Nemotron 3.5 Bolt vs Kimi K3: a 24GB Download Against a 2.8-Trillion-Parameter Reader

The honest short answer is that these two models are not competing for the same job, and the reason is measurable in one number. Kimi K3 holds 1,048,576 tokens of context. NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB holds 262,144. If your task is "read this entire corpus and answer in one pass", the 24GB checkpoint is disqualified before anything else about it matters — and if your task is "run on a workstation with no network, no meter and no vendor in the loop", Kimi K3 is not in the conversation either. Kimi K3 is a 2.8-trillion-parameter mixture-of-experts model published by Moonshot AI on 15 July 2026 under a proprietary hosted API. Nemotron 3.5 Bolt 74B-A9B 24GB is an open-weight file that spends 9 billion of its 74 billion parameters on any given token and fits in 24GB of memory.

The Bolt checkpoint went up on Hugging Face on 7 October 2026 with no announcement attached to it. NVIDIA's own post naming the model — and its 4GB sibling, Nemotron 3.5 Bolt E4B — arrived on 15 October, eight days later. That ordering is worth knowing because it determines what is checkable: eight days of the public record consist of a downloadable file and a model card, and the prose that explains it came afterwards.

What each model is built to do

Kimi K3 is Moonshot's flagship and its most capable release to date: a 2.8-trillion-parameter sparse mixture-of-experts with roughly 104 billion parameters active per token, built for long-horizon coding and end-to-end knowledge work, accepting text and images and returning text. It is served, not shipped. There is no download, no quantised build to run locally, and no version of it small enough for a laptop.

Nemotron 3.5 Bolt is the opposite construction. NVIDIA took its 120B-A12B Nemotron 3.5 Super VL model, applied a router-guided compression it calls Star Elastic to reach 74B-A9B, and then quantised the result to 2.0625 bits per expert element — an IQ2_XXS format — with quantization-aware distillation to recover the accuracy that pricing costs. The backbone is a Mamba-2 and Transformer hybrid the vendor calls LatentMoE, with a C-RADIO v4-H vision encoder for image and video input. It runs under vLLM on Ampere, Hopper and Blackwell parts, and on RTX Spark and DGX Spark hardware through llama.cpp, under the OpenMDW-1.1 licence that permits commercial use.

Reading a long corpus, priced

This is where the comparison becomes arithmetic rather than preference. Kimi K3 is routed on OrcaRouter at Moonshot's list price with nothing added: $3.00 per million input tokens, $15.00 per million output tokens, and cached reads at $0.30. Its p50 time to first token on our catalogue is 7.02 seconds with a p95 of 10.00 seconds, and it carried 990.6 million tokens of routed traffic across our catalogue in the last seven days — long-context reading at scale is a real workload rather than a thought experiment.

Nemotron 3.5 Bolt has no price, because there is nothing to price. There is no hosted endpoint for it, no vendor API, no rate card, and Artificial Analysis lists both its speed and its cost per Intelligence Index task as N/A for exactly that reason. What it has instead is a memory ceiling you buy once, and a marginal cost that rounds to the power bill. For a pipeline that reads a fixed corpus every night, the difference between $3 per million input tokens and zero is not a discount — it is a different accounting model, and the break-even point is a capital-expenditure question your finance team answers, not this article.

Vision is not the tiebreaker people assume

Both models take images. That is where the similarity ends.

• Input — Kimi K3 accepts text and images; Nemotron 3.5 Bolt accepts text, images and video. For an agent watching a screen recording or sampling frames from a camera feed, only one of these can be handed the file directly.

• Precision of the vision path — Bolt carries a purpose-built C-RADIO v4-H vision encoder and inherits its multimodal behaviour from the Super VL client it was compressed from; Kimi K3's visual understanding is documented as native but is not the part of the model Moonshot leads with. NVIDIA's own table lists a WebArena vision score of 50.86 for Bolt; treat it as vendor-reported.

• Reasoning control — Kimi K3 exposes no temperature, no top_p and no seed, with reasoning depth set by a reasoning-effort parameter alone; Bolt's reasoning mode is a chat-template switch, enable_thinking true or false, inherited from the same Super VL client.

• Speculation — Bolt ships with a LiLiCorr drafter driven through DFlash that proposes seven tokens ahead per step, which is the mechanism behind NVIDIA's claim of 84 tokens per second on AIPerf SPEED-Bench. There is no equivalent lever on a hosted endpoint; you get whatever the operator provisions.

The measured gap, and what it is made of

Artificial Analysis puts Kimi K3 at 43.6 on its Intelligence Index and Nemotron 3.5 Bolt at 21. There is no way to read that as close. Kimi K3's supporting rows are strong in the places a long-context reader needs them — an AA Coding Index of 76.2 at ninth of 138 models measured, a GPQA Diamond score of 93.5, and a long-context recall figure of 88.7, which is the number that tells you the million-token window is usable rather than merely declared.

Bolt's 21 is a different kind of result. It is the fourth-highest score among 67 models in its footprint-filtered class, against a class median of 8, and Artificial Analysis describes it as among the leading open-weight models of similar size. The right way to hold both numbers is that one measures what a frontier lab can do without a memory constraint and the other measures what 24GB can hold. Twenty-two index points is the price of the constraint, not evidence that the smaller model was built badly.

What is missing on Bolt's side is any independent measurement of how fast it is. NVIDIA reports at least 80 tokens per second of decoding as the design target and 84 on its own bench, with 83.99 tokens per second in a llama.cpp run on a Surface Laptop Ultra with 128GB. As of writing, no third party has published a throughput or latency figure for the 24GB weights. Kimi K3's latency, by contrast, is measured and public — 7.02 seconds p50 to first token, 10.00 seconds p95 — and those are the numbers to plan against, not the vendor's.

Where a routing layer changes the answer

Kimi K3 is on our catalogue as kimi/kimi-k3 at Moonshot's list price with 0% markup added, sitting behind the same key as 200-plus other models, with automatic failover across upstream providers when one degrades. Nemotron 3.5 Bolt is not on our catalogue — we do not host any NVIDIA checkpoint and there is no Bolt endpoint to call — so the routing question here is not which of the two we serve. It is how the two halves of a real pipeline talk to each other.

The pattern that works is a local reader and a hosted reasoner. Bolt sits on your hardware and handles the high-frequency passes where the data must not leave the building and the volume would make per-token billing painful. Kimi K3 handles the requests that need a million-token window or frontier-scale reasoning, at $3 and $15 per million tokens. Put a routing rule in front of both and the split becomes a line of configuration rather than a refactor: a request that exceeds the local context budget goes to Kimi K3, everything else stays on the workstation. Model fusion is the variant worth knowing about if the expensive calls are the important ones — several models answer the same prompt together and you take the agreement.

What to watch

The specific evidence that would settle this pairing is narrow and imminent. A first independent throughput and latency run on the 24GB Bolt weights would either confirm NVIDIA's 84 tokens per second or shrink it; a published hosted price for Bolt would turn the cost half of this comparison into arithmetic; and a second long-context recall measurement for Kimi K3 would confirm that its 88.7 is a property of the model rather than a property of one harness.

Until then, the summary is narrower than either vendor's framing. Kimi K3 is the better model on every axis that a benchmark can measure, and it is the only one of the two you can call this afternoon. Nemotron 3.5 Bolt is the only one of the two that keeps working when the network does not, and the only one whose marginal cost per token is zero. Pick the second one for the work that never should have been sent anywhere, and the first for the work that needs reach.