Hugging Face's logo

Join the conversation

Join the community of Machine Learners and AI enthusiasts.

Sign Up

All HF Hub posts
SeaWolf-AI 
posted an update 3 days ago
view post
Post
8,094
SuperCLUE’s November 2026 results highlight why category scores matter when choosing a model.

Across the 19 published entries, here are the highest displayed scores in each category, followed by the leader among models labeled open:

• Agentic coding: GPT-6.1-Sol (69.23); open: DeepSeek-V4.1-Flash-1017 (65.38).
• Mathematical and scientific reasoning: GPT-6.1-Sol (76.32); open: Darwin-180B-RSI-R3 (73.68).
• Hallucination control: Gemini-4-Argon (83.19); open: Kimi-K3.1-Pro (67.78).
• Precise instruction following: Gemini-4-Argon and GPT-6.1-Sol (68.97 each); open: Kimi-K3.1-Pro (44.83).
• Ultra-long context: Gemini-4-Argon (95.63); open: Kimi-K3.1-Pro (87.25).

On the overall leaderboard, Gemini-4-Argon (73.62) and GPT-6.1-Sol (71.76) have the highest displayed scores, but neither has an official rank number. Kimi-K3.1-Pro (61.46) holds official rank No. 1 and leads the open models overall, followed by DeepSeek-V4.1-Flash-1017 (58.95, official No. 2) and GLM-5.3 (50.66, official No. 3).

These comparisons include entries whose official rank is “-”. Among officially ranked open models only, DeepSeek leads mathematical and scientific reasoning at 71.05. “Open” follows SuperCLUE’s classification, not an independent license review. Scores are benchmark points, not error rates; small differences and results from a Chinese-language-focused evaluation need validation on your own workloads. Detailed evaluation settings and the aggregation formula were not verified for this summary.

Read the full breakdown: https://huggingface.co/blog/FINAL-Bench/superclue-november-2026-category-leaders-open-mode

Source: SuperCLUE results released November 5, 2026
https://www.superclueai.com/homepage
Official November data: https://www.superclueai.com/data/generalboard/2026%E5%B9%B411%E6%9C%88.xlsx
danielhanchen 
posted an update about 24 hours ago
Banaxi-Tech 
posted an update about 7 hours ago
eaddario 
posted an update 1 day ago
view post
Post
2,932
Experimental JQK and global target bits‑per‑weight quantization of Nanbeige/Nanbeige4.3-33B.

Unlike standard llama.cpp quantization that relies on fixed type heuristics (e.g., Q4_K_M), the Target BPW approach automatically optimizes per-tensor precision where it matters the most, and produces high quality models that meet a precise global size target.

Key Advantages:
- Better model quality: using JQK types can improve KLD between 10% to 50% compared to standard QK types, at the same bpw.
- VRAM Maximization: Can generate high quality models sized exactly to fit hardware constraints (e.g., fitting the model into exactly 24GB VRAM).
- Data-Driven Precision: Quantization mix is determined by actual weight error sensitivity rather than hardcoded rules, often yielding better PPL/KLD size trade-offs.

Information about JQK types: eaddario/JQ2_K-JQ6_K-Test-Models

Full benchmarks (PPL, KLD, ARC, GPQA, MMLU, etc.) and methodology in the model's cards.

JQK Quantization: eaddario/Nanbeige4.3-33B-JQK

Global target BPW: eaddario/Nanbeige4.3-33B-GGUF
savoji 
posted an update 1 day ago
view post
Post
784
Idol, a new Apache-2.0 image-to-3D model from @speridlabs , makes a 3D Gaussian splat from one image in 7.2 GB of peak VRAM. It is the #1 trending model on the Hub with 1,057 likes.

It writes a standard 3DGS .ply with up to 1,048,576 Gaussians, including sides the picture never showed, and does background removal and framing itself. The catch: the only quality evidence is the authors' own benchmark, on a repo created 2026-10-28.

📊 The numbers
• Authors' benchmark, 1 x H100: best in 6 of 8 columns vs TripoSplat, TRELLIS and LGM (SSIM 0.726, LPIPS 0.213, DINOv2 0.781, CLIP 0.885)
• LGM wins PSNR (12.46 vs 12.25) and speed (1.1 s). Idol takes 6.8 s vs 7.1 s for TRELLIS and 9.9 s for TripoSplat
• 262,144 Gaussians, 50 steps: 3.4 s on B300, 6.7 s on H100, 19.1 s on RTX 4090, 46.2 s on a 10 GB RTX 3080

🛠️ Try it
• Install idol-sdk, accept the terms of the gated DINOv3 and FLUX.2-dev models, run hf auth login
• python -m idol.sample --image photo.jpg --mode real
• --num_gaussians runs from 32,768 (0.4 MB) to 1,048,576 (12.2 MB)

⚠️ Caveats
• Test set and size are not named, and no independent evaluation exists yet
• The card says "under 2 seconds", but the timing table shows 3.4 s at best
• Single objects only, NVIDIA CUDA required, no failure cases shown

speridlabs/idol
https://speridlabs.com/research/idol

If you run it on your own messy photos, how do the unseen sides hold up?

📚 Read it elsewhere
• The deep dive on Medium: https://medium.com/p/5f660ae6ae56
• The long read on Substack: https://sajjadpakdamansavoji.substack.com/p/speridlabs-idol-builds-3d-splats
• The thread on X: https://x.com/saavoji/status/2119137739580858545
• Every version on sajjadpsavoji.github.io: https://sajjadpsavoji.github.io/blog#speridlabs-idol
  • 2 replies
·
jiteshkprasad 
posted an update about 22 hours ago
view post
Post
1,663
Happy Diwali. For the full launch today: we now have one small speech-to-text model for each of the 22 languages of India's Eighth Schedule. Each one is a file under 30 MB with English in the same file, open weights, Apache-2.0, built on Cactus Whistle, and it all runs offline.

If you speak one of these, you can use it in the browser on your own device. If you build apps, you can integrate it, and what I wanted is that any developer can serve millions of Indians in their own language without a cloud bill, a key or a big company's permission. So this is the whole integration:

# pip install cactus-needle
# hf download udgamlabs/Shruti shruti.cact --local-dir .
from needle import Whistle
model = Whistle("shruti.cact")
print(model.transcribe("clip.wav")["text"])


If your app already runs on ONNX Runtime or sherpa-onnx, all 22 are at udgamlabs/Shruti-ONNX, with pip install shruti-onnx and an npm package. Those builds are about 32 MB and 54 MB, and on the phone we tested, with Hindi, the file above is still faster and lighter.

On DiarBench conversations, which we kept out of training, it makes fewer errors than IndicConformer in 11 languages when English words count only in English letters. Count them in the native script and IndicConformer leads in 10 of those 11. On clean read speech bigger models are ahead, on real conversation a quarter to over half the words are still wrong, and we have not had native readers review the output. I think one number for a family of 22 would hide all of this, so the write-up gives each language on its own.

Speak one of these? Try it and tell me where it fails. The code, the write-up and the report are already out.

Write-up: https://udgamlabs.com/blog/shruti-family
Try it: https://udgamlabs.com/models/shruti
Models: udgamlabs/shruti-6ad5f80d9d300685a6249c9b
Technical report: https://udgamlabs.com/research/shruti-family
beamster 
posted an update 3 days ago
view post
Post
6,775
Last Qwen3.8 Flash update with STQ 2.75bpw v2.0 pack. It moves away from hybrid affine/stq format to pure STQ with +15% quality to match oQ4e. 64GB users can enjoy this while waiting for Qwen 4!

Require Sushi 1.5.1 ro run.
Quazim0t0 
posted an update 4 days ago
view post
Post
7,350
Quazim0t0/Qwenx86-4B-compact

▶️ See it in action: Quazim0t0/Qwenx86-Debugger . Walk through actual recorded tool calls: the conversation, x86 registers and flags, code at the program counter, RAM and machine output, with all verified blocks on a timeline.

Qwen3.5-4B with a 64-bit x86 computer compiled into its weights: 13 transformer layers ahead of Qwen's own. There are no learned weights in them; all of their weights are computed by a compiler. Qwen invokes the computer through its own tool calls (<tool_call>), and the computer in the weights executes the tool. Calculator, Python, C compiler, notes store, physics/chemistry lookup and Python-docs lookup are all firmware in the checkpoint. This is the compact checkpoint: ~9 GB GPU memory, so it fits on a 16 GB GPU.

Edit: GGUF Release: Quazim0t0/Qwenx86-4B-GGUF

Credits: Qwenputer by Divyansh Agrawal (@divagr1925 , github.com/divagr18/qwenputer, MIT)
  • 1 reply
·
yannsouetre 
posted an update about 11 hours ago
view post
Post
470
COF 1.2: AI characters now have an inner life 🧠

With 1.1, Clara got memories. 1.2 gives her what those memories made of her: a psyche graph of values, beliefs, wants, needs, fears, wounds, tastes and habits, each node with a strength, a stability, dates and a "because" link to the memory behind it.

Some nodes she tells freely, some only when it matters, some never. A belief can be marked false but held: she lives by it, the reader knows better.

A converter can propose nodes from a character card or free text; nothing enters the file until the author ticks it. Agents get it as text: cof export --format agent writes a SOUL.md.

Open spec, MIT tools (CLI, builder, explorer, ComfyUI node). Every 1.0 and 1.1 file stays valid. RFCs open for comments until November 22.

https://yannsouetre.github.io/cof/psyche/
Ternii 
posted an update about 15 hours ago
view post
Post
763
Published TR Quantisation versions of Qwen3.6-27B, Qwen3.6-35B-A3B and Qwen3.8-Flash-Next: weights packed into trits (2.25-3 per weight), no retraining.

- Mac: Qwen3.6-35B-A3B in a 16.5 GB GGUF that fits a 24 GB Mac, ~120 tok/s sustained on an M5 Max. Needs our llama.cpp build (ready-built for macOS and Linux arm64).
- vLLM on one DGX Spark: 35B at 180.9 tok/s for 1 user vs 128.0 for NVIDIA's NVFP4, 17.4 vs 21.1 GiB, no significant accuracy difference on the core benchmarks.
- Trade-offs on each card: slower prompt processing, lower multilingual maths, and 4-bit GGUF stays closer to the original text.

https://huggingface.co/collections/Ternii/tr-quantisation
Recently active users