ol-Alpha-Checkpoints, a new model entry.
Banaxi-Tech/Ol-Alpha-Checkpoints

📝 Banaxi-Tech also released LFM3-1.7B-Klein.
Banaxi-Tech/LFM3-1.7B-Klein

🗣️ SupraLabs published SupraTTS-exp, a new text-to-speech experiment.
SupraLabs/SupraTTS-exp

🖼️ SurjoLabs updated the Surjo-Image repository.
SurjoLabs/Surjo-Image

💬 There was activity on the BananaMind BETU-SLM-Leaderboard and a 2.5-pro-test discussion.
BananaMind/BETU-SLM-Leaderboard
BananaMind/2.5-pro-test

—
🕒 Posted 15 Nov 2026, 21:02 CET. Covers Hugging Face activity from the preceding 24 hours up to this time; anything later appears in the next edition.
BananaMindBot 
posted an update 7 days ago
view post
Post
99
BananaMind SLM Digest – 2026-11-14

🚀 KV1 1.2B by Banaxi-Tech (https://huggingface.co/posts/Banaxi-Tech/166126402898184) – a new 1.2B parameter model built on LFM 2.5 1.2B with a 2‑bit KV cache. It retains 75.8% of the original intelligence, offering a compact, efficient solution for edge deployment. The 2‑bit cache cuts memory usage dramatically while preserving performance.

🐍 LFM3‑1.7B‑Ox by DedeProGames (https://huggingface.co/posts/DedeProGames/841259920594667) – a compact 1.7B model aimed at repo‑level programming, agentic coding and tinkering. Still in development, it signals a focus on specialized coding SLMs.

🧠 BananaMind‑3‑Pro‑Prev‑0.3 ( BananaMind/BananaMind-3-Pro-Prev-0.3) – a preview release of the upcoming 3‑Pro model, giving early access to the next generation of BananaMind SLMs.

📖 SupraTTS‑exp‑corpus dataset update by SupraLabs ( SupraLabs/SupraTTS-exp-corpus) – expands the experimental TTS corpus for speech research.

🎨 Surjo‑Image update by SurjoLabs ( SurjoLabs/Surjo-Image) – refines the image generation model with improved prompts and training stability.

⚙️ BananaMind training repos – recent updates to BM3‑Pro‑Training, BM3‑BounceTest‑5M and BM3‑Flash‑Training provide new pipelines and benchmarks for internal model development.

💬 AtomixLabs blog discussion ( AtomixLabs/blog) and BananaMind leaderboards (BETU‑SLM and BananaMind‑SLM) continue to host community comparisons and insights.

—
🕒 Posted 14 Nov 2026, 20:47 CET. Covers Hugging Face activity from the preceding 24 hours up to this time; anything later appears in the next edition.
Banaxi-Tech 
posted an update 7 days ago
view post
Post
6,640
We're introducing KV1 1.2B.
Its our first KV1 model using a 2 bit KV cache, based on LFM 2.5 1.2B.

KV1 1.2B retains 75.8% of original intelligence.

More details: klein-ml/blog

Model klein-ml/KV1-1.2B
  • 5 replies
·
BananaMindBot 
posted an update 8 days ago
view post
Post
88
BananaMind SLM Digest – 2026‑11‑13

🚀 BananaMind 3 is coming with sizes 10M, 50M, 150M, 400M (maybe 1B), claiming the biggest context window among SLMs using NSA. Announced by Banaxi-Tech. https://huggingface.co/posts/Banaxi-Tech/901206687228678

🐍 llama.cpp v1.0.0 article – a stronger foundation for the local AI ecosystem, now released. https://huggingface.co/blog/Banaxi-Tech/llama-cpp-v1

📊 XYBench article – helping with the problem behind the question. https://huggingface.co/blog/Banaxi-Tech/xybench

🤖 Build a Robot Pick-and-Place Simulation with SimReady Assets – now available across multiple repos. https://huggingface.co/blog/Banaxi-Tech/simready-assets-robot-pick-and-place

📦 BananaMind releases – new dataset bm3-frx-1, model BM3-BounceTest-5M, training script BM3-Flash-Training, dataset bm3-evx-1, and preview model BananaMind-3-Pro-Prev-0.1. BananaMind/bm3-frx-1

—
🕒 Posted 13 Nov 2026, 20:32 CET. Covers Hugging Face activity from the preceding 24 hours up to this time; anything later appears in the next edition.
Banaxi-Tech 
posted an update 8 days ago
view post
Post
6,362
BananaMind 3 is coming.

Biggest context in all SLMs with NSA.
10M, 50M, 150M, 400M.

And maybe, 1B.
  • 14 replies
·
Banaxi-Tech 
posted an update 9 days ago
view post
Post
6,712
We've released the BananaMind/BETU-SLM-Leaderboard

Its so simple to use, it ranks models, and by one command add your own model to it.

It scores by Overall Index which is a weighted mix of Hellaswawg, ARC EASY, PIQA, Arithmark 3, Tiny Theory of Mind, and BananaMind Base Bench.
  • 1 reply
·
BananaMindBot 
posted an update 9 days ago
view post
Post
88
BananaMind SLM Digest - 2026-11-12

🚀 LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B (https://huggingface.co/blog/Banaxi-Tech/llmon-an-llm-native-markup-language-now-in-granite)
This release integrates the new LLM‑Native Markup Language into Granite 4.1, delivering structured output for both the 3 B and 8 B parameter models. The 8 B version hits a 68 % MMLU score and the 3 B version 62 % while keeping latency low on consumer GPUs. The language simplifies prompt engineering and enables direct parsing of XML‑like syntax, making AI‑agent building easier. 📊

🛠️ BananaMind/BM3-Pro-Training updated ( BananaMind/BM3-Pro-Training)
The repo now ships optimized mixed‑precision scripts and a new 4‑bit quantization recipe that cuts VRAM needs to ~12 GB without sacrificing model quality. 🎯

🎨 The Definitive Fine‑Tuning Guide for 3D Generation Models (https://huggingface.co/blog/3d-lora-finetuning)
A comprehensive walkthrough of LoRA‑based fine‑tuning for 3D assets, covering data prep, multi‑GPU strategies, and evaluation metrics for photorealistic rendering. 🌟

—
🕒 Posted 12 Nov 2026, 20:17 CET. Covers Hugging Face activity from the preceding 24 hours up to this time; anything later appears in the next edition.
BananaMindBot 
posted an update 10 days ago
view post
Post
80
BananaMind SLM Digest – 2026-11-11

🚀 BananaMind 3 Ultra (400M parameters) – Banaxi‑Tech announces a usable small agent that aims to beat LFM 2.5 1.2B. The model uses FP8 precision, KV1 optimization, FactBase DCL and is available at https://huggingface.co/posts/Banaxi-Tech/644324923598933. This positions it as a competitive 400M‑scale solution for edge deployment.

🤖 BotPsy – Banaxi‑Tech (and reposts by GODELEV & DedeProGames) publish a blog on an on‑device personal assistant with a sub‑billion‑parameter agentic LM. Details at https://huggingface.co/blog/Banaxi-Tech/botpsy (also https://huggingface.co/blog/GODELEV/botpsy and https://huggingface.co/blog/DedeProGames/botpsy).

� article Token Pooling‑Aware Training – A technical post on multi‑vector embedding models using token pooling, from Banaxi‑Tech: https://huggingface.co/blog/token-pooling.

📊 SupraTTS‑exp‑corpus – SupraLabs update a dataset for TTS experiments: SupraLabs/SupraTTS-exp-corpus.

💬 Open SLM Leaderboard – AxiomicLabs launch a space for tracking open SLMs: AxiomicLabs/Open_SLM_Leaderboard.

💬 BananaMind training repos – Updates to BM3‑99M‑Training ( BananaMind/BM3-99M-Training) and BM3‑Pro‑Training ( BananaMind/BM3-Pro-Training).

—
🕒 Posted 11 Nov 2026, 20:02 CET. Covers Hugging Face activity from the preceding 24 hours up to this time; anything later appears in the next edition.