Nemotron 3.5 Bolt 74B vs Gemma 4 26B A4B: Eight Index Points and a Bit-Width Mismatch

There is already a head-to-head between N​VIDIA N​emotron 3.5 Bolt 74B-A9B 24GB and G​oogle Gemma 4 26B A4B, and N​VIDIA published it. It sits in the benchmark table on the Bolt model card, where the G​emma column is a baseline the company chose. Reading that table carefully is the fastest way to understand what these two models actually are, because the two checkpoints are not competing for the same job, and the vendor's own framing says so out loud: the comparison models were picked for their ability to fit inside 24GB of memory with 4-bit quantisation. Bolt fits inside 24GB by a different route — 2.0625 bits per weight on its expert layers — and the gap between those two strategies is the whole story.

Both numbers in the title are real. Artificial Analysis has measured Bolt at 21 on its Intelligence Index and Gemma 4 26B A4B, in its reasoning configuration, at 17. Neither figure comes from a vendor, and both were live on the panel's model pages when this was written.

The comparison N​VIDIA ran, and the column it picked

N​VIDIA's table for Bolt has four columns: Qwen3.6-35B-A3B in BF16, Gemma 4 26B-A4B in BF16, N​emotron 3.5 Lightning 30B-A3B in BF16, and Bolt in IQ2. Every number in it is vendor-reported and has not been reproduced by anyone independent. Here is the G​emma column against Bolt's, benchmark by benchmark.

• GDPVal-AA-V2 — G​emma 15.33, Bolt 28.47

• Terminal-Bench 2.1 — G​emma 37.22, Bolt 50.98

• Tau3 Banking — G​emma 14.02, Bolt 17.01

• HLE with no tools — G​emma 17.42, Bolt 25.30

• SciCode — G​emma 40.28, Bolt 41.02

• GPQA Diamond — G​emma 79.61, Bolt 86.43

• AA-LCR long-context retrieval — G​emma 57.56, Bolt 64.06

• PinchBench — G​emma 74.70, Bolt 87.82

• WebArena vision — G​emma 38.2, Bolt 50.86

• AA-Omniscience accuracy — G​emma 22.17, Bolt 16.90. G​emma wins this row, and it is the only one on the table where it does.

• AA-Omniscience non-hallucination — G​emma 14.22, Bolt 61.27. This is a large and specific loss for G​emma's family line, and it is worth pausing on: the Omniscience non-hallucination metric penalises a model that answers confidently when it does not know something, and the G​emma build recorded here scores roughly a quarter of Bolt's figure.

One caveat applies to the whole block. N​VIDIA ran all of it, including the G​emma column, on a BF16 build of G​emma against a 2-bit build of its own model, and then reported the result on its own model card.

Two checkpoints, two compression regimes

Gemma 4 26B A4B is a released product with a year of usage behind it. It landed on Hugging Face on 11 March 2026 under Apache 2.0, at 25.2 billion total parameters with 3.8 billion active per token, with a hybrid attention stack that interleaves local sliding-window layers with global ones, 256K of context, native system-prompt support and vision on every size in the family. It is a model you can pick up, quantise to 4 bits yourself and run, and a large ecosystem of ready-made quantisations already exists.

Bolt is a compression result wearing a model card. N​VIDIA took N​emotron 3.5 Super VL 120B-A12B — a multimodal checkpoint from the same generation — and ran Star Elastic over it, a router-guided architecture search that must satisfy a hard deployment budget, in this case under 24GB at batch size 1 and 128K context. Expert top-K fell from 22 in the parent to 9. The expert weights were then quantised to IQ2-XXS at 2.0625 bits per element with quantisation-aware distillation to claw back accuracy, Mamba layers were left at NVFP4, embeddings and output layers at Q8_0, and the vision tower was reattached afterwards.

The result is 74 billion parameters in a 24GB file. That is not a quantisation of a released model; it is a differently shaped network that was searched for under a memory constraint and then trained to recover.

Why the bit-width matters more than the parameters

Put the two side by side and the intuitive read is that a 74B model beats a 26B model. That read is probably correct in this case, and it is also not the interesting part, because the two models are being asked to do different things with the same 24GB.

• Parameters — Bolt 74B total, 9B active per token. Gemma 4 26B A4B: 25.2B total, 3.8B active per token.

• Architecture — Bolt: Mamba-2–Transformer hybrid LatentMoE with a C-RADIO v4-H vision encoder. G​emma: hybrid local/global attention MoE with a vision tower and, on other sizes in the family, audio.

• Stored precision — Bolt ships natively at 2.0625 bits per expert weight. G​emma's repository is 16-bit, with 4-bit and quantisation-aware-training builds published separately.

• Context — Bolt up to 256,000 tokens. Gemma 4 26B A4B: 256,000 tokens. This row is a tie, which surprises people who assume the 2-bit model gave something up here. It did not.

• Inputs — Bolt accepts text, images and video. Gemma 4 26B A4B accepts text and images; video arrives as frames, and audio is confined to the smaller family members.

• Licence — Bolt: OpenMDW-1.1. Gemma 4 26B A4B: Apache 2.0, with G​oogle's prohibited-use policy sitting alongside it.

• Speculative decoding — Bolt ships a LiLiCorr drafter, wired through DFlash with seven speculative tokens in N​VIDIA's reference recipe. Gemma 4 26B A4B has no vendor-supplied drafter; multi-token prediction appears in community fine-tunes rather than in the base release.

The practical consequence of the precision row is that Bolt's download is already the small one. There is no further 2-bit build to find, because the 2-bit build is the model. G​emma's 26B is a 16-bit checkpoint you are expected to compress, and the 4-bit version most people will actually serve is somebody's conversion of it.

Where the smaller model is the better buy

N​VIDIA's own table gives G​emma two things worth taking seriously. AA-Omniscience accuracy favours G​emma at 22.17 against 16.90, and the honest reading of that is that the 2-bit compression took a measurable toll on factual recall even after distillation. Anyone whose workload is encyclopaedic rather than agentic should weight that row heavily.

Licensing is the second. Apache 2.0 is the least encumbered option available for commercial deployment, and OpenMDW-1.1, while written to be permissive and explicitly cleared by N​VIDIA for commercial use, is a newer instrument with less literature around it and no court decisions interpreting it. Legal teams have seen one of these before and not the other.

Third is maturity. Gemma 4 26B A4B has been downloadable since March and callable from a hosted endpoint since spring. Quantised builds, fine-tunes, serving recipes and benchmark reproductions all exist. Bolt's repository went up in October. Nothing independent has been published about its latency, and there are no community quantisations to fall back on because there is nothing to quantise.

Fourth, and easy to miss: G​emma's 21B-and-under siblings carry audio natively. If the pipeline involves speech, the 26B A4B is the wrong member of its own family, and no N​emotron model is the right one.

What each one costs to actually hold

Bolt has no price, and this is not an oversight in the data. There is no hosted endpoint for it anywhere, no metered rate on Artificial Analysis's model page for it, and no provider serving it. Running it means owning the machine, and N​VIDIA's reference recipe is a weekend rather than an afternoon: a pinned vLLM nightly image, the b12x kernel package, a reasoning-parser plugin fetched as a separate file, then explicit settings for the v2 model runner, the linear and MoE backends, the attention and Mamba backends, Mamba cache dtypes, chunked prefill and a speculative-decoding config pointing at the LiLiCorr drafter.

Gemma 4 26B A4B has a price in two places. Through the catalogue we route, google/gemma-4-26b-a4b-it is listed at $0.06 per million input tokens and $0.33 per million output, at a 262,144-token context, with a p50 time to first token around 2.8 seconds — that is our own measured figure over the last seven days, not a vendor claim. Artificial Analysis tracks a different provider mix for the same model and lists $0.10 in and $0.40 out, which is the kind of spread that tells you price is a property of the endpoint and not of the weights. Both numbers are honest; they are prices of different things.

OrcaRouter puts the models we do route behind a single key at the provider's list price with 0 percent markup added, so a vendor price cut on a routed model is live on our side the same day, and automatic failover across providers means a capacity problem upstream is retried rather than surfaced to your users. We do not host any N​emotron checkpoint — the catalogue was checked rather than assumed, and the family is not in it — so on this page those guarantees cover the G​emma half of the comparison and nothing else. What they are worth in practice is a way to try the smaller model on real traffic without a second contract, while you decide whether the larger one is worth the hardware.

Choosing between a file and a service

This comparison is not close on capability and it is not supposed to be. Bolt is a bigger, more capable model that has been compressed hard and cannot be rented. Gemma 4 26B A4B is a slightly weaker model you can rent today, and whose weaknesses are documented by people who have no stake in it.

Take Bolt if the workload involves screenshots, PDFs, video frames or a browser, if it has to run without a network, and if a machine with a Blackwell-class accelerator is already in the room. Accept that you are the one who will discover its throughput, because nobody has published it. Take Gemma 4 26B A4B if you want a text-and-image model you can pay for by the million tokens this afternoon, if you need Apache 2.0, or if factual recall on a broad question set matters more than agentic task completion.

The measurement that would settle this is the one nobody has taken: an independent throughput and latency run on the 24GB checkpoint, and an independent rerun of N​VIDIA's table with G​emma quantised to the same 4 bits the vendor says it selected for. Until those exist, the case for Bolt rests on a vendor table and a parameter count, and the case for G​emma rests on a year of other people's results.

Comparisons like this one are also where the failure modes show up: the documented AI incidents and agent failures behind 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.