Jürgen Schmied

josch15366's activity

• All
• Models
• Datasets
• Spaces
• Buckets
• Papers
• Collections
• Community
• Posts
• Upvotes
• Likes
• Articles

New activity in josch15366/Qwen3.8-Flash-Next-NVFP4-LocalHessian-Experts about 7 hours ago

Calibration corpus: what we got wrong, and the one thing worth copying (position stratification)

#1 opened about 7 hours ago by

josch15366

New activity in RadixArk/Qwen3.8-Flash-Next-NVFP4 about 7 hours ago

Per-expert weight_scale_2 is worth ~0.9 % held-out NLL (and two contracts you got right)

#13 opened about 7 hours ago by

josch15366

New activity in RadixArk/Qwen3.8-Flash-Next-NVFP4 10 days ago

Recipe for vllm

🔥 1

#9 opened 10 days ago by

josch15366

New activity in tcclaviger/Qwen3.8-27B-DFlash2-FP8 19 days ago

rationale for this build

3

#1 opened 19 days ago by

josch15366

New activity in syvai/Qwen3.8-27B-DFlash2-W4A16 20 days ago

Best speed in fleet

1

#1 opened 20 days ago by

josch15366

New activity in RadixArk/Qwen3.8-27B-NVFP4 20 days ago

The DFlash2 restriction behind the BF16 lm_head change was removed upstream a day earlier

2

#3 opened 20 days ago by

josch15366

New activity in unsloth/Qwen3.8-27B-NVFP4 20 days ago

Any chance of Unsloth's Qwen3.8 27B NVFP4 DSPARK/DFLASH2?

12

#20 opened 23 days ago by

MaMo7x

New activity in RadixArk/Qwen3.8-27B-DSpark 21 days ago

Drafter is for fp8 and does not work on nvfp4

➕ 5

3

#4 opened 25 days ago by

josch15366

New activity in Inferact/Qwen3.8-27B-NVFP4 22 days ago

unsloth/Qwen3.8-27B-NVFP4 vs. Inferact/Qwen3.8-27B-NVFP4?

🔥➕ 3

10

#1 opened 28 days ago by

pathosethoslogos

New activity in RadixArk/Qwen3.8-27B-DSpark 25 days ago

[BUG] DSpark slower than EAGLE + severe cold-start penalty

7

#3 opened 25 days ago by

voves

New activity in victor/Qwen3.8-27B-free-endpoint 25 days ago

Have you considered RadixArk/Qwen3.8-27B-DSpark vs native MTP under concurrency?

5

#1 opened 27 days ago by

NodeLinker

New activity in froggeric/Qwen-Fixed-Chat-Templates 26 days ago

QWEN38: `reasoning_effort=xhigh` is an unsafe default: it can consume the entire token budget and return empty content

👍 4

2

#72 opened 27 days ago by

josch15366

New activity in z-lab/Qwen3-Coder-Next-DFlash 2 months ago

Works as GGUF also

❤️ 2

#1 opened 2 months ago by

josch15366

New activity in GestaltLabs/Ornstein3.6-27B-MTP-NSC-ACE-SABER-GGUF 3 months ago

Benchmark with mixed results

#4 opened 3 months ago by

josch15366