Nemotron 3.5 Bolt 74B vs Kimi K3: A 24GB File Against a 2.8-Trillion-Parameter Flagship

The two models in this comparison are not competing for the same slot, and the numbers make that plain before either vendor says a word. N​VIDIA N​emotron 3.5 Bolt 74B-A9B 24GB is a 74-billion-parameter vision-language model that N​VIDIA compressed to a 2-bit checkpoint small enough to sit in 24 gigabytes of memory alongside the operating system. Kimi K3 is M​oonshot AI's 2.8-trillion-parameter flagship, an open-weights model whose smallest viable deployment is a rack. The first is a file you can copy onto a laptop that N​VIDIA began shipping on October 15, 2026. The second is a workload you rent by the token. Reading their scores side by side is useful precisely because the gap is so large that nobody can pretend the smaller one is a substitute.

What is genuinely worth doing on this page is working out where the 22 Intelligence Index points go, and whether any of them land on the tasks a local agent actually performs.

What shipped, and when

The Bolt repository appeared on Hugging Face on 7 October 2026; N​VIDIA's own launch post naming it, "Open From Data Center to Desktop: N​VIDIA Expands N​emotron Family With Highest Efficiency Super VL and Bolt Models," is dated 15 October. The model card itself carries a release date of 10/15/2026. That ordering — weights first, announcement eight days later — matters for anyone reading coverage of this model, because a week of the public record consists of a downloadable checkpoint and a model card with no press release attached to it.

The two models are otherwise separated by everything. Kimi K3 was unveiled on 16 July 2026 with weights following on 27 July, making it roughly three months older and considerably better documented by third parties.

• Architecture — N​VIDIA N​emotron 3.5 Bolt 74B-A9B 24GB is a Mamba-2–Transformer hybrid LatentMoE with a C-RADIO v4-H vision encoder. Kimi K3 is a 2.8-trillion-parameter sparse mixture-of-experts activating 104B parameters per token across 16 of 896 experts, built on M​oonshot's K​imi Delta Attention.

• Parameters — 74B total, 9B active for the N​VIDIA model; 2.8T total, 104B active for Kimi K3. The smaller model activates less than a tenth of the larger model's per-token compute.

• Memory footprint — 24GB at 2-bit quantization (IQ2-XXS at 2.0625 bits per element for the expert weights, NVFP4 for Mamba layers, Q8_0 for embeddings and output) against whatever a 2.8T-parameter deployment costs, which is not a number this page can state in gigabytes.

• Context — 262,144 tokens for the N​VIDIA model; 1,048,576 for Kimi K3, with flat pricing regardless of length.

• Inputs — text, images and video in, text out for both.

• Licence — OpenMDW-1.1 for the N​VIDIA checkpoint; the Kimi K3 License for M​oonshot's weights, permissive for most use with a separate agreement required above $20 million in annual revenue.

The independent numbers, and the one that does not exist

Artificial Analysis has run both models, which is what makes this comparison answerable at all. Not everything on the board was measured for both, and the gaps are as informative as the scores.

• Intelligence Index — N​VIDIA N​emotron 3.5 Bolt 74B-A9B 24GB scores 21.4, fourth of 67 models in its size class on the panel's own comparison set. Kimi K3 scores 43.6 in its maximum reasoning mode and 30.1 in its low mode. Both figures are independently measured, not vendor-reported.

• Cost per Intelligence Index task — the N​VIDIA model has no figure because it has no metered price: Artificial Analysis lists no input or output rate for it at all. Kimi K3 costs $2.00 of API spend per index task at $3.00 per million input tokens and $15.00 per million output, against $0.30 for cached input.

• Output speed — Kimi K3 generates 43.8 tokens per second with 3.6 seconds to first token on the panel's rig. No independent throughput figure exists for the N​VIDIA model; N​VIDIA's own research page publishes 83.99 tokens per second on AIPerf SPEED-Bench-Coding with the LiLiCorr drafter enabled, which is a vendor measurement on a different benchmark and a different machine.

• Terminal-Bench 2.1 — 85.0 for Kimi K3, independently run, against M​oonshot's own claim of 88.3. N​VIDIA reports 50.98 for the Bolt model on its model card, so the two are not on the same footing: one number has been reproduced outside the lab and the other has not.

• GPQA Diamond — 93.5 for Kimi K3 independently; 86.43 for the N​VIDIA model as published by N​VIDIA, with no independent figure on the panel.

• Long-context retrieval (AA-LCR) — 88.7 for Kimi K3 against 69.7 for the N​VIDIA model, both measured.

• Scientific coding (SciCode) — 59.5 for Kimi K3 against 39.7 for the N​VIDIA model, both measured.

• Knowledge accuracy and hallucination — on AA-Omniscience, Kimi K3 lands at 47.6 percent accuracy with 46.8 percent non-hallucination; the N​VIDIA model lands at 17.1 percent accuracy with 56.3 percent non-hallucination. That pairing is the most interesting row on the board and the least flattering to a simple reading of the index: the compressed model knows less and admits it more often.

The missing measurement is the one a buyer would most want. Nobody outside N​VIDIA has published a throughput or latency run on the 24GB checkpoint, which is the entire point of shipping a 24GB checkpoint.

Where the 22 points actually go

A 22-point index gap sounds like a category difference, and on some tasks it is. Kimi K3 is better at scientific coding by twenty points, better at long-context retrieval by nineteen, and better at graduate-level science questions by more than seven. Those are not rounding errors and they are not artifacts of the comparison set.

But the compression story behind the N​VIDIA model is specific enough to explain the shape of the deficit. The checkpoint was produced by Star Elastic, N​VIDIA's router-guided architecture search, which took the 120B-A12B Super VL parent and searched for the expert count and MoE width per layer that would fit under 24GB at batch size one and 128K context, pulling the expert top-K down from 22 to 9 and then quantizing the experts to 2.0625 bits per element with quantization-aware distillation to recover accuracy. That is a deliberate trade of stored knowledge for footprint. A model that has been squeezed into 2 bits per expert weight keeps its procedure and loses its facts, which is why the hallucination row improves while the accuracy row collapses.

N​VIDIA's own benchmark table makes the same argument in the other direction. The company publishes the Bolt model beating Qwen3.6-35B-A3B, Gemma4 26B-A4B and its own N​emotron 3.5 Lightning 30B-A3B on Terminal-Bench 2.1, GPQA Diamond, BrowseComp and WebArena vision — that is, on agentic and tool-use tasks — while losing to Gemma4 on AA-Omniscience accuracy and to Qwen3.6 on PinchBench. Agentic execution is what the compression preserved.

Both are open weights, which makes the choice about electricity

This is the part that separates the matchup from a comparison against a closed API. Kimi K3's weights are downloadable. If you have the hardware, you can run the 2.8T model yourself and pay nothing per token — the same proposition as the 24GB file, at a completely different scale. Neither model here is gated behind a vendor that can change its mind about access.

The practical split is deployment geography. A 24GB IQ2 checkpoint runs under vLLM on a single A100, H100 or Blackwell part, and N​VIDIA's own reference recipe names llama.cpp and a GGUF conversion as well, which is how it ends up inside local agent harnesses on RTX and DGX Spark machines. The recommended serving command is not a one-line toggle: it pins a vLLM nightly build, requires a downloaded reasoning-parser plugin, needs the b12x kernels for the linear and MoE backends, sets explicit Mamba and attention backends, and enables speculative decoding with seven LiLiCorr draft tokens. That is a recipe for someone who intends to own the machine, not for someone who wants a transcript.

Kimi K3 goes the other way. Its whole design assumes you are renting inference, and its price is set accordingly — $3.00 in and $15.00 out per million tokens is a deliberate position at the top of the open-weights market.

Running the comparison instead of reading it

If you want to know whether a 22-point index gap matters for your workload, the cheap experiment is to call the expensive half. Kimi K3 is on OrcaRouter's catalogue as kimi/kimi-k3, on the same key as the rest of the 200-plus models we route, billed at M​oonshot's provider list price with no markup added — so a M​oonshot price change is live on our side the same day rather than at the next billing cycle. Point a prompt set that looks like your real traffic at it and measure the difference against whatever you are running locally. OrcaRouter does not host any N​VIDIA model and has no Bolt endpoint to offer; the N​emotron family is not in our catalogue, and the local half of this comparison stays a hardware question either way.

Where routing does earn its place here is the fallback path. Kimi K3 is a single-model commitment at a high price point, and the failure mode of a production agent wired to one high-capability model is that a capacity problem upstream becomes an incident. Automatic failover across providers behind one endpoint turns that into a retry.

How to decide

The rule is shorter than the gap suggests.

If your requirement includes the words offline, private, on-device or no per-token cost, Kimi K3 is not a candidate and the comparison is over — a 2.8-trillion-parameter model is not going to run on the machine you are reading this on, and the 24GB checkpoint will. Buy the laptop, accept 21.4 on the index, and point it at tool use rather than at questions about biochemistry.

If your requirement is correctness on hard reasoning, long-context work across hundreds of thousands of tokens, or repo-scale coding, the 24GB model is the wrong tool and the index gap is telling you the truth. Pay for Kimi K3.

What neither vendor has settled is the middle case, because it depends on a number nobody has published: how fast the 2-bit checkpoint actually runs on hardware you can buy. Until an independent throughput run exists, the honest answer to "can a 24GB file replace a 2.8T flagship for my agent loop" is that the intelligence gap is measurable and the latency case is not yet.

For a model whose failure modes are still undocumented, the useful counterpart is the record of what has already gone wrong elsewhere: the documented AI incidents and agent failures around models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.