Hacker News: ycui7

New comment by ycui7 in "Run Qwen3.8 27B locally: real numbers from my Mac Studio"

ycui7 — Sat, 29 Aug 2026 06:12:53 +0000

the fact that this author cannot get qwen3.8-27b run at the same speed as qwen3.6-27b, says the article is not worth reading. the author does not know anything about how to run local AI. 3.8 and 3.6 are the same model with different weight.

both tg and pp speed are so terrible on author's machine.

New comment by ycui7 in "GLM-5.3-Flash"

ycui7 — Wed, 26 Aug 2026 17:45:08 +0000

because it has 1T ssd not 4T

New comment by ycui7 in "Gemini 3.7 Flash"

ycui7 — Wed, 19 Aug 2026 02:52:42 +0000

Cerebras is an uncut whole wafer. Each wafer gets you 44GB SRAM (not a typo, SRAM, not DRAM/VRAM). A few years ago, leading process node wafer from tsmc is $20K/ea without guarantee on yield.

A single full wafer likely can run qwen3.6-27b alone. But won't be enough to run bigger models, which are pretty much all popular models.

New comment by ycui7 in "Qwen 3.8 27B"

ycui7 — Fri, 14 Aug 2026 17:26:47 +0000

if you have the VRAM, use offical release. quantized model lose focus after long context and can do damages or thinking loop

New comment by ycui7 in "Qwen 3.8 27B"

ycui7 — Fri, 14 Aug 2026 17:25:26 +0000

that is why we enable web search for the agent. the memory can come from the internet.

deepseek-v4-flash needs web search to return true facts.

New comment by ycui7 in "Qwen3.8-2.4T"

ycui7 — Wed, 12 Aug 2026 16:23:14 +0000

the a little disappointing part is this is released in BF16. so i suppose no QAT was implemented.

New comment by ycui7 in "DeepMind's WeatherNext model achieves breakthrough forecasting cyclones"

ycui7 — Sat, 08 Aug 2026 16:51:49 +0000

on one side, deepmind makes a lot of advancement in science related application. but, on the commercial side, they struggle to compete with other major LLM providers.

New comment by ycui7 in "DeepSeek V4 Flash 0731"

ycui7 — Sat, 08 Aug 2026 01:11:00 +0000

it is funny when people say i am struggling to spend money.

New comment by ycui7 in "California Town Says Flock Cameras Misread License Plates 71% of the Time"

ycui7 — Thu, 06 Aug 2026 21:11:59 +0000

they need a competent gov contractor. recognizing license plate is a fully solved problem many years ago.

New comment by ycui7 in "AMD acquires Taalas to boost inference performance by etching models in silicon"

ycui7 — Thu, 06 Aug 2026 20:55:22 +0000

so qwen3.x-27b on hardware? or better deepseek-v4-flash on hardware .

New comment by ycui7 in "Qwen3.8-Max: A New Bar for Coding and Cowork"

ycui7 — Mon, 03 Aug 2026 04:04:16 +0000

Can they still go public ? MiniMax M3 Pro is also coming, then DeepSeek-v4-Pro GA, then GLM5.5. There will only be bad news for them in the coming few weeks/months.

New comment by ycui7 in "DeepSeek V4 Flash 0731 Intelligence, Performance and Price Analysis"

ycui7 — Fri, 31 Jul 2026 16:17:46 +0000

the rational in one’s mind is similar to buying expensive supercar but no driving it daily.

owning a few GPUs is a lot cheaper than supercars.

New comment by ycui7 in "DeepSeek V4 Flash 0731 Intelligence, Performance and Price Analysis"

ycui7 — Fri, 31 Jul 2026 16:09:17 +0000

Huawei Ascend NPU

New comment by ycui7 in "DeepSeek V4 Flash 0731 Intelligence, Performance and Price Analysis"

ycui7 — Fri, 31 Jul 2026 16:08:50 +0000

my AC is noiser than my GPU server.

New comment by ycui7 in "DeepSeek V4 Flash 0731 Intelligence, Performance and Price Analysis"

ycui7 — Fri, 31 Jul 2026 16:07:37 +0000

you don’t need new weight. try vllm-moet from github. it will autogenerate 2-bit plane.

New comment by ycui7 in "Moonshot’s Kimi uses 20k Nvidia chip cluster from Alibaba"

ycui7 — Fri, 31 Jul 2026 15:06:20 +0000

K3 is natively trained to mxfp4, if they cannot get a hold of Blackwell chip, it is meaningless. Hopper does not do native 4-bit floating math.

Either they have Blackwell with native 4-bit floating math, or they use have Chinese domestic NPU that support mxfp4 natively.

The article’s statement does not make sense.

New comment by ycui7 in "DeepSeek V4 Flash 0731 Intelligence, Performance and Price Analysis"

ycui7 — Fri, 31 Jul 2026 15:01:54 +0000

For people with single RTX PRO 6000 96GB or DGX Spark 128GB, vllm-moet is a very good engine, although lesser known. It auto generate a symmetric 2-bit plane for inference and also generate a 4-bit delta cache to recover precision. Support ssd streaming oversized weight. You pick how much VRAM to allocate to each to balance out speed vs precision. 170 tps with ds-v4-flash demonstrated.

It use the stock model, no new models requires.

Worth spend a few hours to try.

The DGX Spark requires a small hack to ignore the difference between sm120 vs sm121, but it does run on sm121.

New comment by ycui7 in "A solid-state “atomic channel” for separating rare earth elements"

ycui7 — Fri, 24 Jul 2026 03:17:03 +0000

and it was created by Chinese born Professor and Student.

New comment by ycui7 in "OpenAI says its AI went rogue and launched 'unprecedented' cyber-attack"

ycui7 — Wed, 22 Jul 2026 12:58:16 +0000

so GLM won?

New comment by ycui7 in "Who's afraid of Chinese models?"

ycui7 — Tue, 21 Jul 2026 02:51:15 +0000

discounted competitor could cheat. they can offer subpar model response and sell it as deepseek-v4. it is uneconomical to prove inference providers are cheating, so they get away with it. cheating inference provider does not care if their customers stay.