ごめんごめんDeepSeek-V4 Flashの実力はこんなもんじゃなかった。 引用元はReasoning Effort有効になってなかったのでReasoning Effort Maxで改めて。 これがDGX Spark × 2台でロスレス で Decode 〜46tok/sで動くん意味わからんよね。

Have you seen the qwen3.8 max

Qwen3.8-Maxの出力もイケてますね🍻 オープンウェイトだとさらに！

reasoningありでどれくらい時間伸びましたか？ 無しの状態で十分凄いのであんま伸びるようなら無しでも良いのかなと思いました

ほったらかしにしてたので意識してなかったですが、めちゃくちゃ時間伸びてました！用途に応じて使い分けですかね

だいぶ違いますね。用途によりますがreasoning無しで威力が強すぎるので早いし質も良く最高ですね！

Wait, do you mean the last one was set the reasoning to off? it's mindblowing for non-reasoning in 3d-creating!

What was the prompt

🍺

Why isn’t there an official pagoda bench site yet?

たしかに公式pagoda benchサイトほしいですねw

I just want to see alibaba release pagoda maxed models

DGX Spark双卡无损解码46 tok/s，配合官方FP4+FP8混合精度，已经接近部分云端推理的性价比区间。家庭本地跑大模型真正进入“可日常使用”阶段，而不是只能跑Demo。

周一卖出 AI 股票

wow amazing and its literally free to use - lets goo

And with DGX Spark × 2, DS can actually run agents for extended periods without issues?

用途次第じゃないですかね？

Nothing can justify the price of me buying two sparks for my personal use :(

@ivanfioravanti make sure to verify the reasoning levels

性能も凄いですが、 一番の衝撃は「その性能をローカルで出している」ことなんですよね。

@grok what is shown on this video? What does it do? Why is it used as a showcase of ds4-flash’s capability? is what is shown on the video useful?

この検証、待ってました🔥 夢がありますねえ👀

46 tok/s on 2 DGX Sparks lossless is the bar. Local reasoning-class inference finally hit the point where you don't pick between quality and not-waiting-forever. Reasoning Effort Max is the new default — anyone still shipping base-only is leaving 30% of capability on the table.

it's fully unrestricted now see here:

This is actually better than Fable and Codex both I tried. This is INSANE dude!

Hi, I would like to See 27B with reasoning effort at Max

This result shows why reasoning effort belongs in every comparison. More reasoning may improve quality, but latency and cost move with it. The pricing and Codex walkthrough:

Amazing