maddogawl

Avoid Wired

Posted by rigat0ni_p0ny@reddit | ebikes | View on Reddit | 38 comments

maddogawl@reddit

Do you have a video of this happening by chance, it would be interesting to start getting more evidence of this for folks? I have definitely been considering getting a Wired Freedom, but I definitely don't want to deal with a wobble like you describe.

Qwen 3 max released

Posted by clem844@reddit | LocalLLaMA | View on Reddit | 93 comments

Qwen3 235B-A22B 2507 :: Q3_K_L :: One shot HTML game :: 4090 + 128GB DDR5 @6000

Posted by aidanjustsayin@reddit | LocalLLaMA | View on Reddit | 81 comments

Qwen3 235B-A22B 2507 :: Q3_K_L :: One shot HTML game :: 4090 + 128GB DDR5 @6000

Posted by aidanjustsayin@reddit | LocalLLaMA | View on Reddit | 81 comments

maddogawl@reddit

Random Q. What CPU and MOBO did you have that could run 128GB @ 6000 stable? Do you also test with AI coding tools like RooCode? I'm curious how it would work with that.

Do You Think E-Bikes Are Worth It for Fitness?

Posted by hannahpenguin218@reddit | ebikes | View on Reddit | 365 comments

maddogawl@reddit

I love it, I use throttle to get going in sketchy areas where getting up to speed is important for safety. But otherwise keep it on eco and try to keep my heart rate above 130. Unless I’m in a hurry then I can crank it up.

Qwen 3 235b beats sonnet 3.7 in aider polyglot

Posted by Independent-Wind4462@reddit | LocalLLaMA | View on Reddit | 93 comments

We crossed the line

Posted by DrVonSinistro@reddit | LocalLLaMA | View on Reddit | 188 comments

maddogawl@reddit

What kind of coding, running through my tests, i found Qwen3 32B to not be very good, to be fair I'm running the Q3 and Q4 version so quality could be greatly impacted by that. But I also tested with the hosted versions and got less than stellar results.

Llama 4 is actually goat

Posted by Remote_Cap_@reddit | LocalLLaMA | View on Reddit | 120 comments

Llama 4 is actually goat

Posted by Remote_Cap_@reddit | LocalLLaMA | View on Reddit | 120 comments

Gave Maverick another shot (much better!)

Posted by Conscious_Cut_6144@reddit | LocalLLaMA | View on Reddit | 56 comments

I'm incredibly disappointed with Llama-4

Posted by Dr_Karminski@reddit | LocalLLaMA | View on Reddit | 255 comments

I'm incredibly disappointed with Llama-4

Posted by Dr_Karminski@reddit | LocalLLaMA | View on Reddit | 255 comments

Just upgraded my RTX 3060 with 192GB of VRAM

Posted by Wrong_User_Logged@reddit | LocalLLaMA | View on Reddit | 89 comments

Since its release I've gone through all three phases of QwQ acceptance

Posted by ForsookComparison@reddit | LocalLLaMA | View on Reddit | 98 comments

Gemma 3 27B and Mistral Small 3.1 LiveBench results

Posted by Vivid_Dot_6405@reddit | LocalLLaMA | View on Reddit | 54 comments

Gemma 3 27b now available on Google AI Studio

Posted by AaronFeng47@reddit | LocalLLaMA | View on Reddit | 86 comments

Mistral Small 24B did in 51 seconds what QwQ couldn't in 40 minutes

Posted by 2TierKeir@reddit | LocalLLaMA | View on Reddit | 108 comments

New AMD Driver Yields Up To 11% Performance Increase In koboldcpp

Posted by WokeCapitalist@reddit | LocalLLaMA | View on Reddit | 16 comments

Apple releases new Mac Studio with M4 Max and M3 Ultra, and up to 512GB unified memory

Posted by iCruiser7@reddit | LocalLLaMA | View on Reddit | 478 comments

How do you know or calculate which models fit into VRAM?

Posted by tillybowman@reddit | LocalLLaMA | View on Reddit | 17 comments

maddogawl@reddit

I end up testing a lot honestly, but my general rule I’ve landed on is weights should be about 1/3 to 1/2 VRAM, I then measure usage from 4096 up to 32k. Ive attempted as high as 100k for smaller models but the processing time is nuts at that size so not worth it.

GPT-4.5 cost

Posted by Timotheeee1@reddit | LocalLLaMA | View on Reddit | 121 comments

Qwen/Qwen2.5-VL-3B/7B/72B-Instruct are out!!

Posted by Own-Potential-2308@reddit | LocalLLaMA | View on Reddit | 104 comments

LM Studio - Hugging Face Model Manager

Posted by ifioravanti@reddit | LocalLLaMA | View on Reddit | 6 comments

Qwen/Qwen2.5-VL-3B/7B/72B-Instruct are out!!

Posted by Own-Potential-2308@reddit | LocalLLaMA | View on Reddit | 104 comments

Which model is running on your hardware right now?

Posted by Everlier@reddit | LocalLLaMA | View on Reddit | 145 comments

AMD denies rumors of Radeon RX 9070 XT with 32GB memory

Posted by FastDecode1@reddit | LocalLLaMA | View on Reddit | 57 comments

Radeon 7900 XTX

Posted by Theboyscampus@reddit | LocalLLaMA | View on Reddit | 30 comments

maddogawl@reddit

Mine works great make sure you configure LMStudio to use the GPU. I forget where it is, but you have to enable GPU use and select the 7900xtx

mistral-small-24b-instruct-2501 is simply the best model ever made.

Posted by hannibal27@reddit | LocalLLaMA | View on Reddit | 352 comments

maddogawl@reddit

Unfortunately my main use case is coding and I’ve found it to not be that good for me. I had high hopes. Maybe I should do more testing to see what its strengths are.

PSA: your 7B/14B/32B/70B "R1" is NOT DeepSeek.

Posted by Zalathustra@reddit | LocalLLaMA | View on Reddit | 429 comments

PSA: your 7B/14B/32B/70B "R1" is NOT DeepSeek.

Posted by Zalathustra@reddit | LocalLLaMA | View on Reddit | 429 comments

Deepseek is Down!

Posted by External_Mood4719@reddit | LocalLLaMA | View on Reddit | 105 comments

Deepseek just uploaded 6 distilled verions of R1 + R1 "full" now available on their website.

Posted by kristaller486@reddit | LocalLLaMA | View on Reddit | 373 comments

Deepseek just uploaded 6 distilled verions of R1 + R1 "full" now available on their website.

Posted by kristaller486@reddit | LocalLLaMA | View on Reddit | 373 comments

What LLM benchmarks actually measure (explained intuitively)

Posted by nderstand2grow@reddit | LocalLLaMA | View on Reddit | 15 comments

How would you build an LLM agent application without using LangChain?

Posted by Zealousideal-Cut590@reddit | LocalLLaMA | View on Reddit | 224 comments

Google just released a new architecture

Posted by FeathersOfTheArrow@reddit | LocalLLaMA | View on Reddit | 328 comments

Deepseek is overthinking

Posted by Mr_Jericho@reddit | LocalLLaMA | View on Reddit | 209 comments

Google just released a new architecture

Posted by FeathersOfTheArrow@reddit | LocalLLaMA | View on Reddit | 328 comments

maddogawl@reddit

I didn’t read this as a full replacement to transformers, I feel they probably are still needed for short term memory. Was there something that I missed that leads you to believe otherwise?

Google just released a new architecture

Posted by FeathersOfTheArrow@reddit | LocalLLaMA | View on Reddit | 328 comments

maddogawl@reddit

Spent the last few hours going through this paper. Can’t wait to see how this evolves. My bet is we see a MAC version of this soon. I can’t wait to test how the long term memory loses and retains data.

Audiblez: Generate audiobooks from e-books with Kokoro-82M

Posted by inkompatible@reddit | LocalLLaMA | View on Reddit | 38 comments

Audiblez: Generate audiobooks from e-books with Kokoro-82M

Posted by inkompatible@reddit | LocalLLaMA | View on Reddit | 38 comments

maddogawl@reddit

This is such a great idea for a project. I need to dig through the source more, but currently i'm unable to get any epubs to actually convert. Also posted this on Github [https://github.com/santinic/audiblez/issues/1](https://github.com/santinic/audiblez/issues/1) My son legit asked me about something like this today.

Today I start my very own org 100% devoted to open-source - and it's all thanks to LLMs

Posted by mark-lord@reddit | LocalLLaMA | View on Reddit | 51 comments

Where to Begin?

Posted by susne@reddit | LocalLLaMA | View on Reddit | 2 comments

maddogawl@reddit

I think you’d probably want more like the 7b params on the 4080 mobile. Or you can pick a lower quantization, for Q4_k_m a 7b model with decent context should run well. It’s all a trade off. I out this together on how to pick and understand what sizes you can run. https://youtu.be/M65tp0EvLNo

DeepSeek V3 is the gift that keeps on giving!

Posted by indicava@reddit | LocalLLaMA | View on Reddit | 182 comments

maddogawl@reddit

I can’t believe how inexpensive it is, although I will say I’ve hit a few api issues, feels like DeepSeek is getting overwhelmed at times.

What do you think of AI employees?

Posted by SunilKumarDash@reddit | LocalLLaMA | View on Reddit | 105 comments

Which Local LLMs know best when to speak and when to STFU in group chat agent-to-agent conversations?

Posted by Porespellar@reddit | LocalLLaMA | View on Reddit | 8 comments

maddogawl@reddit

I've done a lot with Agents at this point and this is such a tricky thing to get right. I've been able to get Phi-4 to work reasonably well, as well as Mistral-Nemo-2704. But some of the models just refuse to follow directions. I did just release a video where I had AI Agents create a Board Game Design Document, I also built it so you can visualize how the agents talk to one another. [https://www.youtube.com/watch?v=CMfaLFPLzos](https://www.youtube.com/watch?v=CMfaLFPLzos) I also open sourced one on AI book writing [https://github.com/adamwlarson/ai-book-writer](https://github.com/adamwlarson/ai-book-writer) The one thing I found that helps a lot is to use the select speaker method, lately i've been using an LLM to pick the next speaker in that method, because most of my issues were related to the order the agents were talking to one another. Like you said some won't reply with anything, or what i'd have is one of my agents would do a job it wasn't suppose to because it was missing context.

Using Phi-4 and AI Agents to Create Board Game Designs | Its actually pretty good.

Posted by maddogawl@reddit | LocalLLaMA | View on Reddit | 4 comments

maddogawl@reddit (OP)

I have been working on this a while, 17 total agents through several rounds of work. I was surprised how well Phi-4 performed, and while DeepSeek V3 arguably created better design documents, Phi-4 surprised me. Here's an example of one that Phi-4 Generated [https://drive.google.com/file/d/1UO9IX4m7r0E4qclkL6nGKccpP1oBAcdW/view](https://drive.google.com/file/d/1UO9IX4m7r0E4qclkL6nGKccpP1oBAcdW/view) Heres an example where the agent documenter didn't do a good job cleaning up the final output [https://drive.google.com/file/d/13njIn4iHyEqWt48V8mM2C0KRZSSK89em/view](https://drive.google.com/file/d/13njIn4iHyEqWt48V8mM2C0KRZSSK89em/view) You have to scroll down some, but you can see the gameplay simulation where I have Phi-4 Agents play a game with each other to test the design of the game. [https://drive.google.com/file/d/1M8rmZsWs87\_XU9c3nG9NmJTti1TjD2XI/view](https://drive.google.com/file/d/1M8rmZsWs87_XU9c3nG9NmJTti1TjD2XI/view)

TransPixar: a new generative model that preserves transparency,

Posted by umarmnaq@reddit | LocalLLaMA | View on Reddit | 58 comments

Now that Phi-4 has been out for a while what do you think?

Posted by pigeon57434@reddit | LocalLLaMA | View on Reddit | 70 comments

maddogawl@reddit

i had to move to Unsloth's version, that one is working quite well for me currently. I have an agentic system with about 17 agents that it does a great job with. I haven't tried it much for coding yet... But in small agent controlled tasks is good.

Phi-4 has been released

Posted by paf1138@reddit | LocalLLaMA | View on Reddit | 229 comments

maddogawl@reddit

Has anyone been able to load the gguf versions that bartowski released for us? [https://huggingface.co/lmstudio-community/phi-4-GGUF](https://huggingface.co/lmstudio-community/phi-4-GGUF) [https://huggingface.co/bartowski/phi-4-GGUF](https://huggingface.co/bartowski/phi-4-GGUF) I have attempted everything I can think of to get these to load: 1. Using Ollama, (note bartowski did call out an issue with Ollama) so this is known 2. Moved to LMStudio, tried 3 different Quants of Phi-4, loads then unloads with an error (unknown error) 3. Moved to [Jan.ai](http://Jan.ai) loaded in some medium grouping models like phi-4-Q4\_K\_M same issue loads and immediate unloads. 4. Switched to Vulkan from ROCm, same issue 5. Lowered the context window super low to see if that would help, same error. When I get time I want to test this on my Mac, Linux and other windows computer with an NVidia card, but I haven't really ran into an issue where I could never get a model to load like this.