Kimi K3

2026-11-01

MODELS · 1 SOURCE · The Decoder

Vals AI study: Chinese models show varied ‘pro‑China’ lean; language and prompts shift responses

Vals AI tested dozens of Chinese, US and other language models on hundreds of politically charged questions and found that alignment with official Chinese positions varies widely by model, language, prompt framing and response format rather than uniformly reflecting a government ‘mouthpiece’. Examples include GLM‑5.3 Flash acknowledging documented human rights violations, Qwen3.8 Max rendering “Taiwan” as “Taiwan, China” in Simplified Chinese, and Thomson 1.0 Small (a Thomson Reuters fine‑tuning of Qwen3.6 35B) dropping pro‑government alignment from 84.6% to 15.0% in the tests; Vals also found non‑Chinese models sometimes adopted official Chinese positions and noted that refusals were scored as neutral and some tests used third‑party endpoints.

7.0

2026-10-23

MODELS · 1 SOURCE · The Decoder

Moonshot releases Kimi K3.1 Pro, climbs to #4 on Artificial Analysis index

Moonshot AI has released Kimi K3.1 Pro, a post‑trained update to Kimi K3 that scores 55 points on the Artificial Analysis Intelligence Index (an 11‑point gain), placing it fourth and three points behind top proprietary models such as Claude Opus 5.5. The model keeps 2.8 trillion parameters (104B active), a 1M‑token context window, ships a faster ‘Highspeed’ variant, cuts API token prices by ~33%, and the company says it plans to release model weights by November 4, 2026.

7.0

2026-10-20

REGULATION · 1 SOURCE · The Decoder

DHS orders federal civilian agencies to remove and stop buying Deepseek AI

The US Department of Homeland Security has ordered federal civilian agencies to remove Deepseek AI products from their IT systems and barred new purchases, extending the restriction to affiliated firms and third‑party services described as "built with" Deepseek. The notice cites a 2018 procurement supply‑chain law but gives no deadline or specific security concerns and leaves unclear what "built with" covers (for example, whether AI gateways like OpenRouter are included).

7.0

2026-10-17

MODELS · 1 SOURCE · The Decoder

Deepseek releases Deepseek‑V4.1‑Flash‑1017, open-sources weights and opens search APIs

Chinese company Deepseek released Deepseek‑V4.1‑Flash‑1017 and published the model weights on Hugging Face and ModelScope. The company says the model—trained with reinforcement learning using self-feedback—runs faster, produces shorter outputs, beats larger rivals in its internal coding and agent benchmarks, and is paired with new Search & Fetch APIs for developers; independent verification of the benchmark claims is not yet available.

6.0

2026-10-14

MODELS · 1 SOURCE · The Decoder

Artemis Security launches Orion-1, an AI to investigate and respond to corporate cyberattacks

Artemis Security has launched Orion-1, a purpose-built AI model that ingests logs from endpoints, networks, cloud services and SaaS, reconstructs attacks, assigns confidence scores, and can recommend or initiate responses; the company says Orion-1 outperformed general-purpose models like Claude Opus 5.5 and GPT-6 Sol on its internal Decision-Grade Readiness (DGR) benchmark. Orion-1 is in private preview; Artemis says no customer data was used for training and it is unclear whether Orion-1 was built from scratch or fine-tuned from another foundation model.

7.0

2026-10-09

MODELS · 1 SOURCE · InfoQ AI, ML & Data Engineering

Google releases Android Bench 2.0 with long‑horizon tasks, agentic evaluation, and continuous scoring

Google published Android Bench 2.0, a major update to its benchmark for evaluating AI models on Android development tasks that adds long‑horizon tasks (LHTs), agentic evaluation (using providers' agents), and a continuous, multi‑factor scoring system instead of binary pass/fail. The new dashboard lists recent models (e.g., Gemini 3.8/3.7, OpenAI GPT‑6, Anthropic Fable 5.1, Kimi K3, Qwen 3.8 Max); at the time of reporting Claude Opus 5.5 led LHT pass rates at 32%, with GPT 6 Astra at 28%.

7.0

2026-10-08

COMPANIES · 1 SOURCE · TechCrunch AI

Goodfire launches internal-activation monitors on Baseten to catch rogue AI agents

Goodfire launched cheaper “inside-out” monitors that read a model’s internal activations (probes) in real time and are available to Baseten customers; in tests on Kimi K3 the probes flagged 94% of malicious hacking sessions, sent 8.7% of benign sessions for review, and monitoring ~1,500 sessions cost about $51 vs $233–$10,000 for alternative checkers. Baseten’s Base Labs had announced a safety partnership with Goodfire and Hugging Face last month, and Goodfire cites similar research from Google DeepMind informing probe-based detection in Gemini.

7.0

2026-10-06

MODELS · 3 SOURCES · TechCrunch AI · The Decoder · WIRED AI

Reflection unveils Beam, a 501B-parameter open-weight model claiming frontier performance at lower compute

Reflection AI announced Beam, a text-only, open-weight mixture-of-experts model with 501 billion total parameters (23 billion active), a 1 million token context window, and pretraining on 23.8 trillion tokens. The startup says Beam matches top Chinese open models on advanced reasoning benchmarks while using '3–4x less inference compute,' will publish weights and full technical details this month, and is targeting enterprises and sovereign 'AI factory' deployments; the performance claims have not been independently verified.

8.0

2026-09-26

MODELS · 1 SOURCE · The Decoder

OpenAI's GPT-6 Astra scores 80% on Epoch AI's furniture assembly benchmark

Epoch AI's Furniture Assembly Benchmark (FAB) tests models on identifying deliberate assembly errors in photos of IKEA furniture. Previously the best model (Claude Opus 4.5) scored 28% in November 2025; ten months later OpenAI's GPT-6 Astra reaches 80% with a processing time of about three minutes per photo, followed by Claude Fable 5.1 at 70% and Claude Opus 5 at 61%, while some Chinese open-weight models like Kimi K3 lag substantially.

8.0

2026-09-22

MODELS · 1 SOURCE · The Decoder

Xiaomi's MiMo-V2.6-Pro tops open-model rankings; Anthropic alleges Claude data siphoning

Xiaomi released the MiMo‑V2.6 lineup, with flagship MiMo‑V2.6‑Pro (1.02T params, 42B active) scoring 46 on Artificial Analysis's Intelligence Index and claiming very low token costs; the company attributes gains to a major expanded reinforcement‑learning run and published an open RL toolkit, training framework, smaller models, and ~7,000 graded training tasks. Separately, Anthropic's threat report says Xiaomi appears in case GTG‑16008, alleging ~400,000 exchanges in March–April 2026 routed MiMo conversations through OpenClaw/OpenCode to Claude to extract training data as part of a broader set of ~190M exchanges across several labs.

8.0

2026-09-18

MODELS · 1 SOURCE · AWS Machine Learning

Moonshot AI's Kimi K3 (2.8T, 1M-token) now available on Amazon Bedrock

Moonshot AI's Kimi K3 is now available on Amazon Bedrock. Per Moonshot, Kimi K3 is the company's most capable open-weight model (2.8 trillion parameters) with native vision, a 1‑million‑token context window, and ~2.5x scaling-efficiency improvement over Kimi K2; Bedrock also supports explicit prompt caching, Responses/Chat Completions APIs, regional/global inference profiles, and AWS data protections (zero data retention and zero operator access).

7.0