TrendPulse

Hi. Many people think uncensored models are basically the same model that just doesn't refuse, but... I was recently checking whether uncensored models would give me better answers for stock market predictions (my idea was: the uncensored one will tell you the truth and won't be polite where it shouldn't be). And I noticed that abliteration didn't only remove the refusals, it also changed the model's attitude. Generally, after removing censorship the models are more optimistic . More &quot...

I was looking for a strong coding model and a strong general model, both should be 120b or under. After weeks of researching, qwen3.6 27b(general) and qwen3 coder next(coding) are the top choices. I mean there has to be something better. Any reccommendations?

The model was quantized to 8, 4, 2, and 1 bit. Characteristics: Q8 : 8-bit 1.56 TB, lossless Q4 : 4-bit, 1.51 TB Q2 : 2-bit: 861 GB Q1 : 1-bit, 594 GB The smallest Q1 model keeps 78.9% accuracy, while being almost 3 times smaller than the original one. Instruction for running the model is in the model's card

I've got better results than expected for 768gb DDR5 and 2x5090. Using fork https://github.com/pwilkin/llama.cpp/tree/kimi-k3-text and https://huggingface.co/GrEarl/Kimi-K3-GGUF Q2_K quant. Prefill speed for big prompt is 50-70 tps. The most fun thing that decoding tps growing over time. Maybe some kind of warmup or swap thingy. Llama-becnh crashes, so can't share.

No weights yet. I feel sad for them, that training run cost maybe 100s of thousands of dollars and they didn't even beat GPT-OSS 20B in every regard But the ability to train a model from scratch on 3.5 trillion tokens instead of 35 trillion sure gives me hope. They only spent 100s of thousands of $ instead of millions, so maybe soon enough hobbyists will be able to pretrain true LLMs at home

DISCLAIMER: No Ai was prompted in the creation of this post. Today I had an experience that complely blew my mind, I just had to write it down. As a bit of background I have been dabbling prompting local models using LM studio for the better part of 18 months now, keeping up to date with the latest releases but never actually doing anything with them. I am not a software developer, I'm not deeply technical, I wouldnt even call myself a vibe coder. The last couple weeks I got Hermes conn...

Kind of unexpected. Happy for Gemma-4/Google, big win for us, LocalLLMers. Yet Qwen3.6 still does better in Hermes than Gemma-4 somehow. We need Gemma-4.1 fine-tuned on Agentic-tasks. That would be killer.

My testing with android suggest it is below that. Vision on Opus is the reason. But I don't doubt that for non vision tasks Kimi k3 is on par.

This seems like a pretty solid improvement, and should make the more extreme quant setups viable on AMD cards.

pi 0.81.0 now has integrated support for llama.cpp (llama-server router). https://pi.dev/docs/latest/llama-cpp This seems to be able to replace the huggingface/pi-llama extension and/or manually managing models in the config.

I've been working on this project that makes LLM distributions decentralized and fast using torrents. Read more about tech on Github . Website: https://llama.garden Suggested client: Transmission News: - Added more web seed URLs that go through our API that will increase speeds thanks to HTTP being faster than UDP. This is just for initial seeding, then we can rely on peers becoming seeders for broader distribution - Wrote an API that resolves HF CDN to actual URLs and caches thos...

https://huggingface.co/Motif-Technologies/Motif-3-Beta Motif-Technologies is one of the tech company participated South Korea's AI Foundation Model project. Upstage(Solar Series), LG AI Research(EXAONE Series), and SKT(A.X Series) are the competitors. Deepseek V4 Pro: 1.6T-A49B MiniMax-M3: 428B-A23B Motif 3 Beta: 314B-A13B + About South Korea's AI Foundation Model Project. South Korea's AI Foundation Model Project (This will not be official English name.)(aka. K-AI) is one of the na...

(Slides because why not) I've been building a serverless post-quantum group-encryption protocol for the last few weeks, mostly with Opus 4.8 and Fable, with GPT-5.6 Sol running adversarial review on the most challenging part (a recovery/finality mechanism). It went through four rounds of review: each round found something, Claude fixed it, and by the end the formal model was solid. I was trusting the process... But then Kimi K3 comes along and I figured I'd give it a shot. No guardrails on a...

I know DavidAU gets a bad rap, and rightfully so. I've tried some of his fine tunes in the past and they have been... Interesting. I stumbled on this model on accident. A locallama post asking people to name some of longest model names they have found: https://www.reddit.com/r/LocalLLaMA/s/rsvWBHZasI So I hopped on the HF page and noticed a few things that stood out. The model is a collaboration between different people. This model isn't a one man show like most of his models iirc. ...

I’ve seen many people worrying that if Trump moves to block Chinese AI models, aggregators like OpenRouter will no longer be able to host them. As someone from China, let me reassure you: there is really no need for such concern. China has been locked in trade conflicts with Trump for years. Nobody understands better than us Chinese how to deal with him. We can connect with local business partners in Vietnam 🇻🇳 willing to act as distributors. We will fork the model repository and hand it o...

I asked myself where the Bonsai models actually land, so I ran them and compared to the results I already have for qwen-3.6-35b-a3b and qwen-3.5-9b on the same harness. Thought it might interest more people. Setup: little-coder harness via the harbor adapter, all 89 tasks of terminal-bench 2.0, single attempt (k=1), 40-turn cap, temp 0.2. RTX 5070 Laptop 8GB, i9-14900HX, 32GB RAM, CUDA 13.1. Runtime is PrismML's llama.cpp fork (stock llama.cpp can't load the 2-bit kernels). Results: Ternary-...

https://huggingface.co/Nanbeige/Nanbeige4.2-3B Nanbeige4.2-3B is a compact agentic model built on Nanbeige4.2-3B-Base , designed to combine strong agentic behavior with broad reasoning and alignment capabilities. Its Looped Transformer architecture reuses the transformer layers to increase model capacity without adding parameters. With only 3B non-embedding parameters, the model delivers solid performance on general-agent and code-agent tasks.

There are people on twitter/X saying that Chinese models are as good as they are only due to distillation (source: https://x.com/scaling01/status/2079332469501727052 ). I don't buy that even for a minute. While the investigation is interesting, one can see that GPT models do not really "like" themselves, and that is unlikely (at least for 5.4 and 5.5). Especially as Opus and Gemini models like each other. (here I use "like" for "they are not that surprised by the oth...

HF: https://huggingface.co/poolside/Laguna-S-2.1 GGUFs available for use with llama.cpp custom fork: https://huggingface.co/poolside/Laguna-S-2.1-GGUF Posted on X: https://x.com/poolsideai/status/2079613777343848465?s=20

Model Size Terminal-Bench 2.1 SWE-bench Multilingual SWE-Bench Pro (Public Dataset) DeepSWE SWE Atlas (Codebase QnA) Toolathlon Verified Laguna S 2.1 118B-A8B 70.2% 78.5% 59.4% 40.4% 46.2% 49.7% Finally the banger we've been waiting from Laguna. probably will be great for 64GB+ RAM and VRAM setups.

Nearly a year later, we've had safeguard fine-tunes but no general-purpose base model or successor. Will Kimi, Qwen and GLM force their hand?

$1.5 billion settlement largest known payout in U.S. copyright case Case part of a wave of lawsuits from copyright holders against AI companies Some authors and publishers opted out and continue separate ​cases against Anthropic

Kimi-K3’s release, while impressive, is still months behind the closed-source frontier, so all the “it’s over for Anthropic” talk feels overblown. According to Artificial Analysis, though, Kimi-K3 has brought the open-source frontier to just 1.5 months behind closed-source, putting it right on the heels of OpenAI and Anthropic. Also worth noting from the graph: where has Google been since Gemini 3 Pro last November? The top open-source models keep getting bigger, proving scaling laws still hol...

Google hasn't shipped a model recently that is capable of competing with Sol or Fable. The previous models were pretty disappointing and unreliable, it seems the more time goes on that they might have different strategies: - They might be going all-in on on-device inference for their own products. But this is a battle that Apple might win because they just have better hardware and can license a third party open model. - They might be just buried deep into internal politics and nobody is ...

From clem 🤗 on 𝕏: https://x.com/ClementDelangue/status/2079301434357456931 Fortune: Hugging Face says it resorted to a Chinese AI model to battle a fully autonomous cyberattack because U.S. model guardrails stymied its defense: https://fortune.com/2026/07/20/hugging-face-turns-to-chinese-open-source-ai-to-fend-off-autonomous-ai-cyber-attack-after-american-ai-guardrails-stymie-defense/