Ripple-65B Preview vs Mistral Large 4.0: 65 Billion Dense Against a Trillion-Parameter MoE

Put Ripple-65B Preview and Mistral Large 4.0 side by side on a disk and the difference is not subtle: 129.6 GB of BF16 shards against 1,060.6 GB of FP8 shards, for two models that both call themselves open weight and both pitch agentic coding. One is a 65B dense transformer trained from scratch on 12 trillion tokens by a five-person research group that says it spent $32 million. The other is a 1.05-trillion-parameter mixture of experts with 52 billion active parameters and a 1.6-billion-parameter vision encoder, from a company with a product line behind it.

The interesting part is not the size gap. It is that both labs chose to publish a DeepSWE 1.1 number — the same benchmark, the same version — and the smaller model loses by 17.4 points on it. Ripple's own card reports 44.3. Mistral reports 61.7 for Large 4.0. So before any argument about efficiency, licensing or self-hosting, there is one clean, same-version, same-task-set comparison on the board, and it does not favour the new model.

What each lab actually shipped

Ripple-65B Preview is Ricursive Intelligence's first public model, released 4 November 2026 under Apache 2.0 with no gate and no acceptable-use rider beyond the licence itself. The architecture is deliberately thin and deep — 80 layers at a model width of 5,120 — with overprovisioned attention heads, a high ratio of sliding-window to global layers, gated attention, global NoPE and value residuals. The card describes it as relatively lightly post-trained, with 700 RL steps and a stated emphasis on agentic coding. A 4.04B speculative drafter, an NVFP4 checkpoint and an ungated base model ship alongside it.

Mistral Large 4.0 is the fourth generation of a line that has been the company's flagship since 2023, and this one arrived in two stages: an API preview on 6 October 2026, then weights on 29 October under the Mistral Ecosystem Open Weight License, a modified-MIT document with a service-provider clause. It is multimodal where Ripple is not — a 1.6B vision encoder takes images in — and it exposes three reasoning effort levels rather than one.

• Parameters — 65B dense, every parameter active per token, vs 1.05T total with 52B active and 128 routed experts, 4 per token

• Weights on disk — 129.6 GB BF16 across 16 shards vs 1,060.6 GB FP8 across 54 shards, plus a 610.1 GB NVFP4 sibling

• Context — up to 1M claimed, though the first hosted endpoint served 16,384 and a newer listing headlines 128K vs 1M across the board

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

• Licence — Apache 2.0 vs Mistral Ecosystem Open Weight License, which requires a commercial agreement to offer the model as a service

• Price where hosted — $0.20 / $0.60 per million tokens vs $0.70 / $2.10, with $0.07 cached input

The licence is the sharper difference, and it is not close

Apache 2.0 means what it has always meant. You can take Ripple-65B Preview, fine-tune it, rename it, serve it to customers, and publish nothing. There is no revenue threshold, no notification requirement, and no clause that turns a routing layer into a licensee. For a company whose product is a model behind an API, that is the whole ballgame — it is the difference between a model you can build a business on and a model you can build a feature on.

Mistral Large 4.0's licence is more permissive than its name suggests and less permissive than Apache. It is a modified MIT with an added condition: if you provide the model to third parties as a service, you need a commercial agreement. The definition carves out two things that matter — embedding the model inside an end-user product, and "mere relaying of requests to models hosted by others," which is what a gateway does when it forwards a request to Mistral's own inference. Read carefully, that carve-out keeps a router on the right side of the line. Read carelessly, or read only the marketing page that says "open weight," and you can talk yourself into a compliance problem.

So the honest framing is not "Apache versus restrictive." It is "Apache versus a licence you should have a lawyer read once." If your plan is to host the weights yourself and sell access, Ripple's terms are categorically simpler. If your plan is to call an endpoint, both licences accommodate you and the difference mostly disappears.

Terminal-Bench is the benchmark you cannot use here

Both labs publish a Terminal-Bench score, and they are not the same test. Ripple reports 80.52 on Terminal-Bench 2.1 and calls it U.S. open-weight state of the art. Mistral reports 28.3% on Terminal-Bench 4.0. Placed next to each other those two numbers look like a 52-point rout. They are not comparable, and treating them as comparable is the single most common mistake in coverage of either model.

Terminal-Bench 4.0 is a harder, later task set, and the gap between versions is large enough that a good 2.1 score and a mediocre 4.0 score can describe the same model. Our own comparison of Ripple against Claude Opus 5.5 exists mostly to explain why the 80.5 and Opus's 66.4 cannot be subtracted either. The only defensible statement here is that Ripple has not published a 4.0 number and Mistral has not published a 2.1 number, so this particular axis is empty.

What survives is DeepSWE 1.1, and there the picture is legible. Ripple's 44.3 comes from its own model card. Mistral's 61.7 comes from Mistral's own run. Neither is independent, but they are the same version of the same benchmark, which is more than Terminal-Bench can offer. A 17.4-point gap on software-engineering tasks is not noise, and it lines up with the rest of each card: Mistral Large 4.0 also reports 87.6 on GPQA Diamond, 89.4 on MMLU-Pro and 81.3 on Artificial Analysis's long-context reasoning set, against Ripple's 85.2, 84.2 and 75.0.

Ripple's counter-case rests on benchmarks where it leads its own comparison set and on one number Mistral cannot match: 80.52 against Reflection Beam 501B's 80.1 on Terminal-Bench 2.1, from a model fifteen times smaller. That is a real result about training efficiency. It is not a result about being better than Mistral Large 4.0, and the two claims keep getting merged in secondary coverage.

Cost per token flatters Ripple; cost per task is closer

At list, the hosted gap is roughly 3.5× on both directions: $0.20 and $0.60 per million tokens against $0.70 and $2.10. For a high-volume agent that difference is the entire budget conversation, and it is the reason a 65B dense model remains interesting even when it loses benchmarks.

Artificial Analysis puts Mistral Large 4.0 at an Intelligence Index of 26 and a cost of $2.06 per index task. Ripple-65B Preview has no index entry at all, so there is no equivalent figure to divide — which is itself the honest answer to "which is cheaper per unit of capability." One model has been measured on the neutral board and one has not. If you need a planning number today, only one of these columns has anything in it.

The self-hosting arithmetic changes the shape again. A 65B dense model in BF16 fits in tensor-parallel-4 across four 80GB cards, and the merged vLLM support as of 5 November means it now loads without --trust-remote-code. A 1.05T FP8 model needs an order of magnitude more hardware and a serving stack that understands expert parallelism. For a team with four GPUs and a compliance requirement, the 65B is not the compromise option — it is the only option on this list.

Where the routing decision actually lands

Neither model is in the OrcaRouter catalogue. There are no Mistral models in it at all, and Ripple-65B Preview is not there either — so any article telling you to reach both on one key, including this one, would be making it up. What is true is narrower: the open-weight models we do route are the ones a cost-sensitive buyer usually ends up comparing against these two. Kimi K3.1 Pro carries a 1M context and 943K maximum output at $2.00 and $10.00 per million. DeepSeek V4 Pro carries 1M context and 384K output at $0.66 and $1.98. Both are reachable through a single OpenAI-compatible endpoint with automatic failover, and provider list prices pass through with no markup, so a vendor price change is live on our side the same day.

That matters for this comparison in one specific way. Mistral Large 4.0 launched at $1.36 and $4.18, and the $0.70 and $2.10 that Mistral's documentation shows today is the discounted rate that followed. A price that moves by half within a month is a price worth checking rather than memorising — and it is the kind of movement a pass-through gateway reflects without you changing anything.

Who should pick which

Pick Ripple-65B Preview if you want to own the weights outright under terms you do not have to interpret, if your workload is terminal-style agentic coding rather than broad software engineering, and if 65B fits your hardware. Accept that you are trusting a vendor card that no independent lab has reproduced, that the context figure depends entirely on how you serve it, and that a 44.3 on DeepSWE is the number to beat before you displace whatever you run now.

Pick Mistral Large 4.0 if you need images as well as text, if a 1M context with a documented serving story matters more than licence simplicity, and if a measured Intelligence Index of 26 with a $2.06 cost per task is the kind of number your procurement process accepts. Accept that the licence needs one legal read, that 1,060.6 GB of weights is a different class of infrastructure commitment, and that Terminal-Bench 4.0 at 28.3% is the score your own evaluation will most likely reproduce.

The one thing I would not do is choose between them on the headline benchmarks. Two labs, two benchmark versions, two self-reported DeepSWE runs — the comparison that decides this is the one you run on your own repository, and both models are open enough to make that possible without asking anyone's permission.