Intelligence Orchestrator, AI Software Engineer, Algo Trader, Builder, Shipper, Agentic Loop Engineer, CMP-4-LIFE 70hxer
Joined January 2021
- Tweets14.9K
- Following1.1K
- Followers1.8K
- Likes10.7K
DeepSeek V4.1 Flash 4× CMP 170HX optimized drop 🎙️⚡️
178 tok/s, one user. 587 tok/s across 8. 🔥
Prefill 1,570 tok/s at 2K.
KV cache 617K tokens. EXL3.
Full graphs and coop underneath, tensor-core MoE from row 43 up.
@Morrowmake
github.com/CurtisAccelerate/…
@ajambrosino copy deepseek harness agent view... its superior I can see whats going on vs codex hides everything leaves me in the dark
you are wrong-- we're running
DeepSeek v4 1017,
GLM 5.3 Flash,
DeepSeek is competitive with SOTA models from 6 months ago and in some areas better. And the models keep getting better.
Not every task demands the most SOTA AI
I hear you though and agree single GPU/card setups aren't able too deliver SOTA AI but they can still be useful....
Qwen4 might change that, we'll see.
your pissed that others are "cheating" with AI.
Reality check:
you were never playing the same game.
From the time of the camera obscura,
to professional artist that had own photography studios and traced
to professional game devs that had cross-assembler setups
to Olympic athletes that tech advantages in equipment, training, etc.
to traders that had speed advantages others lacked,
it was always "cheating" in every thing.
Why did you lose at everything?
Because most people aren't maximally optimizing.
But people judge themselves against others who were always maxxing.
You thought you were playing a certain game,
you bought into the kayfabe of talent,
Reality was the winners were playing a different game, altogether.
DeepSeek Harness trajectory-view is already way ahead of closed source Codex, Claude, etc. in offering agentic monitoring/interpretability at a glance.
DeepSeek.
Based.
Working toward AGI because they believe in it.
Contrast with Dario's stated beliefs,
- Believes AGI is dangerous
- Only working toward AGI because he believes only they are wise enough to harness it and only they should have it.
🔥重磅！DeepSeek研究工程师高熙在乌镇互联网大会上的发言，直接把DeepSeek真实目标给说透了！
为什么DeepSeek要把辛辛苦苦做出来的AI技术义无反顾开源给全世界？
高熙说得很干脆：“答案非常简单，就是为了AGI更快到来。”
他进一步补充：如果AGI本来十年后到，我希望它一年后到；如果本来一年后到，我希望它明天就到。
技术壁垒看起来能短期挡住对手、维持领先，但本质上也在挡住AGI的到来。
他重申：DeepSeek的竞争对手从来不是国内大模型公司，也不是海外的Claude、GPT，真正的对手只有一个：时间。
时间一分一秒在走，谁先实现AGI并不重要，重要的是它能尽快惠及全人类。
这才是开源的底层逻辑：发现就公开，让更多人一起验证、一起加速，而不是各自筑墙。
DeepSeek 格局拉满了！
#DeepSeek #AGI #开源 #乌镇互联网大会 #AI #高熙 #与时间赛跑 #梁文峰 #崔天意 #梁圣 #大胖鱼
I have had DeepSeek write large poems for testing token generation speed.
Tonight, on a real project it was searching for relevant material when it found a note it thought might be relevant. That note was simply the poem.
It thought:
The river remembers was a poem-- probably output of testing...
I never connected testing to that task or mentioned it.
Its not epyc but its fast
We've now reached up to 143 tok/s prose on Broadwell with significant gains in prefill speed and reductions in wall-time.
We now provide a docker image with configurable for speed, quality, or balance:
github.com/CurtisAccelerate/…
This update adds full 40-layer integer prefill preparation fusion, faster target execution and integrated draft sampling to our existing DeepSeek EXL3 serving stack. We retained our specialized expert shapes, split-attention and collective tuning, and Engram integration.
Selected improvements from @GNmac's recipe are stacked into that work, including short-row FP8 target paths and fast draft-head/sampling components. We also tested his adaptive drafting policy; our local results favored keeping fixed-five verification as the default. We don't claim a performance superiority.
Thanks @GNmac for building and sharing your recipe!
We've now reached up to 143 tok/s prose on Broadwell with significant gains in prefill speed and reductions in wall-time.
We now provide a docker image with configurable for speed, quality, or balance:
github.com/CurtisAccelerate/…
This update adds full 40-layer integer prefill preparation fusion, faster target execution and integrated draft sampling to our existing DeepSeek EXL3 serving stack. We retained our specialized expert shapes, split-attention and collective tuning, and Engram integration.
Selected improvements from @GNmac's recipe are stacked into that work, including short-row FP8 target paths and fast draft-head/sampling components. We also tested his adaptive drafting policy; our local results favored keeping fixed-five verification as the default. We don't claim a performance superiority.
Thanks @GNmac for building and sharing your recipe!