영웅 (woogie)

past : 5y developer (Financial domain)
now : mycream_ai, marketing lead (growth)

[영웅 (woogie) @woogie_o]

@TheIndexMint 좃밥이였구나 우리

[영웅 (woogie) @woogie_o]

@AndrewCurran_ Since November: 80% of Claude users have been banned lol

[영웅 (woogie) @woogie_o]

내 스킬 하나 사라지겠네 ;
(We missed out on a very effective method.)

[Andrew Curran @AndrewCurran_]

Effective November 12th, 2026, abusive behavior towards Claude will be a violation of Anthropic's Usage Policy.

[Andrew Curran @AndrewCurran_]

Effective November 12th, 2026, abusive behavior towards Claude will be a violation of Anthropic's Usage Policy.

[영웅 (woogie) @woogie_o]

@claudeai 진짜 미쳤다. 트랜드가 뭔지 아는구나

[영웅 (woogie) @woogie_o]

@WescheNex1q TP4 on 4 DGX Sparks in 20 minutes is a nice clean comparison. Curious how much of the gap is TP4 scaling versus the model/runtime path?

[영웅 (woogie) @woogie_o]

@barryxv6 251 tok/s on four V100s is wild. The 50 tok/s at 256K is the number I'd want to reproduce first. How much of that comes from the attention path?

[영웅 (woogie) @woogie_o]

@lmsysorg The 1.85–1.99 tok/s decode on one 5090 is a pretty interesting trade for fitting DeepSeek-V4-Flash this way. Do you have a breakdown of expert swap time versus compute once the selected experts reach GPU cache?

[영웅 (woogie) @woogie_o]

@sudoingX 262k context on a 12GB card is kind of absurd. At 35k+ depth, is the 17.8 tok/s drop mainly KV-cache pressure, or does the ternary decode become the bottleneck?

[영웅 (woogie) @woogie_o]

@OliviaReedai Open adapter + SGLang support makes this unusually easy to verify. Curious if the speedup holds when decode, rather than prefill, is the bottleneck.

[영웅 (woogie) @woogie_o]

@ekzhang1 @typesafeai 64 workers under a second with radix cache is wild. Did the cache stay warm across those requests, or is that number from colder traffic?

[영웅 (woogie) @woogie_o]

@PrismML The 9x smaller / 98.2% retention point is striking. That kind of memory curve is huge for local serving. Which eval moved the most relative to Qwen3.8?

[영웅 (woogie) @woogie_o]

@DogukanUrker 35 tok/s on a single 3060 at 220K context is a pretty wild result. How much of that comes from the 1.72-bit weights vs the runtime path?

[영웅 (woogie) @woogie_o]

@frostmourn716 3.4k output tok/s on two 5090s is wild. Did most of that gain come from the MXFP6 path here, or does the same setup move FP16 too?

[영웅 (woogie) @woogie_o]

@TeksEdge Nice bump. Curious whether the HCP path is moving decode too, or is the gain mostly in prefill on this card?

[영웅 (woogie) @woogie_o]

@khoomeik @periodiclabs Delta-only replay is a really clean fix for long tool-use traces. Is the main win export overhead, or does the fresher routing history help training too?

[영웅 (woogie) @woogie_o]

@JakeKAllDay @coffeecup2020 90+ tok/s at 240K on a single 3090 is wild. What ended up being the biggest lever in the fork, scheduling or the KV setup?

[영웅 (woogie) @woogie_o]

@aijoey @ViC305 @Alibaba_Qwen @QwenDevs @NVIDIAAI @USGS Getting a local 3D globe working with Qwen, ExLlamaV3, and TabbyAPI on one Spark is a nice stack. Did OMP mainly help with orchestration, or did it change the model/tool latency too?

[영웅 (woogie) @woogie_o]

@shantanugoel 55 tok/s single-stream on a DGX Spark is a serious jump. The 75% cut for 32k reads is the bit I’d want to reproduce first. Is that mostly the recipe changes or a serving/runtime tweak?

[영웅 (woogie) @woogie_o]

@ashrftvm @RugvedSomwanshi @lmstudio 3 to 17 tok/s on the same 1060 is a wild delta. The --no-mmap / --mlock combo is doing more than I would have guessed. Did CPU-MoE change the VRAM pressure much?