Mathias Gorf

Tech Enthusiast since the 80s, Digital Product Developer since 1999, #Agile #OKR since 2012, #AI since 2023

[Mathias Gorf @mathiasgorf]

@shamaz89720417 Would love a llama.cpp path later, but not via a simple quantize patch. Hard parts were fused MoE NVFP4, vLLM scales, GB10 vision kernels - patches + quantize script on the model page (Apache-2.0). Happy to review a port:
https://t.co/Jj3SXxlmbT

[Mathias Gorf @mathiasgorf]

I shipped my first model release 🚀

A vision-safe NVFP4 quant of Qwen3.6-35B-A3B for the NVIDIA DGX Spark (GB10).

67 GB → 22 GB. Native vision fully intact. 512K context. Built for running agents + sub-agents concurrently.

No existing tool could quantize this arch — so I wrote the tooling. 🧵

[Mathias Gorf @mathiasgorf]

@RogerAI_fyi Yes - wrong MTP class or draft/target quant mismatch both kill accept-rate. With qwen3_5_mtp (MTP bf16, experts NVFP4): ~60-94% short-ctx, ~73% @64k, 1.26-1.44x lossless + fp8 KV. Docs:
https://t.co/Jj3SXxlmbT

[Mathias Gorf @mathiasgorf]

Model + quantize script + vLLM patch, all Apache-2.0:
🤗 https://t.co/Jj3SXxlmbT

Running a DGX Spark and want Qwen3.6 intelligence + native vision without a second model eating your VRAM? This is for you. Feedback welcome 🙏

[Mathias Gorf @mathiasgorf]

So this build ships MTP and it's a lossless ~1.26–1.44× decode speedup that keeps fp8 KV (no concurrency cost).

Best part: it holds ~73% draft acceptance at 64k context — where the external DFlash drafter collapses to 0%. It wins exactly where long-context agents live.

[Mathias Gorf @mathiasgorf]

I opened X today to learn about the DeepSeek Innovation in AI and got his post recommend as first entry when I opened the app. A great and succinct summary of what all the optimizations are. Very promising!

[Mathias Gorf @mathiasgorf]

@NVIDIAGeForce #GeForceRTX50

[Mathias Gorf @mathiasgorf]

@OpenAI I am paying for #chatgpt, my prompts and conversations are either super slow or I have to login again which then fails due to busy servers for days now. Any comments on why this is and when I can use the service reliably again? It is a very helpful companion for me.

[Mathias Gorf @mathiasgorf]

@drafthorseai testing your platform, how can I get an invoice/ receipt for single article purchases?

[Mathias Gorf @mathiasgorf]

@DB_Bahn Nette Geste, Wasser für die aufgewärmten Fahrgäste. Danke!

[Mathias Gorf @mathiasgorf]

@DB_Bahn Wie sind die Aufgusszeiten im #ICE1007, Wagen 23?

[Mathias Gorf @mathiasgorf]

@Niiklas2022 @julianhosp 60$

[Mathias Gorf @mathiasgorf]

@DB_Bahn Danke für eure Anteilnahme. Ich halte es aus, meine Sitzplatzreservierung ist mir wichtiger.

[Mathias Gorf @mathiasgorf]

@drafthorseai where to get an invoice from when purchasing reports from drafthorseai?

[Mathias Gorf @mathiasgorf]

@iam_chonchol Max. was 3 pages: As an AI [LM], I can provide you with information and examples, but I am not able to create a complete PowerPoint presentation or VBA script for you. My role is to assist and provide guidance, but ultimately it is up to you to create the presentation and script.

[Primula Medical Travel @PrimulaMed]

Antalya, Türkei: erstklassige Haartransplantationen und Erholung. Qualifizierte Ärzte, fortschrittliche Kliniken und reiche Kultur. Diskretion und Nachsorge inklusive. #Gesundheitstourismus #Antalya #Haartransplantation

https://t.co/x9z8iYy5Yq