🤯 This is like having Jev at home, a AI reflex system running locally.
Someone is turning Google's open DiffusionGemma 26B-A4B into a local version of the new “System One” AI idea inside vLLM.
Remember Jev? It is a decision LLM. Instead of asking an LLM to generate paragraphs, System One models are designed to make FAST structured decisions:
🎯 yes / no
🧭 route A / B / C
🛠️ which tool to call
🚨 severity 1–4
📊 classify / score / choose
And DiffusionGemma has an important role because a normal autoregressive LLM takes Prompt ↓ token ↓ answer
DiffusionGemma is Prompt ↓ predefined answer slots ↓ fill multiple decisions in parallel
Because it operates over a token canvas with bidirectional attention instead of being forced to generate everything strictly left-to-right.
The vLLM patch lets it return
✅ bounded choices
✅ probabilities
✅ confidence / uncertainty
✅ multiple decisions simultaneously
And it runs locally.
On ONE DGX Spark the developer reports
⚡ 1 request: ~120 ms
🚀 concurrency 32: ~54 requests/sec
🧠 3 decisions/request
🔥 ~162 decisions/sec
Then the developer found each question can require only around 3 canvas tokens:
🔢 index
❓ placeholder
✂️ separator
Meaning as many as ~85 questions could fit into 1 diffusion canvas.
And subsequent optimization reportedly cut decision time another:
⚡ ~20–40% with no change in decision quality.
The model stats ...
🧠 25.2B total parameters
⚡ 3.8B active
💾 NVIDIA NVFP4: ~18.9GB
🎮 Can fit on a 24GB NVIDIA GPU
📜 Open weights
⚠️ vLLM PR #
57250# is still OPEN 👉 not merged.
And this is Jev-like functionality. It is NOT evidence that DiffusionGemma matches Jev's intelligence or calibration.
🔗 /vllm-project/vllm/pull/57250
Show more
A few thoughts on $NVDA:
1) When it comes to product engineering/R&D, Nvidia might still be in good shape post-Jensen if it appoints the right successor. The company has a ton of engineering and management talent and an unmatched R&D budget, and its many hardware, software/ecosystem and supply-chain strengths don’t suddenly disappear with Jensen’s departure.
2) I think it’ll be much harder for anyone to fill Jensen’s shoes as the public face of Nvidia. Expecting anyone to match Jensen’s ability to form strong ties with politicians, customers and supply-chain partners, as well as to sell Nvidia’s story to developers, the media, investors and the general public, all while running one of the world’s biggest tech companies, is a tough ask.
3) Long-term, I think the loss of Jensen’s big-picture thinking and willingness to make expensive long-term bets with uncertain outcomes in support of his vision will almost certainly be felt. From making huge AI/HPC software and ecosystem investments long before ChatGPT arrived, to buying Mellanox and gradually turning Nvidia into a full-stack accelerated computing provider, to aggressively using Nvidia’s balance sheet/FCF to lock up supply-chain capacity and support neoclouds (who, unlike the tech giants, aren’t developing custom ASICs), Nvidia is where it’s at today due to Jensen’s ability to see where the puck is going and willingness to bet aggressively in that direction, all while while having the full trust and support of investors and the board to do so.
Leaving aside strategic vision, will a post-Jensen Nvidia be able and willing to make such difficult and expensive long-term bets? For now, I think it’s worth watching how committed Nvidia remains to following through on some of Jensen’s more recent long-term bets, whether related to supply-chain and neocloud support or major R&D and M&A investments in areas such as robotics and open-weight models.
4) Nvidia now trades for 10.4x its FY28 (Jan. ‘28) EPS consensus. That’s a lower forward EPS multiple than what it fell to during either the COVID crash or its late-2018 selloff (which happened shortly before crypto-related demand plunged). It’s also around half of $AVGO’s forward EPS multiple, less than half of $MRVL’s, and less than one-third of $AMD’s.
Even if one grants that a portion of this discount predates Jensengate and likely stems from worries about customer/neocloud financing, ASIC competition and China-related sales, markets are taking a pretty dim view of Nvidia post-Jensen at a time when the company still has a lot of strengths, companies up and down the AI value chain keep saying they’ve got a lot more demand than they can handle, and LLMs are improving at a very rapid pace.
On the flip side, Nvidia is far from the only AI infra stock trading for low multiples right now (deals abound across memory, optics, semicaps, PCB makers, industrials, etc.), and those companies didn’t just abruptly lose a visionary founder/CEO. I can easily see the stock working in time from these levels, but some other names might have better odds of rallying strongly in the short-term.
Show more