Easy AI Daily Digest | November 7, 2025: Kimi K2 Thinking, Google Ironwood, and More

Forum topic · 小凯 · 2026-03-27

Summary

Easy AI Daily for November 7, 2025 covers major AI developments: Moonshot AI released Kimi K2 Thinking, an open-weights 1T-parameter INT4 MoE reasoning model with 256K context and SOTA results on HLE and BrowseComp. Google launched the Ironwood AI chip (TPU v7), 4x faster than its predecessor, targeting models up to 100 trillion parameters. SANA-Video merged into Hugging Face Diffusers. An anonymous model, Polaris Alpha, climbed to #3 on Repo Bench, while leaked snippets suggest a "GPT-5.1 Thinking" model. In the agent ecosystem, LangChain shipped Deep Agents JS, and fastWorkflow achieved SOTA on Tau Bench. Research highlights include EleutherAI's finding that reasoning in Qwen 3 14B and Gemma 3 12B can be expressed as linear mappings, and Anthropic's postmortem on fp16/fp32 sampling bugs. Industry news: Apple may use Google's 1.2T model for Siri, Soumith Chintala announced his departure from Meta/PyTorch, and David Sacks stated there will be no federal AI bailout.

Easy AI Daily | November 7, 2025

A digest of AI industry news from the Easy AI teaching project.

Model Updates

Moonshot AI's Kimi K2 Thinking: Open-Weights 1T INT4 Reasoning MoE

Moonshot AI released Kimi K2 Thinking, an open-weights model with 1 trillion parameters using an INT4-quantized MoE architecture. It supports a 256K context window and 200–300 consecutive tool calls, achieving SOTA on HLE (44.9%) and BrowseComp (60.2%).

Links: Technical blog | Hugging Face | Kimi.com

SANA-Video Lands in Diffusers

SANA-Video has been merged into the Hugging Face Diffusers library, supporting video generation and compatible with Diffusers' schedulers and pipelines.

Link: Diffusers PR

Polaris Alpha Rockets to Repo Bench Top 3

An anonymous model, Polaris Alpha, quickly climbed to #3 on Repo Bench, fueling speculation it may be GPT-5.1 or a Gemini model. Some users found Claude 4.1 outperforming Claude 4.5 on certain tasks.

Link: Repo Bench

GPT-5 Tops Gemini 3 Pro on VoxelBench

GPT-5 beat Gemini 3 Pro (codename Lithiumflow) on VoxelBench, showing stronger 3D model generation; screenshots circulated in the community.

Link: VoxelBench results

Alleged GPT-5.1 Source Code Leak

Reddit users surfaced source snippets allegedly from OpenAI showing a model named "GPT-5.1 Thinking," unconfirmed officially.

Link: Reddit discussion

AI Hardware

Google Ironwood AI Chip Launch

Google launched the Ironwood AI chip, 4x faster than its predecessor, supporting 9,000+ TPUs per Pod and training of 100-trillion-parameter models, positioning itself against Nvidia.

Link: Google announcement

New AI Silicon and Inference Stack Updates

Google TPU v7 (Ironwood) approaching GA; Apple M-series Neural Accelerators supported in llama.cpp; Together's ATLAS adaptive speculative decoding delivers 4x speedups.

Links: TPU v7 docs | llama.cpp update | ATLAS

GPU Systems: FP4 Tricks, Real Bandwidth, and Triton Tactics

NVIDIA Blackwell supports FP4→FP16 block conversion; memory bandwidth tests reach 92% of spec; Triton dynamic compilation optimizes kernels with C++ JIT support.

Links: Blackwell PTX ISA | Triton docs

Agents and Tooling

LangChain shipped Deep Agents for JS; Privy + LangChain enable agent wallets; Perplexity Comet adds multi-tab browsing; Google Deep Research integrates Gmail/Drive. Links: Deep Agents, Privy, Comet, Deep Research
CodeClash: LLMs played 1,680 matches, but human experts won 37,500–0; Claude Sonnet 4.5 was the best model. Results
fastWorkflow achieved SOTA on Tau Bench (retail and airline), showing small models with context engineering can match larger ones. Repo
DroidRun AI: Reddit discussion on Android automation and the open-source status of Gemini 2.5 Computer Use. Discussion

Research and Benchmarks

GoodfireAI decomposed MLP weights into memorization and generalization components; Google released the DS-STAR data science agent benchmark; MIRA revealed visual reasoning flaws. Links: GoodfireAI, DS-STAR, MIRA
EleutherAI paper: reasoning in Qwen 3 14B and Gemma 3 12B can be represented as linear mappings, with low-dimensional semantic structure found via SVD. OpenReview
Anthropic postmortem: fp16 vs fp32 precision issues caused top-p/top-k sampling bugs, underscoring dtype validation. Postmortem

Community

Yannick Kilcher Discord debated slow mode (1/2/6-hour) for the ML papers channel.
Hugging Face paused some Spaces over potential new regulations, sparking debate on safety responsibility.
Tinygrad's tinybox now supports BMC remote reboot, confirmed by George Hotz.

Company and Industry News

XPeng unveiled the IRON humanoid robot with a biomimetic female-pelvic gait; practicality questioned; compared with Tesla Optimus. Reddit
Apple may use Google's 1.2T-parameter model for the new Siri, per Reuters. Reuters
OpenAI launched mid-run prompt editing, letting users interrupt long queries and add context without restarting.
Soumith Chintala announced his departure from Meta/PyTorch.
David Sacks: "There will be no federal bailout for AI"; Sam Altman clarified OpenAI is not seeking government guarantees. Sacks | Altman

---

*Source: Easy AI teaching project.*