Ripple 65B Preview vs Kimi K3.1 Pro: Two Open-Weight Releases, Zero Shared Benchmarks

On 4 November 2026, two open-weight models landed within hours of each other, and almost nobody compared them properly. Ripple 65B Preview is a 65-billion-parameter dense model from Ricursive Intelligence, pretrained from scratch on 12 trillion tokens and released under Apache 2.0. Kimi K3.1 Pro is Moonshot AI's trillion-parameter sparse flagship, whose API opened on 23 October and whose weights — about 1.56 TB of them — became public on 4 November. One is small enough to sit on a desk; the other needs a datacentre. The obvious question is which is better, and the honest answer is that no published number lets you answer it, because the two labs did not run a single benchmark at the same version.

That is not a rhetorical dodge. It is the most useful thing on this page: if you are choosing between these two, you are choosing between two different kinds of bet, and the benchmark table cannot make that choice for you.

What each model actually is

Ripple 65B Preview is dense, not sparse. Every one of its 65 billion parameters fires on every token, in an 80-layer "thin-and-deep" transformer with a 5,120-wide residual stream. Ricursive describes the design as a move away from the Olmo 3 architecture toward a Gemma 3N-like one — overprovisioned attention heads, a higher sliding-window-to-global attention ratio, gated attention, global NoPE, value residuals and FFN activation clipping. Training ran 877,500 steps at 8K sequence length on a text-only, mostly English-and-code corpus, followed by midtraining, a long-context stage and light RL focused on agentic coding. Ricursive puts the compute bill at $32M.

Kimi K3.1 Pro is the opposite shape. It is a Mixture-of-Experts model in the trillion-parameter class — third-party trackers put it around 2.8T total with roughly 104B active per token — and Moonshot's release notes describe it as a post-trained K3 with a much larger RL budget spread across hundreds of domains. The company's own launch material calls it a post-trained version of K3, but the model card tells a more specific story: the backbone moved from 24 Gated MLA layers to 24 Gated DSA layers. It ships with text, image and video input, configurable reasoning effort and a 1,048,576-token context window.

• Parameters — 65B dense (all active) vs ~2.8T total / ~104B active, per third-party trackers

• Context — 1M tokens advertised vs 1,048,576 tokens, advertised and served

• Licence — Apache 2.0 vs the Kimi K3 licence, renamed, with a $20M/12-month MaaS threshold

• Weights on disk — a single-repo BF16 checkpoint plus NVFP4 and base variants vs ~1.56 TB across 96 MXFP4 shards

• Modality — text in, text out vs text, image and video in, text out

• Price — $0.20 / $0.60 per million input/output tokens on third-party listings vs $2.00 / $10.00, which is the rate our own catalogue carries

The number that looks like a blowout isn't one

Ripple's card reports 80.5 on Terminal-Bench 2.1. Kimi's launch reports 65.7 on Terminal-Bench 4.0. Put those side by side and Ripple appears to win by nearly 15 points for a fraction of the size. That comparison is meaningless, and it is being made constantly in the replies to both launch threads.

Terminal-Bench 4.0 is a later, harder revision of the harness than 2.1. A model that scores 80 on the 2.1 task set has not been shown to be within fifteen points of a model that scores 65.7 on 4.0 — it has been shown to have sat a different exam. The same trap swallows the other comparisons floating around: Ripple's 79.1 on SWE-bench Pro v2.0 Hard, its 44.3 on DeepSWE 1.1, its 76.5 on SWE-bench Verified, against Kimi's 51.7 on SWE Marathon v1.1 and 61.2 on Kimi Code Bench 3.0 Long-horizon. Those are six benchmarks, not one of them shared.

Even a single lab measuring a single benchmark disagrees with itself across harnesses. Kimi K3.1 Pro's Terminal-Bench 4.0 number appears as 65.7 from Moonshot, 63.6 from Vals AI and 67 from Artificial Analysis — a 3.4-point spread on the same model and the same benchmark version, purely from tooling. If three runs of one benchmark on one model vary that much, a cross-version comparison between two models is not a measurement. It is decoration.

Ricursive says as much in its own card, more quietly than the launch thread did. The card notes that "replicating benchmark scores has been challenging" and that scores were taken at maximum effort in BF16. Read that sentence twice before you put 80.5 in a slide.

The axis you can actually compare

Two things are directly comparable because they are not benchmark scores: what the model costs and what it takes to run.

On price, the gap is roughly ten to one. Kimi K3.1 Pro is $2.00 per million input tokens and $10.00 per million output on our catalogue, which passes through Moonshot's list price with no markup. Ripple 65B Preview is listed at $0.20 and $0.60 on its first hosted endpoint — a figure worth treating as provisional, because Ricursive has not published a rate card of its own. That same endpoint serves a 16,384-token context against a 1M advertised window, and at least one tracker labels the model proprietary while its weights sit publicly on Hugging Face under Apache 2.0. When the trackers are that wrong about the licence, their pricing deserves the same scepticism.

The bigger cost difference is not per token. It is per answer. Kimi K3.1 Pro is a reasoning model that emits on the order of 55,000 output tokens per Artificial Analysis index task; at $10 per million output that is most of the $1.98 the index costs per task. Ripple's numbers come from a model with light post-training and no configurable reasoning ladder, so it should emit far less — but until someone runs the same task set through both, the per-task figure is an estimate, not a fact. Tokens per answer, not dollars per token, is where small models usually win and where the comparisons on the launch threads never go.

Deployment is where the two diverge completely

Kimi K3.1 Pro's download is about 1.56 TB across 96 safetensors shards in MXFP4, with QAT keeping quality at that width. You are not running it on a workstation. The realistic options are a hosted endpoint or a multi-node cluster, and the licence adds a commercial threshold that matters at scale: a $20M/12-month MaaS trigger, plus a naming clause at 100M monthly users or $20M monthly revenue.

Ripple 65B Preview is built for the other end of that spectrum. In BF16 the checkpoint is roughly 130 GB, which is why the "runs on a MacBook" line in the launch coverage is doing more work than it admits — 130 GB does not fit in a laptop's unified memory without quantisation. What Ricursive's card actually claims is an NVIDIA DGX Spark, plus 250 tokens per second on two B200s at batch size 8 before custom kernels. The repo also ships an NVFP4 variant, which is the one that makes small-machine deployment plausible.

So the deployment question answers itself before the benchmarks do. If you have a DGX Spark and no GPU budget, Ripple is a model you can own. If you have neither a spare rack nor the appetite for a 1.56 TB pull, Kimi is a model you rent.

Calling either one

This is where the two stop being symmetrical. Kimi K3.1 Pro is on OrcaRouter today at its list rate — $2.00 in, $10.00 out, 0% markup, so when Moonshot cuts its price the change reaches your key the same day rather than on the next billing cycle. One key covers that model alongside the rest of the catalogue, which matters if you are running Kimi as the expensive escalation tier behind a cheaper default.

Ripple 65B Preview is not on our catalogue, and we are not going to pretend otherwise. It reaches an API through Ricursive's own endpoint and third-party platforms, and its weights are on Hugging Face if you want them directly. A brand-new Preview release with a card that admits its own scores are hard to replicate is exactly the case where you want the model behind a router rather than wired into a production path — but that only helps once a vendor opens it to third-party routing.

Which one to pick

Pick Kimi K3.1 Pro if the work is long-horizon and open-ended: large refactors across a big repository, document analysis at 200K tokens and up, anything where image input saves you a separate OCR step, or anything where you need the model to hold a plan across hundreds of tool calls. You are paying frontier-adjacent prices and you should expect frontier-adjacent behaviour. The 3.4-point spread across harnesses is noise; the 1M context and the multimodal input are not.

Pick Ripple 65B Preview if the work is bounded and you want to stop paying per token: a coding agent on a fixed task set, a local assistant on hardware you already own, a fine-tuning target — Apache 2.0 and a 65B dense checkpoint are a far friendlier starting point than a 1.56 TB MoE. Treat the 80.5 as a claim from the lab that made the model, on a benchmark revision nobody else has published at, and validate it on your own tasks before it goes anywhere near a production decision.

And if you want the comparison that would actually settle this, it does not exist yet. Nobody has run Ripple 65B Preview and Kimi K3.1 Pro through the same harness at the same version. Until someone does, the only honest ranking is by shape: a small model you can own against a large model you can rent, with a price ratio of about ten to one and no shared exam to break the tie.