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? 👀
3
371
Replying to @SandhuJonathan
Man I don't have to show you anything 🤣.
If you like my work and support local ai interface then Start testing stuff yourself, don't waste my time challenging me or all the other contributors from all different projects. My focus is to unlock local hardware potential. This is the key point you keep missing. Do something constructive dude, i get it you're doing it for engagement and impressions, but if you haven't tested yourself then you're unable to provide a constructive feedback.
Unreleased beta engine at 128K context.
A single RTX 5090. 🚀
I've pushed Qwen 27B NVFP4 on #TensorFold to previous 309 to 326.93 tokens/ in a real cold-decode test.
Part that matters most: 2,048 generated tokens.
Byte-for-byte identical output. Zero output drift.
Replying to @mimikku3080
Your AMD CPU controller it's shit. I got the same problem with my 99503DX.
Replying to @SandhuJonathan
The aim was to test local inference performance and agent behavior with the new BLACKFURY kernel on TensorFold. BenchLedger was the workload. App crash recovery and bug repair are separate tests. Users choose workloads relevant to their own use cases.
1
13
@SandhuJonathan does this real agentic work satisfying your expectations?
1
20
5/5
The agent reported 6/6 public tests and 132/132 of its own tests.
My independent check: 6/7 hidden suites passed. One exposed a real consistency bug.
Promising, but not flawless.
29
4/5
Across the run:
• 169 successful requests
• 277,887 generated tokens
• ~2,574 tok/s fresh-token prefill
The ~344 tok/s decode average is total output divided by generation time. It excludes prefill and tool execution. Context length varied.
1
30
3/5
The setup:
• TensorFold inference engine
• NVIDIA Qwen3.8-27B-NVFP4
• FP8 KV cache + DFlash2
• One RTX 5090 (32GB)
• Claude Code, low thinking effort
• Real tools, including sandboxed Bash
Peak input: 118,414 / 131,072 tokens.
1
44
2/5
The task: build BenchLedger, an offline benchmark registry that syncs data across machines.
SQLite migration. CSV/JSON imports. Duplicate events. Replica conflicts. Crash recovery. Reports.
The challenge: keeping the data consistent through all of it.
2
52
~344 tokens/sec average decode. ONE RTX 5090.
26 minutes of local AI coding in Claude Code.
THREE automatic context compactions.
Qwen3.8 27b NVFP4 - The agent kept working. 🔥
Experimental Unreleased BLACKFURY Exact on #TensorFold
Setup + independent test results below 🧵
Replying to @martinwanderer @ashxhart
Go for it Mate, load the model set-up a smaller context that fits with in your vram and let it cook
1
13
Replying to @TacticaLJes @ashxhart
We are already another 10% above 10%😁. Still working on making it stable.
13
Replying to @SandhuJonathan @ashxhart
First get steep, make it fast, then polish it. Either way the model works as normal, tool calling and the rest of the stack works as normal. As nothing has been changed on that side. Each part of the kernel delivers everything byte by byte, I'm only optimising each branches from the kernel so the model knows better how to communicate with the GPU.
18
Replying to @Night_Diver2543 @ashxhart
My guy, that's a demo on 2.8k - 4.7k tok... A demo has no utility in real world use case. Go with 128k context and get back to me with real work demo 👀
1
56
Replying to @Yves_From_BE @ashxhart
I'm sure that in the future Ash can push it to 300tps
It's so much room to optimise our local hardware that ee haven't explored yet
114


