NVIDIA Announces New Jetson Thor Modules "T3000" and "T2000" — Taking Aim at Mass-Market Robotics

On July 15, NVIDIA announced two new modules based on the Jetson Thor architecture: the T3000 and the T2000. The T3000 delivers 865 FP4 teraflops of AI compute in a compact form factor, coming in at roughly half the size and power consumption of the T5000. With 32GB of LPDDR5X memory, 273GB/s of memory bandwidth, and 25 GbE connectivity, it targets volume deployment of humanoid robots and autonomous machines.

The T2000, meanwhile, offers 400 FP4 teraflops and 16GB of memory, positioning it as an entry point for a broader range of edge AI applications. The overall Jetson portfolio now scales from 70 TOPS all the way to 2,000 teraflops.

What stands out is the automated memory optimization enabled by the new "Jetson agent skills." AI agents optimize the entire software stack, freeing up as much as 15GB of memory. As a result, UBTech and Agile Robots can now migrate from 64GB configurations down to 32GB modules. With memory prices continuing to climb, the approach offers a practical way to cut costs.

NVIDIA also announced Cosmos 3 Edge, a lightweight 4B-parameter world foundation model. It runs on-device inference on the Thor platform and can reportedly be post-trained for a specific robot implementation in about a day. An emulation mode for the T3000 is due later this month, with one for the T2000 to follow, and the modules themselves are slated for release in Q1 2027.

Microsoft Teaches Sales Staff the "Weak Points" of OpenAI and Anthropic — The AI Model Race Enters Its Next Phase

Microsoft has reportedly been training its sales staff to tell customers that OpenAI and Anthropic models "fall short on efficiency and cost" as part of a push to sell its in-house AI models. The company positions its own models as a more efficient and cost-effective alternative to the competition.

Even as Microsoft pours billions of dollars into OpenAI, it keeps building out its own family of AI models — a sign that the two companies' relationship has settled into a parallel state of "cooperation and competition." As price competition and differentiation intensify in the AI model market, how major cloud providers make the case for their proprietary models is likely to become the next focal point.

Kimi K3 Release Imminent? — Anticipation Builds in the Community

Speculation is spreading that the release of Kimi K3, the next-generation model of Moonshot AI's conversational AI Kimi, is just around the corner. Kimi's official X account has posted a video peppered with "3"s, and reports suggest it is already being tested on LM Arena under the codename "kivine."

Community reactions include remarks that it is "good but slower than Fable," so the performance debate is already underway ahead of launch. Given how well Kimi K2 was received, all eyes are on how much the successor improves.

Hy3's 1-Bit Quantization "Holds Up Better Than Expected" — A Glimmer of Hope for Running Large Models Locally

Test results for a 1-bit quantized build of the Hy3 model (89–93GB) are drawing attention on Reddit's LocalLLaMA community. The poster judged it "surprisingly coherent even at the minimum configuration" and confirmed its practicality by generating an HTML-based flight simulator and creating SVG illustrations.

As models grow ever larger, the finding that extreme compression like 1-bit quantization can still preserve a decent level of quality looks like a promising lead for running LLMs under hardware-constrained conditions.

SDABench — How Ready Are LLMs for "Scientific Discovery"?

A new benchmark, SDABench, evaluates LLMs' scientific discovery capabilities across six capability axes (descriptive, exploratory, inferential, predictive, causal, and mechanistic) and five domains (biology, chemistry, environmental science, geography, and physics). After testing 15 representative LLMs, it found that while they handle descriptive analysis well, performance drops sharply on tasks that require hypothesis selection, latent-process modeling, or mechanistic reasoning.

More capable models get better at identifying the relevant scope and variables, but they still struggle to select appropriate analysis procedures, model relationships between variables, and draw valid conclusions. At a time when AI is increasingly positioned as a partner in scientific research, this stands out as an important contribution for exposing the limits of LLMs' "scientific thinking."

---