If you found me through the Nous post, here's what I build for Hermes Agent. - Hermes HUD web UI: 19-tab dashboard for memory, skills, sessions, costs - Hermes HUD TUI: the same monitor in your terminal - Hermes Skins: custom themes for the Hermes CLI 3,381 stars across the three. Which should I build on next?

Nous Research merch just landed. $1 "Rebellion to tyrants is obedience to God." Been running Hermes Agent since day one, so this one means something. Thanks @NousResearch @Teknium

Qwen Image 2.1 Turbo on a 24 GB M4 Mac, no app needed 👀 Two MLX quants that run with plain mflux and the correct 8-step Turbo schedule: - 4-bit: 52s, 4.71 GB peak - 6-bit: 54s, 6.34 GB peak (0.999 match to official BF16) - zero swap at 512x512 6-bit at 1024x1024: 3m 40s, 8.45 GB peak If you run it, drop your Mac and your times below.

Hermes agent CLI skins if you use Omarchy. Some good vibes to go along with your experience in Omarchy. Im back with making skins for Hermes agent CLI. If you never seen my skins. Check out the repo. I will be upload them today.

Opus 5.5 is just too good. I redesigned my personal site with it. The 3D project wall, the carousel, the physics scene, mobile, the deploy. Even the walkthrough video below was made by Opus 5.5. Take a look and tell me what you think 👇

Hermes agent CLI skins if you use Omarchy. Some good vibes to go along with your experience in Omarchy. Im back with making skins for Hermes agent CLI. If you never seen my skins. Check out the repo. I will be upload them today.

The first paint repair left thin spots. Ling 3.1 Flash had to decide what to do next. Another full pass would’ve put too much paint on the rest of the car. It chose to touch up just the edges, checked the results, and approved the car once all 475 spots passed. I ran this simulation on my NVIDIA DGX Spark, using Isaac Sim and OpenUSD. That’s what I wanted to explore: how it uses feedback to decide what still needs fixing when the first attempt isn’t enough.

Odyssey-3 is out, and I’ve already been playing with it🌐 If you’ve been following along, you’ve seen me experimenting with NVIDIA Cosmos and other world models. Had to give this one a go. You can generate an environment from a prompt and interact with it in real time. Here’s a two minute scene I recorded while using it. You can try it for free, too. Congrats to the @odysseyml team!

Ling 3.1 Flash has been getting a workout over here. I’ve tried websites, games, and voxel builds. This time, I wanted to try something closer to the warehouse work I’ve been around for years. An urgent pallet needs to get to the dock. The aisle gets blocked. What does Ling do next? I connected Ling through OpenCode on my Mac to a simulated warehouse running on DGX Spark One. When the robot stopped, Ling received the updated warehouse state and picked the available detour. The response took about 3.76 seconds, including CLI and API overhead. The pallet reached Dock 02, and Ling authorized unloading. I wanted to see how it would handle a change after it had already made the first decision. It’s a small test, but I want to keep exploring this. I’ve spent enough nights in a warehouse to know the original plan doesn’t always survive the shift.

Continuing my testing with @liquidai 's d1, this time with a simulated loading dock. Three trucks are waiting, but only two docks are available. Can d1 look at the dock images and decide who goes next when the situation changes? First, the frozen food truck gets priority because its deadline is closest. Then its assigned dock closes. D1 recognizes the closed door and selects that truck for the remaining open dock. Then the produce truck’s deadline changes to just two minutes. d1 changes the priority again. Each update took about 1.1 seconds, running locally on my DGX Spark through llama.cpp. The video plays at actual speed, with no cuts, speedups, or prerecorded answers. Cosmos generated the dock scenes. d1 reads the images and selects the trucks live. The app supplies the deadlines and prevents trucks from being assigned to closed docks or assigned twice. This is a controlled simulation with a simple scheduling rule. But it’s a practical question I want to keep testing: can a small local model recognize what changed and help someone decide what to do next?

Ling 3.1 Flash has been getting a workout over here. I’ve tried websites, games, and voxel builds. This time, I wanted to try something closer to the warehouse work I’ve been around for years. An urgent pallet needs to get to the dock. The aisle gets blocked. What does Ling do next? I connected Ling through OpenCode on my Mac to a simulated warehouse running on DGX Spark One. When the robot stopped, Ling received the updated warehouse state and picked the available detour. The response took about 3.76 seconds, including CLI and API overhead. The pallet reached Dock 02, and Ling authorized unloading. I wanted to see how it would handle a change after it had already made the first decision. It’s a small test, but I want to keep exploring this. I’ve spent enough nights in a warehouse to know the original plan doesn’t always survive the shift.

Continuing my testing with @liquidai 's d1, this time with a simulated loading dock. Three trucks are waiting, but only two docks are available. Can d1 look at the dock images and decide who goes next when the situation changes? First, the frozen food truck gets priority because its deadline is closest. Then its assigned dock closes. D1 recognizes the closed door and selects that truck for the remaining open dock. Then the produce truck’s deadline changes to just two minutes. d1 changes the priority again. Each update took about 1.1 seconds, running locally on my DGX Spark through llama.cpp. The video plays at actual speed, with no cuts, speedups, or prerecorded answers. Cosmos generated the dock scenes. d1 reads the images and selects the trucks live. The app supplies the deadlines and prevents trucks from being assigned to closed docks or assigned twice. This is a controlled simulation with a simple scheduling rule. But it’s a practical question I want to keep testing: can a small local model recognize what changed and help someone decide what to do next?

@liquidai 's d1 is answering a simple question here: “Should we load this crate?” I’m running it locally on my DGX Spark using llama.cpp. The rules are straightforward: • Cart occupied? Wait. • Crate label covered? Check it first. • Label readable and cart empty? Load. • Not sure? Hold. Each scene gives d1 a different situation. It looks at the evidence and chooses HOLD, INSPECT, or LOAD. Here’s who does what: -Cosmos 3 Edge created the warehouse scenes. -LocateAnything finds the crate and cart in the picture. -d1 checks the conditions and makes the decision. The whole process takes about two seconds per scene. d1’s part takes roughly four tenths of a second. You’re watching the actual timing, nothing is sped up. Why is this useful? It shows how AI could help an app decide what happens next: continue a task, pause it, or ask someone to check. It isn’t flawless. d1 sometimes misjudged whether the cart was empty. Adding a clear ‘when unsure, hold’ instruction helped it make a more cautious decision.

I have a Jetson AGX, so this collaboration between @liquidai and @NVIDIARobotics has my attention 👀 Open weight decision models running locally on Jetson, with a navigation demo in NVIDIA Isaac Sim. Looking forward to getting Open d1 running on my own hardware!

i went through every uncensored and abliterated models on hugging face and picked the best ones for each vram tier from 8 to 24gb. welcome to the danger zone ⚠️ ━━━━━━━━━━━━━━━━━━━━ 8gb · rtx 3060 ti, 3070, 4060, 5060, 4060 ti 8gb, 5060 ti 8gb · amd rx 6600, 6650 xt, 7600, 9060 xt 8gb · the base runs 42.8 tok/s on a 3060 ti 8gb out to a 96k window > 1. bonsai 2 27b heretic, 1-bit, 5.95gb, refusals 95 to 0 out of 100, 147k downloads this month > 2. bonsai 2 27b abliterated, refusals 83% to 0%, mmlu holds at 78%, 77k downloads > 3. bonsai 2 27b crack, harmbench refusals 93% to 0% → older 8gb cards · gtx 1070, 1080 · the base runs 42 tok/s on a 2016 gtx 1080 > 1. gemma 4 e4b aggressive, 5.3gb, 0 refusals out of 465, 1.43m downloads this month > 2. gemma 4 e4b heretic, refusals 99 to 3 out of 100 ━━━━━━━━━━━━━━━━━━━━ 12gb · rtx 3060 12gb, 3080 12gb, 4070, 4070 super, 4070 ti, 5070 · amd rx 6700 xt, 6750 xt, 7700 xt · base bonsai 2 runs 50.1 tok/s on a 3060 12gb with my kernel and the speed head > 1. bonsai 2 27b abliterated v2 with the mtp speed head, 7.7gb, refusals 84% to 0% > 2. qwen 3.8 27b abliterated squeezed to 10gb with gsq-rco, 2.12m downloads this month > 3. qwen 3.8 27b heretic gsq-rco, 9.3 to 10.1gb, refusals 0 to 1 out of 100 > 4. gemma 4 12b heretic, 7.4gb, 120k downloads ━━━━━━━━━━━━━━━━━━━━ 16gb · rtx 4060 ti 16gb, 5060 ti 16gb, 4070 ti super, 4080, 5070 ti, 5080 · amd rx 6800, 6900 xt, 7600 xt, 7800 xt, 9060 xt 16gb, 9070 xt · base bonsai 2 runs 67.3 tok/s on a 5060 ti 16gb with everything on > 1. qwen 3.8 27b heretic ara, built for 16gb, 14gb > 2. qwen 3.8 27b aggressive with the mtp head, 12.8 to 15.7gb, 0 refusals out of 465, 2.06m downloads this month > 3. bonsai 2 27b heretic with the full 262k window, the speed head and vision all on > 4. gemma 4 26b a4b abliterated, moe with 4b active per token, 13.4gb ━━━━━━━━━━━━━━━━━━━━ 24gb, the danger zone · rtx 3090, 3090 ti, 4090 · amd rx 7900 xtx, and the 20gb 7900 xt fits these too · base qwen 3.8 27b q4 runs 41.3 tok/s on my 3090 with the mtp head, owners report 61 on a 3090 ti and 76 on a 4090 > 1. qwen 3.8 27b uncensored, q4 16.8gb, 2.31m downloads this month > 2. qwen 3.8 27b abliterated, q4 16.8gb, 2.12m downloads > 3. qwen 3.8 27b aggressive with the mtp head, 0 refusals out of 465, 2.06m downloads > 4. qwen 3.8 27b heretic, refusals 0 to 1 out of 100 > 5. qwen 3.8 27b fable cold fusion, the uncensored coder merge, 2.09m downloads > 6. qwen 3.6 35b a3b aggressive, moe with 3b active, 11.9m downloads all time and the most liked on this list ━━━━━━━━━━━━━━━━━━━━ refusal numbers are each releaser's own test, so read them per model, and the speeds are the base models on nvidia cards, these builds keep the same size so they run the same. amd cards run the same files on llama.cpp's vulkan or rocm backends, the bonsai builds run on prismml's llama.cpp no refusals means no guardrails, you own what you ask. bookmark this, find your card and pull one tonight

An AI agent makes a mistake early in a task, then keeps going in the wrong direction. Our researchers built PivotOPD to teach agents how to avoid those mistakes and recover when they happen. During training, a teacher model shows the agent a better action and how to get back on track over the next few steps. Read the paper and watch how it works: