Good morning. The open-weight pile got heavier overnight: Moonshot dropped Kimi K3.1 Pro at 2.8T parameters, Tencent released Hy4, and a US lab called Ricursive put out Ripple-65B with a carefully-hedged “SOTA” claim. Meanwhile Meta and OpenAI each quietly absorbed a coding infrastructure startup, and Anthropic’s provable-inference research — which we flagged yesterday — has drawn an attribution complaint from a prior researcher.
Moonshot releases Kimi K3.1 Pro weights. Moonshot AI published the weights for a 2.8T-parameter MoE (104B activated) with a 1M token context, native multimodality, and MXFP4/MXFP8 quantization. It’s already live on OpenRouter and Together, with Modal reportedly offering double the throughput of Moonshot’s own endpoint at the same price. The HN thread is split between people calling it benchmaxxed — it reportedly underperforms on internal evals where Opus 5.5 does well — and testers who found it genuinely strong for most use cases. The 1,453GB download settles the question of local inference for anyone without an 8x B300 node lying around.
Tencent’s Hy4 lands in the middle of the pack, cheaply. Hy4 is positioned between GPT-6.1 Sol and Terra on cost-accuracy, with specific investment in game development through Tencent’s IEG division (feature implementation, gameplay logic, asset pipelines). Tencent is also open-sourcing two co-developed scientific models: HyCell for single-cell representation and HyPDE for physical field modeling. Plotly’s benchmarkers found it thorough but slow, spending ~26 seconds per question. One HN commenter noted it’s slightly worse than Kimi K3.1 Pro but considerably cheaper, and like Kimi, it shows degraded cybersecurity capabilities.
Ripple-65B Preview’s SOTA claim gets picked apart. Ricursive Intelligence released Ripple-65B, a dense 65B model trained on 12T tokens for $32M, claiming “U.S. open-weight SOTA on agentic coding” with 80.52 on TerminalBench 2.1. The HN response was blunt: Qwen 3.8 Flash Next scores 84.3 on the same benchmark, the parameter chart compares dense-to-MoE in a misleading way, and the “U.S. open-weight” category-narrowing drew mockery (“state of the art for my neighborhood”). Commenters did concede legitimate demand for US-jurisdiction-only deployments, and the dense architecture is explicitly optimized for NVIDIA hardware.
Meta buys Conductor; OpenAI buys Pierre. Meta acquired Conductor, a multi-agent coding workspace tool launched in July 2025, folding it into Superintelligence Labs. Conductor Cloud goes read-only November 13 and shuts down shortly after, prompting visible frustration from users who’d just committed to the platform — one team requested a refund the day of the announcement and migrated to T3.codes. Separately, OpenAI acquired Pierre Computer, the Git infrastructure company from Bootstrap creators Jacob Thornton and Mark Otto, whose Code.Storage already underpins Lovable, Bolt, and Amp. The deal is almost certainly tied to OpenAI’s published struggles scaling Git — as one commenter put it, cheaper to buy the company than keep paying the subscription.
Anthropic’s FlashProve draws an attribution complaint. Anthropic published FlashProve, a zkSNARK system that cryptographically proves which model generated a response, hitting ~110 tokens/sec on Llama 4 Maverick (400B) at 70-95x overhead. On HN, James Petrie claimed the work builds heavily on his VerInf project from a MARS fellowship “with very little attribution” — down to the specific choice of Maverick as the test model. Separately, several commenters questioned who actually needs this: for open-weight models you can just redo the deterministic computation, and for closed models you’re trusting the provider’s attestation anyway.
RTX Spark laptops throttle hard on Linux. A MindStudio writeup shows Nvidia RTX Spark laptops (MSI Prestige, Asus ProArt, Dell XPS) drop GPU clocks from 2.2 GHz to ~780 MHz on Linux because the power management firmware is encrypted and only the Windows driver holds the key. Inference falls from ~65 to ~14 tokens/second. WSL is the only practical workaround for now. HN commenters were quick to note the article is about the RTX Spark laptop platform, not the DGX Spark workstation (which ships Linux by default) — though several used it as a reminder that Nvidia’s Linux posture hasn’t meaningfully changed.
Qualcomm’s SRAM-heavy accelerator resurfaces. The Cloud AI 100 Ultra — 150W, 128GB at 548GB/s, using SRAM instead of HBM — is being discussed on HN, though the card actually launched in 2024. It’s available at Cirrascale for $5.15/GPU/hr or ~$3,759/month on annual terms. The design is unusual (half a terabyte of on-package SRAM plus regular RAM replacing HBM), but commenters felt the bandwidth-to-TOPS ratio feels dated versus current alternatives, and nobody seems sure how it holds up under sustained inference.
AI PACs had a mixed midterm. AI-focused political groups spent millions on the 2026 midterms with uneven returns. Greg Brockman and a16z-backed groups put $4.8M behind a South Carolina Senate candidate who lost in the broader Democratic wave; Anthropic-aligned Public First had better luck with Democrats. Leading the Future, the pro-industry group, has concluded primaries (where it spent $27M+) are a more efficient use of money than general elections — worth watching as 2028 cycles begin.
That’s four model releases, two acquihires, and one attribution dispute before breakfast. See you tomorrow.