Nvidia Shipped an Open Local Model in a Format Nobody's Local Tools Can Read
On October 15, Nvidia added two models to its Nemotron 3.5 line: Super VL, a 120-billion-parameter vision-language model built for server-scale agent deployments, and Bolt 74B-A9B, a 74-billion-parameter model with 9 billion active parameters aimed at a single workstation. Both ship as open weights, and both are pitched as agentic models that handle tool use, OCR, charts, and multi-step workflows rather than plain chat. A third variant, a 4GB edge model called Bolt E4B, is due later this week.
A checkpoint built for hardware most practitioners don't have
Bolt's headline feature is that it is supposed to fit in 24GB of memory, which is the real draw for anyone running it on a single RTX card or a DGX Spark box. But Nvidia's own upload gets there through a format tuned for its own stack: Bolt's Mamba-2–Transformer hybrid MoE layers are kept in NVFP4, most of the rest of the model sits at roughly 2.06 bits per weight, and the output head is left at Q8_0. That mixed-precision scheme is native to Nvidia's TensorRT and NIM pipeline. It is not a format llama.cpp, Ollama, or LM Studio know how to load — which happen to be the tools most individual developers actually use to run an open model on their own hardware.
Unsloth's response was a full GGUF repack, built with its own four-bit 'Dynamic' quantization scheme (UD-Q4_K_XL), published with run commands for llama.cpp, Ollama, vLLM, LM Studio, Jan, and several other local runtimes. That repack — not Nvidia's original BF16 or native-quant upload — is the version of Bolt that is actually spreading through the tools practitioners use day to day, because it's the only version most of them can run at all.
The hosted version is drawing a different crowd
Super VL's distribution problem runs the opposite direction: it's cheap enough on OpenRouter, at $0.10 per million input tokens and $0.40 per million output tokens with a 1-million-token context window, that it undercuts most comparable vision-capable agent models by a wide margin. OpenRouter's own public traffic breakdown for the model lists meaningful volume coming through client integrations tagged 'Claude Code' and 'Codex' alongside dedicated agent frameworks like Hermes Agent — which indicates that at least some developers are pointing existing coding-agent tooling, built around other vendors' models, at Nemotron instead. That's a pricing-driven substitution, not a loyalty signal, and it's worth watching whether it persists once novelty wears off.
The performance numbers are still one-sided
Nvidia's own marketing claims Super VL completes tasks '5x faster than the nearest competitor in its class' and reaches accuracy 'comparable with its largest reasoning model at about one-fifth the parameter count.' Neither claim names the competitor or publishes test conditions. The one third-party number available so far, Artificial Analysis's Intelligence Index, scores Super VL at 26.5 and Bolt at 21.4 — respectable, but solidly mid-pack among current open models rather than the field-leading result Nvidia's framing implies. Nvidia announced the release through its own channels, including a post from its AI account describing Super VL as combining 'reasoning and tool use' with multimodal understanding for long-running agents; none of the three outlets covering the launch obtained an on-record quote beyond that post.
The split is a useful reminder of how open-weight distribution actually works in practice: a lab's primary upload is tuned for its own inference stack, and the formats individual developers can run — GGUF, AWQ, exl2 — typically arrive from the community within a day or two of the original drop. On Hugging Face, that usually means the 'fastest-growing' artifact tied to a new release isn't the lab's own checkpoint but whichever third-party repack first made the model runnable on ordinary hardware. Bolt's native 2-bit NVFP4 format made that gap unusually wide, and widened the role Unsloth's repack is playing in getting it onto local machines at all.