Skip to content
Agents · Nov 5, 2026

Open-weight models and agentic coding updates from Moonshot, Ricursive, and Cognition

Moonshot releases Kimi K3.1 Pro with open weights, Ricursive unveils Ripple-65B for agentic coding, and Cognition launches SWE-2 Flash with restricted weight access.

Trust78
HypeLow hype

1 source · cross-referenced

ShareXLinkedInEmail
TL;DR
  • Moonshot released open weights for Kimi K3.1 Pro, a post-trained model with expanded RL for coding, financial research, and long-horizon agentic work, featuring 1M context and 2.8T total parameters with 104B active.
  • Ricursive launched Ripple-65B, a 65B dense model pretrained from scratch, claiming the best US open-weight model on agentic coding with 80.5% on Terminal-Bench 2.1 and 12T training tokens.
  • Cognition introduced SWE-2 Flash, a coding model post-trained on Inkling-Small, outperforming larger American open-weight models while costing 20% of SWE-2 and running 3x faster, with weights restricted due to cyber-misuse risk.

Moonshot released open weights for Kimi K3.1 Pro, a post-training update to its K3 model family. The release expands reinforcement learning applications in coding, financial research, and long-horizon agentic workflows. The model supports text and image inputs with a 1 million token context window. Specifications include 2.8 trillion total parameters with 104 billion active parameters. Independent evaluations on Agent Arena rank Kimi K3.1 Pro first among open-weight models and 13th overall, with a reported net improvement of +5.1% across more than 17,500 agentic sessions. This represents a +4.2 percentage point gain over its predecessor, K3. Day-zero support for inference frameworks such as vLLM and SGLang has been confirmed across AMD, NVIDIA, Rubin, and TPU hardware, with additional support for Trainium 2/3 in SGLang. Hosted access is available via Ollama Cloud, Databricks, and SkyPilot, with Ollama Cloud offering zero data retention options.

Ricursive introduced Ripple-65B, a 65 billion parameter dense model pretrained from scratch. Ricursive claims it is the best US open-weight model for agentic coding, surpassing larger mixture-of-experts models including Muse Spark Open-816B, Beam-501B, Inkling-975B, and Nemotron 3 Ultra-550B. Reported performance on Terminal-Bench 2.1 reaches 80.5%, compared to Beam-501B’s 80.1%, with Beam activating 23 billion parameters per token. The model was trained on 12 trillion tokens with a 1 million token context window and released under the Apache 2.0 license. Ricursive cites a $32 million budget covering research, scaling infrastructure, data curation, and final training runs. Prime Intellect, which operated the pretraining clusters, reports 16 times pretraining compute efficiency. The architecture incorporates value residuals and sliding window attention, and internal ablations indicate approximately a 3 times gain from a new architecture on the same data as Olmo3 and a 4 times gain from a new dataset on the same architecture. Skepticism has been noted regarding comparisons with larger proprietary models such as the 284 billion parameter V4 Flash, and some critics argue that the "US-only SOTA" framing omits performance gaps with leading Chinese models or omits comparisons with Qwen models.

Cognition launched SWE-2 Flash, a coding-focused model post-trained on Thinking Machines' Inkling-Small. The company states it outperforms all American open-weight models, including those three times larger, while reducing costs to 20% of SWE-2 and improving speed by three times. SWE-2 Flash is Cognition’s first model to include a mid-training stage and was developed in a six-week cycle. Weights are not publicly available; instead, Cognition plans a secure weight-sharing program for partners and researchers, citing cyber-misuse risks as the rationale for restricted access. The model is offered at no additional cost through November for users of Devin Pro, Max, and Teams tiers.

Sources
  1. 01Latent Space — swyx — [AINews] It's Swapping Season
Also on Agents

Stories may contain errors. Dispatch is assembled with AI assistance and curated by human editors; despite the trust-score filter, mistakes happen. We correct publicly — every article links to its revision history. Nothing here is financial, legal, or medical advice. Verify before relying on any claim.

© 2026 Dispatch. No ads. No sponsorships. No paid placement. Reader-supported via Ko-fi.

Built by a person who cares about honest AI news.