TrendPulse

I can't imagine how arrogant one must be to make such a decision. People pay $200 a month for Anthropic to mess with their codebase. Imagine how they would humiliate their customers if the world didn't have an open-source model. https://preview.redd.it/6qr2ymt25d6h1.png?width=1646&format=png&auto=webp&s=bdc349c68bbf92285d5d5fbca39afa6494868aa8

Thank you to everyone who contributed to my previous post, providing feedback and various models to add, and questioning the rating system. You can now participate in a live blind voting to create a proper ELO for all the models that are added. Each new model that we add will automatically go into the voting pool. https://5uck1ess-tts-arena.hf.space/ Please let me know other things to improve. Local TTS should hopefully be a little easier for everyone. https://github.com/5uck1ess/tt...

Previously I did post a thread on this. Now with some more details. GGUF downloads: Gemma-4-12B-it: https://huggingface.co/Zhongzhu/OSCAR-LLAMACPP-Gemma-4-12B-it-INT2-KV Qwen3-32B: https://huggingface.co/Zhongzhu/OSCAR-LLAMACPP-Qwen3-32B-INT2-KV Qwen3-4B-Thinking-2507: https://huggingface.co/Zhongzhu/OSCAR-LLAMACPP-Qwen3-4B-Thinking-2507-INT2-KV Code: llamacpp: https://github.com/FutureMLS-Lab/OSCAR/tree/zhongzhu/llamacpp sglang: https://github.com/FutureMLS-Lab/OSCAR/tree/z...

SCAIL-2: Unifying Controlled Character Animation with End-to-end In-Context Conditioning SCAIL-2 is an open-source model for end-to-end controlled character animation . It animates a reference character with a driving video, and also supports character replacement and multi-character scenarios without relying on intermediate pose representations. Overview Prior approaches to character animation depend heavily on intermediate representations such as skeleton maps or inpainting masks. These...

Just a warning to anyone thinking about signing up to OpenCode Go/Zen. It appears that you are unable to delete your account. There are various GitHub issues open regarding this, and they do not seem to be getting any response. https://github.com/anomalyco/opencode/issues/29090 https://github.com/anomalyco/opencode/issues/24303 This one does have a response, saying they'll 'probably' add the ability to delete your account yourself: https://github.com/anomalyco/opencode/issues/29101 ...

Having a chat bot you can ask questions is cool and all but for more advanced stuff like tool calling, agents etc. you will need some kind of harness. so far heard of opencode, hermes, openclaw, claude code, pi. Also interested in the use cases what harness do you use for what task what is each good for.

Hey everyone, TinySearch v0.2.0 (first stable beta) is out. The first version used DuckDuckGo directly, which worked well enough to prove the idea, but yeah.. relying on one search source was way too fragile lol. DDG started throwing limits/CAPTCHAs more often in the last 2 weeks, I guess they realized how many of us where doing exactly this, and for an MCP tool that agents depend on, that’s not really good enough. So in v0.2.0, TinySearch now uses SearXNG as the default search backend. ...

I just feel i need to post this here again so more people see: Test around with throttling the power limits of your GPUs, you will often find that you can save tons of power with only minor performance deficits. On my dual Radeon VII setup, i went from 250 to 100 watts per card, and the speeds diminished by not even 10%.

https://marketplace.nvidia.com/en-us/enterprise/laptops-workstations/nvidia-rtx-pro-6000-blackwell-workstation-edition/

Early access was linked here a few days ago, but final release seems to be now. 30B A3B coding model. Weights: https://huggingface.co/CohereLabs/North-Mini-Code-1.0 Blog: https://huggingface.co/blog/CohereLabs/introducing-north-mini-code Artificial analysis score of 28 is pretty weak against Qwen 3.6 35B (43), but it's more competitive in coding index (33 vs 35) and well above Gemma 4 26B (22) there: https://artificialanalysis.ai/models/north-mini-code submitted by ...

Hi folks! Jay here from Cohere. we just officially launched North Mini Code after getting some great feedback from you guys this weekend on the unreleased version. I wanted to come here and answer some of the questions you asked and provide some extra detail about the model itself. You can download the weights on Hugging Face ( fp8 here ) or try it on OpenCode for free . if you want to read more about what I mentioned in the video, feel free to look at our technical blog post on Huggi...

They're both available as q8_0 models named mtp-gemma-4-*.gguf on the root of the directory and in both q8_0 and larger quants within an MTP folder. https://huggingface.co/unsloth/gemma-4-12B-it-qat-GGUF/tree/main https://huggingface.co/unsloth/gemma-4-26B-A4B-it-qat-GGUF/tree/main https://huggingface.co/unsloth/gemma-4-31B-it-qat-GGUF/tree/main https://huggingface.co/unsloth/gemma-4-E2B-it-qat-GGUF/tree/main https://huggingface.co/unsloth/gemma-4-E2B-it-qat-mobile-GGUF/...

Key Technical Advantages: Performance: The M5 chip's neural accelerators significantly boost prompt processing Concurrency: MLX LM Server utilizes continuous batching to handle multiple sub-agent requests simultaneously without stalling Scaling: For massive models that exceed local memory, MLX supports distributed inference across multiple Macs using Thunderbolt RDMA To get started, developers can install MLX LM via pip and point their preferred agent tool to the...

https://huggingface.co/silx-ai/Quasar-Preview https://preview.redd.it/ur27udpzy66h1.jpg?width=900&format=pjpg&auto=webp&s=5ce3a8f2f5829fb4a94d9d3563145dadd2d63309

This is still a work in progress, but since recording the video, I added callbacks for tool use, more tests, and published it as a cargo crate . Currently working on speeding up the prefill. The decode speed is almost the same on my Ryzen 7950x (~37 tokens/s), but the prefill speed is not yet optimized (almost the same as decode). This model can comfortably run on a machine with 16GB of RAM. Its memory usage will fit within ~7GB. You can reuse the weights between multiple Agent instances, e...

Hopefully this isn't too low effort of a post. I just finished the benchmarks and I figured I'd post them online because they certainly were insightful for me. I did not use any AI other than asking Gemini 3.1 Pro if it was statistically significant because I was too tired to do inferential statistics. Methodology: oMLX used to run Gemma 4 26BA4B IT from mlx-community. I used the following models: Gemma 26B 4 Bit: https://huggingface.co/mlx-community/gemma-4-26b-a4b-it-4bit Gemma 26B ...

I'm just getting started using local LLMs for code. I'm not interested vibe coding, but I am hoping to increase my productivity in the publish or perish world of academia. My existing code from past projects is a mess and LLMs often fail to understand my code because I work with niche models, don't comment much, and sometimes have misleading variable names that LLMs over index on (if I redesign things as I learn new information I might not rename variables as I change their use). So, I'm moving ...

https://preview.redd.it/cugpphztz96h1.jpg?width=899&format=pjpg&auto=webp&s=2aa10f8b8f2a0ff666cdc2c63c1775ffd2ed7e7b https://preview.redd.it/14wncc3tz96h1.jpg?width=850&format=pjpg&auto=webp&s=85d3bc9c19ef458a0578159d5dc709552e92dad9 https://preview.redd.it/bj4fmubsz96h1.png?width=869&format=png&auto=webp&s=a10531edb03cd65344c2e8eb76f8303d85a1ed98 https://preview.redd.it/vh3y6wt9y96h1.jpg?width=857&format=pjpg&auto=webp&s=e0d509cd207d3c...

This news seem to have flown under the radar. Apple announced CoreAI on WWDC which is basically a future replacement for CoreML and an alternative to MLX/llama.cpp/torch for on-device optimized inference, especially on phones and tablets. The model weights need to be converted similarly to CoreML via python script, atm the list of supported models is mostly from mid 2025 year though https://github.com/apple/coreai-models/tree/main/models . For anyone wondering how is that anything new - Cor...

I had a huge LLM server , and now I have a tiny one! I had a Jetson Orin NX gathering dust from a long dead robotics project, from back in the Llama-7B days. I figured now with MoE and smaller models doing well, it was time to mess with it again. Goal: As silent as possible (given they bumped the power from 25W -> 40W) Greater than 10 tok/s TG and 300 tok/s PP at least 65K context for Hermes Agent Must look cool AF 👌🏻 With those constraints, I had to take a hacksaw to the...

nobody expected HF there

I'm trying to find out if anyone has done any benchmarking comparing the Gemma 4 4-bit QAT models (via Unsloth) against standard 8-bit non-QAT quants. I know QAT is supposed to retain a ton of accuracy compared to the baseline BF16, but I'm curious how a 4-bit QAT model actually fares against a traditional 8-bit PTQ. I've read some mixed feedback across different threads, but I haven't been able to find hard numbers or a direct head to head comparison between the two. Has anyone run any eval...

The question I’m asking myself is whether open-source LLMs are now “ just good enough ” to meet 95% of requirements. I know, of course, that they still need to and will get even better, but where does the added value of the remaining 5% come from? a) Better answer quality? Okay, but does that justify the extra cost? b) Cleaner automated loops? Do the extra costs justify the effort of manual interventions to produce the same or similar quality? c) Reduced risk of facing internal/external ...

No accounts or sign up. No email requirements. No pop-ups and no cookies. No ads. Info is saved locally in your browser so you dont lose any progress. Its got every single flag and argument that could be found in the documentation. Tool tips are added to everything. Every field is editable. Once you build the CLI or server command you can add your run info and add the run to the log and track which configuration works best for your hardware. Only Linux support currently. Maybe Mac and w...

Oh Hey Folks, I took the Mellum 2 model for a spin, so I wanted to share my impressions here. Disclaimer: the tests presented here are not cientific nor have those nice names like perplexity,etc. These tests are somewhat more akin to what Im working in a daily basis or how useful a model is helping me on a given task. Just saying. First of all, being a 12b moe model with 2.5b params activated is somewhat uncommon but look at the speed: Model JetBrains/Mellum2-12B-A2.5B-Think...

This PR improves matmul performance for k-quants. The following table shows the improvement on the pp512 test in M2 pro. quant model master (t/s) PR (t/s) speedup Q2_K qwen3 0.6B Q2_K - Medium 817.86 ± 6.14 1991.81 ± 6.87 2.44x Q3_K qwen35 4B Q3_K - Medium 92.54 ± 0.13 302.24 ± 0.37 3.27x gemma4 E4B Q3_K - Medium 79.06 ± 0.08 298.73 ± 0.90 3.78x Q4_K qwen35 4B Q4_K - Medium 243.82 ± 0.09 327.24 ± 0.59 1.34x g...

By default, llama.cpp enables pipeline parallelism, presumably to speed up inference. In my testing, I found that pipeline parallelism has no speed benefit and comes at a significant cost of VRAM. This cost can be avoided by compiling llama.cpp with the -DGGML_SCHED_MAX_COPIES=1 option. This prevents llama.cpp from allocating a much larger compute buffer when pipeline parallelism is enabled. Pipeline parallelism is enabled when --split-mode layer is used (the default) and all model layer...

I fine-tuned NVIDIA's Parakeet TDT 0.6B v2 for clinical speech and am releasing the weights as Omi Med STT v1 (CC-BY-4.0). Disclosure: I'm the founder of Omi Health and built this. Happy to dig into the training mix, benchmark, failure cases, quantization, or anything else. The goal was simple: get a small local ASR model close enough to the strong cloud systems that patient audio doesn't have to leave the device for transcription. There's also a runtime for Mac, Windows and Linux. Insta...

MODS: if you wanna remove for slop, that's cool - once I have something like a llama.cpp patch I'll repost something people can use. I'll write a full article on my Medium account with how it works. Just got excited and wanted to share. *** Forgive the claude summary, in the readme, but the base works. I'm still working on the hip kernal and having it combine with MTP. I hope to get up near 80 tk/s. All started because I realized every Q8 (INT8 or F8) calculation was using f32 of compute an...