Kurt Schwitters

Politisch : grünlinks ver(s|k)ifft
Sprachen: de, en,ru: я могу говорить и ругаться только со швабским акцентом!

[Kurt Schwitters @kurt_dwitters]

Die Quelle des Videos ist leider verschwunden/К сожалению, источник видео исчез.

[Kurt Schwitters @kurt_dwitters]

@MirovichMedia Позвони-ка этому умнику из Санкт-Петербурга и передай ему, что он уже давненько кое-что неправильно понимает

[Kurt Schwitters @kurt_dwitters]

@Kachelmann Wann sprechen Sie eigentlich mal wieder mit Ronzheimer?

[Kurt Schwitters @kurt_dwitters]

@by_kalesnikava @Tsihanouskaya @POTUS @johnpcoale Besuch doch mal deinen alten Studienort. Ich glaube, viele Leute haben während deines Gefängnisaufenthalts mit dir mitgelitten – auch diejenigen, die Russisch sprechen. Viele Grüße!

[Kurt Schwitters @kurt_dwitters]

If you're unlucky enough to be stuck with an Informix database, check out https://t.co/nuUSpLgK0S — an independent implementation of the horrid Informix wire protocol.

[Kurt Schwitters @kurt_dwitters]

@wieslawsoltes Hehe, I can already hear the EPLAN folks wailing and the sales guys sweating. Tell your assistant he should make EPLAN for humans.

[Kurt Schwitters @kurt_dwitters]

@wieslawsoltes I trust you exiled Lisp and COM from your design doc on sight — can't have that kind of riff-raff cluttering up a serious architecture.

[Kurt Schwitters @kurt_dwitters]

Ach ja, bevor ich's vergesse – kurz in eigener Sache: Wer seine SolidWorks-Modelle mal zumindest auf Metadaten-Ebene inspizieren und daraus eventuell Stücklisten ziehen will, wird hier fündig: https://t.co/bCF5f6VwS5

[Kurt Schwitters @kurt_dwitters]

https://t.co/Vxcj9WdFhx — big shoutout to my very best friends in the CAD mold industry, still lovingly perched atop their sacred 'trade secrets.' Truly inspiring stuff.

[SudoX7 @sudox7]

wrote a playable bare-metal Snake game packed inside a 512-byte BIOS bootsector

just 16-bit real mode assembly, BIOS keyboard polling via INT 0x16, and direct memory writes to VGA text-mode RAM (0xB800)

[Alok @analogalok]

The VRAM barrier is officially dead.

I just ran Qwen 3.8 Flash Next (MoE) 125B A6B with a 250,000 context window on a single 24GB RTX 4090.

21 tokens/sec decode. 364 t/s prefill.

no mtp. no dflash. no kv cache quantization!

We are running datacenter models on consumer hardware.

Tested on Ubuntu 22 | CUDA 13.0 | PCIe 4.0 x16 | 110 GB DDR4 System RAM with a continuous 28k prompt across all runs.

### The Benchmarks & Scaling

# 1. Hybrid Offload (-ncmoe 40 @ 80k Context)
Offloaded 40 expert layers to the GPU, pushing VRAM to the ceiling.

./build/bin/llama-server -m Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf -c 80000 --port 8080 -v --fit off -b 4096 -ub 4096 -ncmoe 40

Prefill: 383.85 t/s | Decode: 22.52 t/s
Footprint: 23.85 GB VRAM | 97 GB RAM

# 2. Full CPU MoE Offload (-cmoe @ 80k Context)

Pinned all 512 expert layers to DDR4 RAM (-cmoe), keeping attention on the 4090.

llama.cpp flags: (Same as above, replace -ncmoe 40 with -cmoe)

Prefill: 355.72 t/s | Decode: 20.84 t/s
Footprint: 11.66 GB VRAM (12GB+ VRAM freed up!) | 110 GB RAM

# 3. The 180,000 Context Run

Prefill: 357.75 t/s | Decode: 20.98 t/s | VRAM: 15.6 GB | RAM: 110 GB

# 4. The 250,000 Context Absolute Ceiling

./build/bin/llama-server -m Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf -c 250000 --port 8080 -v --fit off -b 4096 -ub 4096 -cmoe

Prefill: 364.29 t/s | Decode: 20.97 t/s

Footprint: 18.3 GB VRAM (Still ~5.7 GB of VRAM headroom!) | 110 GB RAM

### Key Insights:
-b 4096 -ub 4096: doubles the prompt ingestion from ~150 to 364+ t/s.
-cmoe Free Lunch: Shifting expert layers to DDR4 RAM slashes VRAM from 24GB to 11.6GB with virtually zero decode penalty (22.5 -> 20.9 t/s), enabling the 250k context ceiling.

Qwen 3.8 Flash-Next (UD-Q4_K_XL) is a massive 111.4 GB model split across 4 shards. To run this architecture, you must build from the experimental PR branch (#27742) by @danielhanchen:

git clone && cd llama.cpp
git fetch origin pull/27742/head:qwen-next && git checkout qwen-next
cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native -DBUILD_SHARED_LIBS=OFF
cmake --build build --config Release -j $(nproc) --target llama-server

A single 4090 paired with 100 GB of cheap DDR4 RAM will comfortably serve production grade 125B inference.

While Qwen 3.8 27B (dense) still holds the crown for single 3090/4090 rigs, Flash Next proves 125B hybrid models are officially viable on consumer hardware.

Hugging Face GGUF link and complete performance telemetry graphs are dropped in the replies below.

GLM 5.3 Flash VS Qwen 3.8 Flash Next, which one takes the open weights crown this week?

[Пан Пачковский @Q0MT6pFmbVqynsM]

Европа загнивает, русские процветают. И это можно отлично обсудить в очередях за индийским 95м

[Kurt Schwitters @kurt_dwitters]

@yuliaisc https://t.co/bCF5f6VwS5

[Kurt Schwitters @kurt_dwitters]

@RaminNasibov

[Kurt Schwitters @kurt_dwitters]

@vgmax На мой взгляд, коровы и быки — идеальное средство для снижения интенсивности дорожного движения, например, возле школ или детских садов.

[Kurt Schwitters @kurt_dwitters]

@sterntaler1970 Kommst du auch nach Süddeutschland?

[Kurt Schwitters @kurt_dwitters]

@vgmax Подождите, скоро он захочет спать у вашей кровати!

[Kurt Schwitters @kurt_dwitters]

@mikailme Бензин на 3 дня

[Kurt Schwitters @kurt_dwitters]

@Zn_Portnova Бензин на 3 дня

[Kurt Schwitters @kurt_dwitters]

@Q0MT6pFmbVqynsM Помимо видео с ударами дро��ов — новый жанр, позволяющий научиться ругаться по-русски...