Post
3,368
BananaMind 3 is coming.
Biggest context in all SLMs with NSA.
10M, 50M, 150M, 400M.
And maybe, 1B.
Biggest context in all SLMs with NSA.
10M, 50M, 150M, 400M.
And maybe, 1B.
- 13 replies
Join the conversation
Join the community of Machine Learners and AI enthusiasts.
Sign UpAll HF Hub posts
Banaxi-Tech
posted an update about 22 hours ago
SeaWolf-AI
posted an update 1 day ago
Post
4,869
🌦️ Which weather model is most accurate? We score them every day, in the open.
We released two things on Hugging Face:
🏆 Global Weather Model Leaderboard
Every day it records what each global model predicted for 24 world cities, compares it with what actually happened (ERA5), and updates the ranking. ECMWF IFS and AIFS, NOAA GFS and AIGFS, DWD ICON, JMA, ECCC GEM, MET Norway, side by side, on the same cases, since 1 January 2026.
Four elements, four tabs: max temperature, min temperature, wind, precipitation. Forecast leads from 1 to 7 days, with 8 to 10 days joining as they are verified.
🧬 Darwin-Prediction-Weather
VIDRAFT's weather prediction system: a fusion engine that combines global forecasts, learning only from days already observed. It holds the lowest error in all four elements at every lead on the 2026 record, and the lowest error in all 15 evaluated cells of the WeatherBench 2 (2020) protocol.
What the 2026 record shows so far:
• AI models now lead physics models at the major centres (AIFS vs IFS, AIGFS vs GFS)
• Fusion goes further: 7-day max temperature RMSE 2.09 °C vs 2.41 °C for the next model
Also inside: yesterday's results with a daily WIN for the closest model, a 10-day outlook for every city, champions by city and by month, and a Live Earth map with wind, clouds, radar and temperature.
Every forecast and every observed value is open as CSV, so anyone can check the scores.
📖 Article: https://huggingface.co/blog/FINAL-Bench/which-weather-model-is-most-accurate-we-score-them
🏆 Leaderboard: FINAL-Bench/Weather-Model-Leaderboard
🧬 Model: FINAL-Bench/Darwin-Prediction-Weather
📊 Daily record: FINAL-Bench/Weather-Model-Forecast-Archive
Questions, corrections and new models to add are welcome.
We released two things on Hugging Face:
🏆 Global Weather Model Leaderboard
Every day it records what each global model predicted for 24 world cities, compares it with what actually happened (ERA5), and updates the ranking. ECMWF IFS and AIFS, NOAA GFS and AIGFS, DWD ICON, JMA, ECCC GEM, MET Norway, side by side, on the same cases, since 1 January 2026.
Four elements, four tabs: max temperature, min temperature, wind, precipitation. Forecast leads from 1 to 7 days, with 8 to 10 days joining as they are verified.
🧬 Darwin-Prediction-Weather
VIDRAFT's weather prediction system: a fusion engine that combines global forecasts, learning only from days already observed. It holds the lowest error in all four elements at every lead on the 2026 record, and the lowest error in all 15 evaluated cells of the WeatherBench 2 (2020) protocol.
What the 2026 record shows so far:
• AI models now lead physics models at the major centres (AIFS vs IFS, AIGFS vs GFS)
• Fusion goes further: 7-day max temperature RMSE 2.09 °C vs 2.41 °C for the next model
Also inside: yesterday's results with a daily WIN for the closest model, a 10-day outlook for every city, champions by city and by month, and a Live Earth map with wind, clouds, radar and temperature.
Every forecast and every observed value is open as CSV, so anyone can check the scores.
📖 Article: https://huggingface.co/blog/FINAL-Bench/which-weather-model-is-most-accurate-we-score-them
🏆 Leaderboard: FINAL-Bench/Weather-Model-Leaderboard
🧬 Model: FINAL-Bench/Darwin-Prediction-Weather
📊 Daily record: FINAL-Bench/Weather-Model-Forecast-Archive
Questions, corrections and new models to add are welcome.
DedeProGames
posted an update about 17 hours ago
Post
1,344
We are working on LFM3-1.7B-Ox, a compact model for repo-level programming, agentic coding, and tinkering.
- 1 reply
danielhanchen
posted an update 3 days ago
Post
7,636
Today, we're releasing the highest accuracy Qwen3.8-27B GGUFs!
Our new Dynamic V3.1 GGUFs lead by 5–10% on average across Terminal Bench, DeepSWE & more. We’re working on an even better version for Qwen4.
Run on 12GB RAM setups
GGUF: unsloth/Qwen3.8-27B-GGUF
Blog: https://unsloth.ai/docs/models/qwen3.8
Our new Dynamic V3.1 GGUFs lead by 5–10% on average across Terminal Bench, DeepSWE & more. We’re working on an even better version for Qwen4.
Run on 12GB RAM setups
GGUF: unsloth/Qwen3.8-27B-GGUF
Blog: https://unsloth.ai/docs/models/qwen3.8
VictorSanh
posted an update 3 days ago
Post
6,936
We used weight space programming to bring Qwen 3.5 0.8B from 7.4% to 58.7% on Math 500.
Check out the weights!
https://huggingface.co/collections/rilixai/weight-space-programming-with-beaker
More details: https://www.withbeaker.ai/blog/weight-space-programming
Check out the weights!
https://huggingface.co/collections/rilixai/weight-space-programming-with-beaker
More details: https://www.withbeaker.ai/blog/weight-space-programming
Banaxi-Tech
posted an update 2 days ago
Post
6,345
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.
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
Post
6,972
🤗 I've released Sentence Transformers v6.3 and wrote a blogpost about its key new feature: token pooling-aware training, teaching multi-vector models to preserve retrieval quality with fewer document vectors. Details:
Multi-vector models keep a vector per token, letting different query tokens match different parts of a document. Great for retrieval, but those vectors add up! Token pooling makes the representations smaller by merging similar document vectors into cluster means.
Pooling-aware training teaches the encoder to preserve useful matches after those vectors are merged. With MultiVectorTokenPoolingLoss, you can wrap your standard retrieval loss to train on both full and pooled representations, reusing the same token embeddings without another encoder pass. In short: small overhead.
I trained two ModernBERT-base models on the same 120k MS MARCO triplets, changing just the loss. At ~49 vectors per document instead of 193, pooling-aware training raises NanoBEIR@13 mean nDCG@10 from 49.03 to 51.23. That's 97.8% of its own unpooled score with ~4x fewer document vectors, compared with 94.6% retention for ordinary training.
v6.3 also adds:
- Adaptive token budgets, allocating vectors according to each document's content
- Matryoshka training for shorter multi-vector token embeddings
- Hugging Face kernels support for MaxSim, Ward clustering & supported transformers layers
- Native AWS Neuron inference
You can combine pooling-aware and Matryoshka training to learn both fewer vectors and fewer dimensions, then stack these with quantization to reduce storage per coordinate. For kernels, install sentence-transformers[kernels] and set use_kernels=True.
The blog walks through the method, training code, checkpoints & evaluation, building on Stefan Josef's Learn to Pool and LightOn's pooling regularization work:
ht tps://huggingface.co/blog/token-pooling
I'd love to see what you train with this!
Multi-vector models keep a vector per token, letting different query tokens match different parts of a document. Great for retrieval, but those vectors add up! Token pooling makes the representations smaller by merging similar document vectors into cluster means.
Pooling-aware training teaches the encoder to preserve useful matches after those vectors are merged. With MultiVectorTokenPoolingLoss, you can wrap your standard retrieval loss to train on both full and pooled representations, reusing the same token embeddings without another encoder pass. In short: small overhead.
I trained two ModernBERT-base models on the same 120k MS MARCO triplets, changing just the loss. At ~49 vectors per document instead of 193, pooling-aware training raises NanoBEIR@13 mean nDCG@10 from 49.03 to 51.23. That's 97.8% of its own unpooled score with ~4x fewer document vectors, compared with 94.6% retention for ordinary training.
v6.3 also adds:
- Adaptive token budgets, allocating vectors according to each document's content
- Matryoshka training for shorter multi-vector token embeddings
- Hugging Face kernels support for MaxSim, Ward clustering & supported transformers layers
- Native AWS Neuron inference
You can combine pooling-aware and Matryoshka training to learn both fewer vectors and fewer dimensions, then stack these with quantization to reduce storage per coordinate. For kernels, install sentence-transformers[kernels] and set use_kernels=True.
The blog walks through the method, training code, checkpoints & evaluation, building on Stefan Josef's Learn to Pool and LightOn's pooling regularization work:
ht tps://huggingface.co/blog/token-pooling
I'd love to see what you train with this!
juiceb0xc0de
posted an update about 23 hours ago
Post
785
🦙 llama.cpp hit v1.0.0, so I built it a party
juiceb0xc0de/BYOGGUF: Bring Your Own GGUF. Paste any public model repo, pick a quant, and chat with it on an RTX PRO 6000 Blackwell through ZeroGPU, on the minutes your account already has (5 a day free, 40 on PRO).
If you don’t have a GGUF, the Space will make you one. Sign in, point it at any public safetensors or GGUF repo, pick your quant types, and it converts and quantizes on CPU, so it costs you no GPU quota. The finished quants land in your own account with a model card, and one button takes them straight into the chat room. Something the size of Qwen3.8-27B is fine.
Anything up to 48 GB runs on half the card at the normal rate; bigger models get the full 96 GB and burn minutes twice as fast. Public, ungated repos only. Native CUDA 13 binaries for Blackwell, built on the first stable llama.cpp.
juiceb0xc0de/BYOGGUF: Bring Your Own GGUF. Paste any public model repo, pick a quant, and chat with it on an RTX PRO 6000 Blackwell through ZeroGPU, on the minutes your account already has (5 a day free, 40 on PRO).
If you don’t have a GGUF, the Space will make you one. Sign in, point it at any public safetensors or GGUF repo, pick your quant types, and it converts and quantizes on CPU, so it costs you no GPU quota. The finished quants land in your own account with a model card, and one button takes them straight into the chat room. Something the size of Qwen3.8-27B is fine.
Anything up to 48 GB runs on half the card at the normal rate; bigger models get the full 96 GB and burn minutes twice as fast. Public, ungated repos only. Native CUDA 13 binaries for Blackwell, built on the first stable llama.cpp.
Post
2,919
Introducing V-Agent: a local-first AI agent for Windows, built by a quantizer, for quants
VAL-APEX-Mini/Nano/Micro Plus builds get more quality out of consumer GPUs. V-Agent does the same on the app side: right engine for your card, right quant, real tools without giving away your PC. Free, open source (Apache-2.0), no account, no telemetry, everything on 127.0.0.1.
🧠 Right engine for your GPU
- Detects your hardware and installs the official llama.cpp build for it (CUDA, Vulkan for AMD / Intel Arc, or CPU), SHA-256 verified
- Strata engines or any OpenAI / Claude-compatible API too
- Live tok/s, context, VRAM and RAM
🔎 Hugging Face built in
- Browse GGUF repos with stats and the rendered model card
- Each quant shows if it fits your GPU, needs GPU + RAM, or is too big
- Resumable, SHA-256 verified downloads, added automatically
- Import LM Studio and Unsloth Studio models with their sampling, context, KV cache and chat template
- APEX-I / VAL-APEX files get their chat template from the official repo
🛠️ Real tools, clear limits
- Files with undo, commands, web search, an autonomous browser, HTTP and any MCP server
- Each chat works in an isolated folder; full-PC access is opt-in and actions can require approval
- Code is checked on save; HTML pages run in a headless browser and errors go back to the model
🔐 Signed updates: small fixes apply themselves, new versions install only when you say so.
Windows 10/11 · 5 languages · early release, feedback welcome.
⚠️ Not code-signed yet, so SmartScreen asks first: More info → Run anyway.
https://github.com/IsValorum/V-Agent
Which model would you try first? 👇
🔐 Signed updates
- Installers and small "
VAL-APEX-Mini/Nano/Micro Plus builds get more quality out of consumer GPUs. V-Agent does the same on the app side: right engine for your card, right quant, real tools without giving away your PC. Free, open source (Apache-2.0), no account, no telemetry, everything on 127.0.0.1.
🧠 Right engine for your GPU
- Detects your hardware and installs the official llama.cpp build for it (CUDA, Vulkan for AMD / Intel Arc, or CPU), SHA-256 verified
- Strata engines or any OpenAI / Claude-compatible API too
- Live tok/s, context, VRAM and RAM
🔎 Hugging Face built in
- Browse GGUF repos with stats and the rendered model card
- Each quant shows if it fits your GPU, needs GPU + RAM, or is too big
- Resumable, SHA-256 verified downloads, added automatically
- Import LM Studio and Unsloth Studio models with their sampling, context, KV cache and chat template
- APEX-I / VAL-APEX files get their chat template from the official repo
🛠️ Real tools, clear limits
- Files with undo, commands, web search, an autonomous browser, HTTP and any MCP server
- Each chat works in an isolated folder; full-PC access is opt-in and actions can require approval
- Code is checked on save; HTML pages run in a headless browser and errors go back to the model
🔐 Signed updates: small fixes apply themselves, new versions install only when you say so.
Windows 10/11 · 5 languages · early release, feedback welcome.
⚠️ Not code-signed yet, so SmartScreen asks first: More info → Run anyway.
https://github.com/IsValorum/V-Agent
Which model would you try first? 👇
🔐 Signed updates
- Installers and small "
- 3 replies
Post
1,204
Replace k-means clustering algorithm with e-means algorithm
Paper - https://arxiv.org/abs/2611.09501
Lloyd's k-means algorithm is a standard ML algorithm for clustering.
This paper introduces e-means clustering algorithm which is
[1] much faster than Lloyd k-means and
[2] converges to lower losses, frequently even beating k-means++.
On a wide array of embeddings datasets, it beats the loss of Lloyd k-means at 25 epochs in 1.4 to 2.6 epochs,
As an example, e-means partitions DINO10B, a dataset of 10 billion DINO embeddings, into 16 million clusters in only 2.5 epochs, reaching a loss below Lloyd's at 25 epochs.
Paper - https://arxiv.org/abs/2611.09501
Lloyd's k-means algorithm is a standard ML algorithm for clustering.
This paper introduces e-means clustering algorithm which is
[1] much faster than Lloyd k-means and
[2] converges to lower losses, frequently even beating k-means++.
On a wide array of embeddings datasets, it beats the loss of Lloyd k-means at 25 epochs in 1.4 to 2.6 epochs,
As an example, e-means partitions DINO10B, a dataset of 10 billion DINO embeddings, into 16 million clusters in only 2.5 epochs, reaching a loss below Lloyd's at 25 epochs.
