What Nvidia actually announced
Nvidia describes Underdog as a new desktop agent app built for fully local, private AI, now coming to RTX Spark. It is powered by a Husky model-specific inference engine and runs models entirely on the device. The stated benefit is that personal agents can get work done without user data ever leaving the PC. Nvidia says Underdog is available now for download.
The wording is "desktop agent app" and "inference engine". It is not "model library manager". Nvidia's announcement doesn't document a model browser, update handling, or other management features. It also doesn't list installation requirements or which models Underdog supports. XDA Developers' framing of "managing your local AI models" is therefore a reading of the announcement, not something Nvidia stated.
The privacy claim, with caveats
Local inference is a real distinction from cloud assistants. It means no prompt round-trips and no dependence on a connection. But "data never leaves the PC" is Nvidia's product claim, and nobody has independently audited it. The announcement doesn't cover several things:
- Telemetry or update pings
- Optional network features
- How data is stored on disk
- What permissions an agent gets
- Where the models come from
IT admins should read it as a design intent to verify, not a compliance guarantee. Agents that can touch files and applications deserve the same scrutiny as any other software. Remio's analysis of the platform notes that Nvidia's OpenShell integration with Windows security can enforce policies and mask personally identifiable information before selected requests reach cloud services. In other words, the wider RTX Spark story includes hybrid workflows, not only fully offline ones.
The hardware context
Nvidia says RTX Spark, which launched last week, powers personal agents on Windows PCs. DGX Spark and DGX Station are the Linux-based systems. Laptops from ASUS, Dell, HP, Lenovo, Microsoft and MSI are available now, with Acer coming soon. The platform was released on October 16.
The headline numbers are up to 128GB of unified memory and 1 petaflop of AI compute. These are platform maximums, not what every configuration ships with. Forkast, covering the launch, put the entry price at $2,599 across major OEMs. Engadget notes that a slightly less powerful configuration with 18 CPU cores and 5,120 GPU cores also exists.
Nvidia also says many RTX Spark laptops ship with an unlocked bootloader, to help developers and enthusiasts customize local agents. Nvidia doesn't say Underdog needs it.
Three separate announcements, easy to conflate
The coverage tends to blur these together, so here they are apart.
| Item | What Nvidia says | Boundary |
|---|---|---|
| Underdog | Local desktop agent app, Husky engine, available to download for RTX Spark | No compatibility list or feature list published |
| llama.cpp update | MTP speculative decoding for Qwen 3.8 Flash Next, up to 1.4x faster agentic tasks | Measured on RTX PRO 6000 Blackwell, not RTX Spark |
| Day-zero model support | Nemotron Bolt E4B, GLM, DeepSeek, Qwen, MiniMax, LFM2.5-24B-A2B, FLUX.3 Klein | Doesn't say every model runs in Underdog |
The 1.4x figure
In llama.cpp, Nvidia says new speculative decoding support with MTP for Qwen 3.8 Flash Next makes end-to-end agentic tasks up to 1.4x faster on RTX PRO 6000 Blackwell. That is a workstation GPU, not a laptop SoC. The announcement gives no baseline, methodology, or task definition, so treat "up to" as a ceiling from the vendor. Nvidia also lists separate vLLM gains: up to 1.3x on DGX Spark for Qwen3.8 27B, and up to 1.2x higher throughput for GLM 5.3 Flash on DGX Station. Those are different engines on different hardware.
What "day-zero" does and doesn't mean
Nvidia lists day-zero support for Nemotron Bolt E4B 4B, GLM, DeepSeek, Qwen, MiniMax, Liquid AI's LFM2.5-24B-A2B and the FLUX.3 Klein diffusion model. It means support is meant to be ready when the model releases. The XDA piece leaves out the Liquid AI model. Neither source says each model works across RTX Spark, GeForce RTX, DGX Spark and DGX Station, or inside Underdog.
Some model-specific details from Nvidia:
- Nemotron Bolt E4B: Nvidia says it fits natural conversation and multi-turn agent skills into as little as 4GB of memory. It releases Friday, October 23.
- FLUX.3 Klein: an open-weights, 5.6-billion-parameter image model. Nvidia says optimized builds run on GeForce RTX GPUs with as little as 8GB of VRAM. It also says they make a one-megapixel image in about four seconds on an RTX 5090.
- LFM2.5-24B-A2B: activates 2 billion of its 24 billion parameters per token. Nvidia pitches it for private, tool-calling agents on PCs.
- Quantized checkpoints: Nvidia also released RAM-efficient checkpoints. They start with Qwen3.8-Flash-Next-FP4-Local on a single DGX Spark 128GB or RTX PRO 6000, and GLM-5.3-Flash-FP4-Local on two systems.
What to make of it
For Windows buyers, the useful signal is that RTX Spark now has a first-party-backed local agent app at launch. That's a better situation than a powerful machine with no obvious starting point. Even so, "easier" is unproven. There are no independent hands-on tests of Underdog yet, and the app's feature set is thinly documented. Nvidia's own announcement is the only primary source, so its performance and privacy claims should be read as marketing until reviewers test them.
For IT pros, watch three things:
- How Underdog's permissions and network behavior look in practice.
- How it fits Microsoft's agent governance tooling on Windows.
- Whether the unlocked-bootloader approach fits your device policy.
Nvidia also says Microsoft Intune is among the integrations for managing DGX Station. That's a different product line from RTX Spark laptops, so don't assume it carries over.
Until Nvidia publishes system requirements and a feature list, think of Underdog as a promising front door to local AI on RTX Spark, not a proven model manager.