Ripple 65B Preview vs Nemotron 3.5 Bolt 74B: The VRAM Bill Comes First

Both of these models are sold on the same promise — you can run a serious agentic coder on hardware you own. Ripple 65B Preview asks for about 130 GB of memory in BF16 before you have loaded a single token of context. Nemotron 3.5 Bolt 74B-A9B asks for 24 GB, because NVIDIA compressed it to 2.0625 bits. That is a five-fold difference in the hardware you have to buy, and it decides this comparison before the benchmark table gets a word in.

Which is the interesting part, because on the one benchmark both labs published at the same version, the small-footprint model is not the winner. NVIDIA's Nemotron 3.5 Bolt 74B-A9B reports 50.98 on Terminal-Bench 2.1. Ricursive's Ripple 65B Preview reports 80.5 on the same revision. Thirty points, same exam — but taken under conditions so different that the honest reading is narrower than the headline suggests.

The arithmetic, before anything else

Ripple 65B Preview is dense. All 65 billion parameters are live on every token, so at BF16 the weights alone are 65 × 2 bytes = 130 GB, and that is before KV cache. Ricursive does ship an NVFP4 variant, which brings the weights to roughly a quarter of that — call it 33 GB plus cache — and that is the build that makes a single-GPU or DGX Spark deployment realistic. The company's stated target is a DGX Spark, with 250 tokens per second on two B200s at batch size 8 before custom kernels.

Nemotron 3.5 Bolt 74B-A9B is a Mixture-of-Experts model with 9B active parameters, derived from NVIDIA's Nemotron 3.5 Super VL 120B-A12B parent by cutting expert top-K from 22 to 9 with a router-guided search the company calls Star Elastic. At 2.0625 bits in IQ2_XXS it lands at roughly 19 GB of weights and fits a 24 GB card at batch size 1. That is the whole design goal, and it is met.

• Memory — ~130 GB BF16 / ~33 GB NVFP4 vs ~19 GB at 2.06 bits, 24 GB card at batch 1

• Shape — 65B dense vs 74B total / 9B active MoE

• Context — 1M advertised vs 262,144, with vLLM recipes recommending 64K on a 28 GB card

• Modality — text only vs text, image and video input

• Licence — Apache 2.0 vs OpenMDW-1.1

• Price — $0.20 / $0.60 per million on the first hosted endpoint vs no hosted per-token price at all

Same harness, different conditions

Terminal-Bench 2.1 is the rare case where both labs quoted the same version, which makes the 80.5 against 50.98 comparison legitimate in a way that most cross-model benchmark talk is not. But legitimacy is not the same as fairness, and there are two asymmetries you have to hold in mind.

The first is quantisation. NVIDIA's 50.98 is a vendor-reported figure for a model distributed mainly as a 2-bit build; the card's headline GPQA number, 86.43, is measured on that compressed checkpoint. Ricursive explicitly states its scores were taken "at maximum effort levels in BF16" — the full-precision build. Comparing a 2-bit run to a BF16 run is comparing two different models that happen to share a repository.

The second is that NVIDIA's number is the more conservative claim. The company published a sober 50.98 on Terminal-Bench while also publishing 86.43 on GPQA Diamond, and independent evaluators have since placed the model at 21 on Artificial Analysis's Intelligence Index — scored on v4.3.2, which is a different index version from the v5 rebuild that other models in this price bracket are quoted on. Do not put that 21 next to a v5 number and subtract. Index versions are not interchangeable, and mixing them is the most common error in comparison posts written this week.

Where the 24 GB model wins outright

GPQA Diamond is the clearest case. Ripple 65B Preview reports 85.2. Nemotron 3.5 Bolt 74B-A9B reports 86.43. The model that fits on a single consumer card is ahead on graduate-level science questions by 1.2 points — a gap small enough to be noise, but not a gap in Ripple's favour. A 65B dense model trained on 12 trillion tokens starting from scratch has not opened a knowledge lead over a compressed 74B MoE distilled from a larger parent, and there is no reason it should have.

Nemotron also takes images and video. Ripple takes text. If your agent needs to look at a rendered page, a chart or a screenshot of a failing CI job, that is not a benchmark difference you can prompt your way around.

Ripple's advantages show up elsewhere on its own card: 76.5 on SWE-bench Verified, 71.67 on MCP Atlas, 79.2 on IFBench, 83.8 on FORTRESS Adversarial. Those are strong numbers for a 65B model and they point at the same thing the Terminal-Bench score does — this is a model post-trained specifically for tool use and agentic coding, at the cost of breadth.

Speed claims that cannot be compared

Ricursive's 250 tokens per second is two B200s at batch size 8. NVIDIA's 83.99 tokens per second is a laptop-class RTX Spark machine with a speculative drafter, on AIPerf's coding benchmark. Those are not the same measurement on the same hardware, and quoting them in one sentence would be dishonest. What can be said is directional: Ripple needs datacentre-class accelerators to hit its number, and Nemotron was tuned to produce a usable rate on a machine that fits under a desk.

The practical translation is throughput per dollar of hardware, and nobody has published that for either model. If you are choosing between them, it is the number you should generate yourself, because it depends on your batch size, your context length and whether you are running the NVFP4 build or the BF16 one.

Two permissive licences, one meaningful difference

Ripple is Apache 2.0, which needs no explanation and imposes no threshold. Nemotron 3.5 Bolt ships under OpenMDW-1.1, NVIDIA's own open model licence, which permits commercial use but is not the OSI-standard default that legal teams have pre-cleared. In practice both are usable; in procurement, one of them will take an extra two weeks.

Nemotron's knowledge cutoff is May 2026 and its language coverage is English and code first, with Spanish, French, German, Italian and Japanese supported. Ripple's card says its corpus is predominantly English and code, full stop. Neither is a multilingual model, and if that is your requirement, neither belongs on the shortlist.

When not to run either one yourself

The honest case against both is operational. A model on your own GPU is a model you patch, monitor, capacity-plan and keep warm — and a 24 GB card running a 2-bit MoE at batch size 1 is not a production serving tier, it is a very capable development box. The moment your traffic is bursty or multi-tenant, per-token hosting wins on total cost even at a worse rate, because you stop paying for idle VRAM.

That is the gap OrcaRouter fills for the part of your stack that does need to be hosted: one API key across 200-plus models at provider list price with 0% markup, so a vendor price cut reaches you the same day rather than at the next contract renewal, and automatic failover when one provider degrades. Neither of these two models is on our catalogue today — Nemotron 3.5 Bolt has no metered endpoint published at all, and Ripple reaches an API through Ricursive's own endpoint and third-party platforms. What the router gives you is the exit ramp: prototype locally on the weights, then move the production path to a hosted model without rewriting the client.

The rule of thumb

Buy the 24 GB model if the machine is the constraint. Nemotron 3.5 Bolt 74B-A9B is the only one of these two that runs on a single consumer card without quantising it yourself, it reads images, and it is within a point of Ripple on GPQA. You give up roughly thirty points of Terminal-Bench 2.1 and you accept a 2-bit checkpoint, which is a real cost in agentic reliability that the benchmark gap probably understates.

Buy the 130 GB model if the work is text-only agentic coding and you have the hardware or a cheap hosted endpoint. Ripple 65B Preview's 80.5 on Terminal-Bench 2.1 is the strongest agentic-coding claim either lab made this month, and its Apache 2.0 licence is the least friction you will find anywhere. Just run the NVFP4 build, budget for the memory, and reproduce the number on your own task set before you believe it — the card itself warns that replication has been difficult.