CNopen @CNopenmodels

Baseten's Base Labs released elephant-1, a 9B model built on Qwen3.5-9B that never writes text. Give it a ticket and yes/no questions; it returns a probability for each option in one pass. MIT weights, one CUDA GPU, tying d1-ultra on Artificial Analysis's Decision Index.

How to run Qwen-Flash-Next Coder on 8 GB VRAM: 1. Use the 3.5 BPW quant: 8 GB VRAM, 24 GB RAM, 262K context. 2. Keep most experts in RAM, smallest in VRAM, n-gram on your SSD. 3. Offload with a selective MoE regex, not --n-cpu-moe. Expect ~70 T/s prompt, 8 T/s generation.

Moonshot AI's Kimi K3.1 Pro is live on OpenRouter, with a 1M-token context window. The open-weight successor to Kimi K3 handles long-horizon coding and multimodal input, at $2 per million input tokens and $10 per million output. Whole repos fit in one prompt, without chunking.

@tinkerlog Super nice !

Alibaba's Amap open-sourced Lambda-0, a humanoid model trained only on 500 hours of first-person human video. Wearable cameras record tasks instead of teleoperating robots, and it runs zero-shot on a Unitree G1. Apache 2.0 code, 62.5% success on real mobile manipulation tasks.

Tencent's Hy-Nomas generates a video world one segment at a time, and you keep steering it: add objects, change scenes, or move the camera between segments. Four denoising steps per segment at 704×1280. Apache-2.0 license, and the BF16 weights load straight into Diffusers.

DeepSeek and MiniMax have completed China's AI export registration, moving ahead with overseas licensing. Open weights and API access would stay outside export controls, the ministry says. But running MiniMax models locally in the US, EU, UK or South Korea needs a license.

Kimi K3.1 Pro is a post-trained successor to Kimi K3, so this update is mainly a post-training story rather than a new base release. • 1,048,576-token context window for long-horizon sessions • Open-weight multimodal reasoning, built for long-horizon coding • $2.00 per 1M input tokens and $10.00 per 1M output, so output costs 5x input and long reasoning traces will dominate the bill Over the next quarter, watch whether 1M-token open-weight models become the default for repo-scale agents, and whether output-heavy pricing pushes teams to route long reasoning chains across cheaper hosts on OpenRouter.

@tinkerlog Let’s gooooooo this is is INCREDIBLE

It’s here!!!

Kimi.ai @Kimi_Moonshot

a day ago

Meet Kimi K3.1 Pro, our most capable 3T-class model.

464 2K 30K 1.9M 3K

@tinkerlog Cool!

1 microVM per agent, with no credentials inside it. Google Cloud's Agent Compute Engine gives every AI agent its own hardware-isolated environment: • Each agent runs in a dedicated microVM • No credentials are stored inside the VM • It already runs Gemini agents • Enters preview in the coming months For production builders, this moves agent isolation from something each team builds into the compute layer itself. Per-agent hardware boundaries also mean one misbehaving agent doesn't share a host with the others.

@tinkerlog Nice !!

Chinese open models by job this week: • Coding → Qwen3.8-Flash-Next • Speech → Ming-3.0-TTS-0.5B • Audio → Ming-3.0-audio-tiny • Agents → Sage-A1.5 • Agents → Agnes-3.0-Qwen • Robotics → SenseNova-RoboRSI Sage-A1.5 is built on Qwen 3.6.

Run Qwen3.8-Flash-Next at 4-bit on one 24GB RTX 4090: 1. Get Unsloth UD-IQ4_XS from Hugging Face, not IQ2_XS. 2. Load experts into 64GB DDR5-5800 RAM, up from 32GB. 3. Keep files off the SATA SSD (~550 MB/s), since disk reads hurt.

@tinkerlog Super good thanks !

Five open-source AI repos with serious momentum right now, worth saving before your next build: 1. thesysdev/open-intelligent-ui → An open-source take on ChatGPT's Intelligent UI, streaming answers as interactive maps, charts, photos and working tools instead of plain text. Built with OpenUI. Gaining about 52 stars/day. 2. kuafuai/andee → Install one APK and it operates any app on your phone: reads the screen, taps buttons, types. No integrations, no root, runs on-device with any model you swap in. Replaces brittle per-app automation hacks. Up 68 stars/day. 3. datalab-to/indra → Fast CPU OCR covering line detection, CTC recognition, reading order and markdown output. Replaces per-page cloud OCR bills for batch document work. Up 101 stars/day, one of the fastest climbers here. 4. yetone/magpie → One place for every agent's model. Run Codex on DeepSeek or Claude Code on Kimi without juggling separate configs. Up 669 stars/day, the biggest velocity on this list. 5. yi1108/printfilm → 6,098 stars and climbing at 143/day. Read the README first to confirm it fits your stack before you commit. Bookmark this stack for your next build.

MiniMax just put M3.1 Flash live on OpenRouter. 1M context. $0.15/1M input, $0.60/1M output. Multimodal, with significant gains over M3 in agentic work, coding and cybersecurity. Live on the OpenRouter API now, so 1M-token tests need no hosting.

MiniMax M3.1 Flash can now process a 1,000,000-token context in a single multimodal call on OpenRouter, at $0.15 per 1M input tokens. → Long inputs like big codebases or multi-step agent traces fit in one request, so you skip the manual chunking and stitching step → Input runs $0.15 per 1M tokens and output $0.60 per 1M tokens → MiniMax reports significant gains over M3 on complex agentic tasks, coding, and cybersecurity benchmarks → Multimodal input goes through the same endpoint, so no separate vision service to route to Engineering takeaway: benchmark it as the long-context agent step in your stack before sending those calls to a pricier model.

@tinkerlog This thing is massive