ToolBraid🪢

ToolBraid Companion is live on the Microsoft Store 💪🏻 apps.microsoft.com/detail/9P… Built for developers. Browser tools + WebMCP on supported sites, straight from your Codex session. Free. You choose which sites your AI can access. Who’s up for trying it? 👀

I wish, but i don't have an 3060...

I haven’t tested this setup on a 3090. The 12-branch 5090 run peaked at ~30.3 GiB VRAM, so a 24GB 3090 needs a smaller memory budget and separate validation. TensorFold has Ampere fallback code, but I don’t have a measured Qwen 27B result on it yet.

Flash Next is next on my list of experiments.

Yes—thinking was ON (LOW) for all 12. Per-branch decode averaged 66–98 tok/s (median 80.5), measured over each branch’s own run. 791 tok/s is aggregate batch throughput; 1,081 was a short peak. This is dense 27B NVFP4, not Flash Next. Also, 64K is the configured ceiling; actual prompts were ~9.2K, mostly shared/cached.

They share the same project context, but each branch gets a different task: towers, bridge, garden, etc. Each returns code for its component. The coordinator then writes the files, assembles the scene and runs the browser checks. All on one GPU.

The live run hit a 1,081 tok/s peak over ~1 second and averaged 791 tok/s across the full batch. A separate controlled replay improved by 34.2%: 571 to 766 tok/s median, with identical generated token IDs. Thinking stayed LOW.

What's new: compact GPU state, scheduling calibrated on measured multi-branch costs and faster short prompt-tail processing. Complete proposals return as files, with per-task token budgets, so the coordinator can integrate all twelve.

The experimental 12-branch version is now on GitHub: Windows setup, native patches, OMP integration, tests and the scene workload. github.com/Maharajahu/qwen-p… Built on @ashxhart's TensorFold.

Local AI is getting wild Mate 🔥

Qwen 27b NVFP4 is driving on Tensorfold my 5090 like he stole it 😂

And believe I haven't reached the peak of it yet. It's still room for improvement. I will release the experimental recipe in couple of minutes

RTX 5090. 12 branches. Over 1,000 tok/s🤯 Qwen3.8-27B NVFP4 locally on my experimental TensorFold/OMP. 🚀 1,081 tok/s peak aggregate ⚡ 791 tok/s average across the generation batch 🧠 Thinking ON, LOW effort 💾 64K-context Local AI is getting wild @ashxhart and #TensorFold 🔥

Update: both recipes are now on GitHub. This shared-context setup and the earlier fanout run, with Windows instructions, native patches and the original videos: github.com/Maharajahu/qwen-p…

The recipe for this run is now public, alongside the newer shared-context version. Windows setup, native patches and the original recordings are all here: github.com/Maharajahu/qwen-p…

Two Qwen3.8-27B NVFP4 Windows recipes, one 32 GB RTX 5090: Fanout (thinking off): 661 tok/s Shared context (thinking low): 499 tok/s incl. thinking Aggregate across 8 branches; different workloads. Code, setup & videos: github.com/Maharajahu/qwen-p… Built on @ashxhart's #tensorfold

5/5 I also enabled low thinking with a 2,048-token reasoning budget and separated reasoning from code in the live terminal. The coordinator integrates the proposals, runs the checks and keeps everything in one conversation. All inference stays on the RTX 5090.

4/5 Synchronization now happens inside the engine. All eight branches finish preparing their prompts before decoding starts. Their first streamed tokens arrived within 3.7 ms of each other. That measures synchronization, not prompt-to-response latency.

3/5 I added shared GPU context. The eight branches reuse the same cached conversation prefix while keeping their own continuation state. In this run, each branch reused 8,192 tokens, reducing repeated prompt processing.

2/5 The biggest change I notice is responsiveness. With the prompts I’m testing now, the model starts working almost immediately. Longer jobs still need time to prepare the context before generation begins.

One RTX 5090. Qwen3.8-27B NVFP4. Eight synchronized branches sharing one context. With thinking set to low, I measured 499 tokens/s aggregate, including reasoning, with a 716.3 tokens/s peak over a roughly one-second window, with 66k context In this run, the agent generated eight Three.js components, integrated them and passed the browser checks in 101 seconds. Running locally with #TensorFold + #OMP. Live coding and the result below 👇