AI 快讯：V4.1-Flash-1017 上线，763.2B 参数面向编程 harness 调优

Kris Krakowiak @KrakowiakK原推发布于 2026.10.17

在 X 上打开原推

AI in short 🐋 Saturday - ⌨️ DeepSeek-V4.1-Flash-1017 live: 763.2B params, tuned for coding harnesses 🇨🇳 - 📦 Reflection AI puts Beam on HF: ~500B params, close to 4x bigger models - ⛔ China draft: export controls on open weights at 2T params, 10^26 FLOPs 🇨🇳 - ⊞ Microsoft MAI-Base-2: 2T params, 80B active, 50T tokens, 38 days on 16k GB200 - ♟️ RoTA-Bench: GPT-5.6-Sol solves 17% of 3D rotation tasks, undergrads 100% - 🧮 Claude cracks 25-year-old NIST ECDSA seeds: lrand48 + Unix timestamp - 🎮 llama.cpp on jailbroken PS5: Qwen 3.5 9B Q4 at 55 tok/s, Claude Code works - 🪦 Postgres 19 slips to ~Nov 19, 8-10 features pushed to v20 over AI review finds - 🖥️ OpenGCUA: code-writing computer-use agent, +19.6 pts on OSWorld-v2, 53% fewer steps 🇨🇳 - 🧪 FDE-Gym: best of 13 configs scores 0.56, fully solves 14 of 95 delivery tasks - 📱 MAI-Gym: 9B GUI agent 78.4% MobileWorld, 76.2% OSWorld-Verified, 20k parallel envs 🇨🇳 - 🎓 Waypoint-OPD distillation beats other methods in all 6 math/code settings - 🏅 MLExpert 35B MoE: MLE-Bench Lite medal rate 37.9% to 72.7% 🇨🇳 - 🧠 NVIDIA midtraining recipe: SWE-bench Verified 60.7 to 64.5 on Nemotron 3.5 Lightning - 🤖 REKI 2.4B VLA: matches 2x bigger models, cuts pose oscillations ~4x 🇫🇷 - 🎬 OmniCompass: 50 MLLMs tested, open models trail on audio-visual video tasks - 🧊 InfiniSplat v2 cuts LPIPS 25.8% and DISTS 32.9% on 3D Gaussian splatting - 🧱 3DCodeVerse: 2M-program 3D code dataset, wins 90% of human comparisons - ⚡ MJev-Spatial: 65.64% spatial VQA, 2.5x faster than fastest autoregressive baseline - 🎧 DBS dynamic band splitting: 29.0% relative SI-SDR gain on DnR v2 - 🗂️ Mosaic: multi-session agentic RL beats broadcasting on BrowseComp with fewer sessions - 🔧 AutoSFT: +6.5 IFEval over single agent, +3.7 MBPP+ in code transfer - ֍ ChatGPT Chrome extension: @chrome reads all open tabs, Computer Use drives them, all plans - ✻ Anthropic gives 6 METR staff internal access to audit models - 🔒 Perplexity open-sources SPACE: one Firecracker microVM per agent task, Apache 2.0 - 🐧 Tencent Hy-Memory-30B-A3B: 30B total, 3B active, Apache 2.0 on HF 🇨🇳 - 🗜️ NVIDIA GLM-5.3-Flash-FP4-Local: 175.2B params, quantized for local runs Postgres 19 losing features to AI code review is the most honest AI story of the week. Which one are you actually running this weekend?