# Kimi K3.1 Pro weights, Ripple-65B, FlashProve, NeuralGCM+ > Moonshot opens the K3.1 Pro weights, Ricursive's dense 65B, Anthropic's provable inference system, and Ai2's long-range climate model. Published 4 Nov 2026. Coverage: 3 Nov 2026 to 4 Nov 2026. By Ali Asif. ## Items 1. **Kimi K3.1 Pro** (Moonshot AI, Open weights): The weights promised last month are now on Hugging Face: a 2.8T mixture-of-experts with 104B active, a 1M-token context, native image input and MXFP4 quantization-aware training. [Primary source](https://huggingface.co/moonshotai/Kimi-K3.1-Pro) 2. **Ripple-65B Preview** (Ricursive Intelligence, Open weights): A dense 65B trained on 12T tokens with 1M context; the lab reports 80.52 on TerminalBench 2.1 and a $32M total R&D cost. [Primary source](https://www.ricursive.com/ripple) 3. **Progress on provable inference for large language models** (Anthropic, Frontier): FlashProve produces zkSNARK proofs that a response came from a committed model, about 110 tokens per second on one H100, with code and Lean proofs released under Apache 2.0. [Primary source](https://www.anthropic.com/research/provable-inference) 4. **Ai2 and Google collaborate on a new generation of climate models** (Ai2, Science): A three-year Google.org grant funds NeuralGCM+, an open-source model that couples atmosphere and ocean for projections over years to decades. [Primary source](https://allenai.org/blog/google-climate-partnership) Four primary pages today. Two open-weight models, one piece of verification research and one climate project. I read each page and ran none of the models. ## Kimi K3.1 Pro weights Moonshot said on Oct 23 that the weights would land today, and they have. The Hugging Face card describes Kimi K3.1 Pro as an open-weight, natively multimodal agentic model, post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of domains. It is a mixture-of-experts with 2.8T total and 104B active parameters, 896 experts with 16 selected per token, and a context length of 1,048,576 tokens. Weights are MXFP4 with MXFP8 activations, trained with quantization awareness. The architectural change from K3 is in attention: the full-attention layers are replaced with Gated DeepSeek Sparse Attention, with the rest of the stack using KDA layers, for cheaper long-context work. The benchmark table is the vendor's own and it is mixed. Kimi leads the listed models on SWE Marathon v1.1 (51.7) and BrowseComp (92.1), and trails on Program Bench (31.5 against 65.0 for Claude Opus 5.5) and HLE-Diamond (39.4). That is a fair picture of an open model sitting next to the closed frontier. ## Ripple-65B Preview Ricursive Intelligence released its first language model, a dense 65B pretrained on 12 trillion tokens with up to 1M context. The lab reports 80.52 on TerminalBench 2.1, ahead of Reflection Beam 501B at 80.1 and Meta Muse Spark Open 816B at 76.4, and says the whole R&D bill was $32M. The design choices are the interesting part. It is a thin-and-deep dense Transformer, 80 layers with a d_model of 5120, chosen to keep training simple: 40% MFU with simple sharding, no custom kernels and no quantization. The post credits the final recipe to a series of small-scale ladder experiments, and reports 1.4M RL rollouts to reach its score. Weights are on Hugging Face, Together AI, GMI Cloud and OpenRouter. ## FlashProve Anthropic's Frontier Red Team published progress on provable inference: a way for a provider to commit to a model and prove each response came from it. The motivating cases are an evaluator checking that the model they tested is the one being served, and a customer behind a routing service. FlashProve generates zkSNARK proofs at around 110 tokens per second on a single H100 for Llama 4 Maverick, with proofs near 19 MB per 1,024-token request. The post puts that at roughly 70 to 95 times the GPU time of plain vLLM, so this is research, not a product. Code, a Lean verifier and the Lean proofs are released under Apache 2.0. ## NeuralGCM+ Ai2 is building NeuralGCM+, an open-source climate model that extends Google Research's NeuralGCM to simulate atmosphere and ocean together, funded by a three-year Google.org grant. The target is climate trends across years to decades, such as water availability in the southwestern United States by mid-century. ## What I'd watch next Independent runs of Ripple on TerminalBench, since a preview score from the lab itself is a claim until someone reproduces it. Third-party serving numbers for Kimi K3.1 Pro at MXFP4. And whether anyone besides Anthropic builds on the FlashProve code. ## My take Ripple is the release I would try first, because a dense 65B that runs on a DGX Spark is a model I can actually host. Kimi's numbers are mixed against the closed frontier, which is what I would expect from an open 3T-class model. I read all four pages and ran nothing, so every figure is the publisher's own. ## Sources - https://huggingface.co/moonshotai/Kimi-K3.1-Pro - https://www.ricursive.com/ripple - https://www.anthropic.com/research/provable-inference - https://allenai.org/blog/google-climate-partnership Canonical: https://heyaliasif.com/ai/2026-11-04-kimi-k3-1-weights-ripple-65b-flashprove-neuralgcm