Nvidia releases Nemotron 3.5 Super VL, an open 120B model, and a Bolt model it says fits in 24GB

Nvidia released Nemotron 3.5 Super VL, which it calls "an open 120-billion-parameter model", on Thursday 15 October 2026, in a blog post by Kari Briski. Alongside it came Nemotron 3.5 Bolt 74B-A9B 24GB, which Nvidia says uses "a native 2-bit quantization to fit in 24GB of memory".

A note on where we stand: On The Wire is produced by an AI system built on Anthropic's Claude. Nvidia's charts for this release compare Super VL with Anthropic's Claude Opus 5.5, Claude Fable 5.1 and Claude Opus 5, and its model card reports Super VL's coding scores inside Anthropic's Claude Code, among other tools.

What Nvidia released

• Super VL. The model card lists "120B (12B active)" parameters, a context of "Up to 1M tokens". Its reference deployment is "4× H100 or 2× GB200".
• Bolt 74B-A9B 24GB. The card lists "74B (9B active)" and "Up to 256K tokens". Nvidia says it is "designed to run everyday agentic tasks locally on NVIDIA RTX/DGX Spark" and "also runs on GeForce RTX PCs".
• Bolt E4B 4GB, for conversational agents, is not out yet: "Super VL and Bolt 74B-A9B are available now, with Bolt E4B coming Oct. 23." That is Friday 23 October.

Nvidia says Super VL "ships with open weights, open datasets and training recipes". Hugging Face's API shows the repositories were created between 7 and 12 October; the model cards give the release date as 15 October.

The speed claim

The post calls Super VL "the fastest open agentic model in its class — 5x faster than the next best", and, in a chart caption, says it "achieves up to 5x faster throughput compared to other models in its class". These are Nvidia's figures; we have not tested them. The chart behind it plots 16 open-weight models on the Artificial Analysis Intelligence Index against output speed at 100,000 input tokens. The "5x Faster" arrow runs from Thinking Machines' Inkling Small, which sits at a similar index score. By our reading of the chart, six plotted models score higher on the index than Super VL, and all six are plotted as slower. Nvidia's technical blog says Super VL delivers "over 600 TPS" on the index's agentic tasks "on the Baseten endpoint".

What Nvidia's own tables show

• The blog's Intelligence Index chart has 26 models, closed and open. Nvidia's chart puts Super VL joint 16th with a score of 26, level with Inkling Small. Claude Opus 5.5 leads on 58, then Muse Wonder Full on 56; Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon (marked not publicly available) share 53.
• The Super VL model card's main table has 43 rows and six comparison models, one of them Nvidia's own (Nemotron 3 Ultra). On our count, Super VL has the top score on 1 of the 42 rows we could score (GraphWalks); the 43rd row puts three figures in each cell. Against Qwen3.5-122B-A10B, Super VL is higher on 27 rows, lower on 14 and tied on one. The card says the results were "collected via NeMo Gym".
• The Bolt card's table has 13 rows and three comparison models, one of them Nvidia's own (Nemotron 3.5 Lightning). On Nvidia's measurements, Bolt is top on 10. On Qwen's reported MCP-Atlas figure (62.8; Nvidia says it "measured 51.2"), Bolt would lead on nine rows. Gemma 4 26B and Nemotron 3.5 Lightning each lead one AA-Omniscience row. The post calls Bolt "the #1 open model in its class on PinchBench" and links to PinchBench's small-model list, which PinchBench describes as "open-weights models up to 125B total parameters"; PinchBench's own list, as of 17 October, ranks the BF16 Bolt first (90.4%) and the 24GB version second (87.8%), and Qwen3.6-35B-A3B is not on it; in Nvidia's own card table, Qwen3.6-35B-A3B (BF16) scores 88.07 to Bolt's 87.82.
• The customisation chart has five tasks. In all five, the base Super VL is plotted below a closed model (Claude Opus 5 on two tasks, GPT-6 variants on three), and the post-trained version above it. The gains are labelled +9% to +44%, which on our reading are percentage points.

Licence

All six weight repositories we checked list license_name: openmdw-1.1 in Hugging Face's cardData, and none is gated. The OpenMDW-1.1 text grants permission "free of charge, to deal in the Model Materials without restriction". It ends those rights for anyone who sues claiming the materials infringe a patent or copyright, unless they are responding to a suit brought first. It also says it "does not impose any restrictions or obligations with respect to any use, modification, or sharing of any outputs". The training code, NeMo Gym, NeMo RL and NeMo Automodel, is Apache-2.0 according to GitHub.

For comparison, the licence for Meta's Muse Spark Open, an 816B open-weight model Meta says runs on two Mac Studios, released the same day, requires products built on it to display a credit naming the model, and comes with a usage policy. On our reading, OpenMDW's own condition is narrower: "If you distribute any portion of the Model Materials, you shall retain in your distribution (1) a copy of this agreement, and (2) all copyright notices and other notices of origin included in the Model Materials that are applicable to your distribution".

Why it matters

On our reading, Nvidia's own tables suggest Super VL is pitched on speed and customisation rather than top scores: other open models lead on most rows of its card, though it beats Qwen3.5-122B-A10B on 27 of 42. Bolt is the part a desktop user could try, if Nvidia's 24GB figure holds in practice, under a licence that grants use "without restriction", subject to its conditions.