AI MODEL MONITOR
AI Model Release Monitor — BASELINE RUN (2026-09-11)
2026-09-11
This is the first run. The database has been seeded with the current state of major Chinese open-weight labs plus other tracked sources. From tomorrow, this email will only list models newly seen since the previous run. Today’s seeded models are listed below for reference.
CHINA / LOCAL MODELS
Qwen3.8-27B
(Alibaba (Qwen))
HuggingFace
A native multimodal dense model aimed at real-world coding and office workflows, with a 262K native context window extendable to 1M and broad inference-engine support.
Hardware: ~61GB VRAM at full precision (FP16/BF16), ~31GB at INT8, ~15GB at 4-bit (GGUF/AWQ) quantization. Fits on a single high-end consumer GPU (RTX 4090/5090, 16-24GB) at Q4.
Baichuan-M3-235B
(Baichuan)
HuggingFace
Baichuan’s latest general-purpose model, continuing the lab’s enterprise and medical-domain focus, released with GGUF/GPTQ/FP8 quantized variants for easier local deployment.
Hardware: ~517GB VRAM at full precision (FP16/BF16), ~259GB at INT8, ~129GB at 4-bit (GGUF/AWQ) quantization. Requires a multi-GPU server (e.g. 2-4x A100/H100 80GB) even at 4-bit, or slow CPU/RAM offload.
ERNIE-4.5-21B-A3B-Thinking (Baidu)
HuggingFace
A reasoning-focused ERNIE variant using a small-active-parameter MoE design (3B active of 21.8B total), making it one of the more locally-runnable Chinese reasoning models.
Hardware: ~48GB VRAM at full precision (FP16/BF16), ~24GB at INT8, ~12GB at 4-bit (GGUF/AWQ) quantization. Fits on a single high-end consumer GPU (RTX 4090/5090, 16-24GB) at Q4. (MoE: only ~3B active per token, so inference speed is much faster than the total size implies, but full weights still need to be loaded.)
DeepSeek-V4.1-Flash (DeepSeek)
HuggingFace + MarkTechPost
The smallest model in DeepSeek’s new V4.1 architecture line, adding native multimodal (vision) input, a 1M-token context window, FP4 KV cache, and cross-layer attention reuse for cheaper long-context serving.
Hardware: ~1679GB VRAM at full precision (FP16/BF16), ~840GB at INT8, ~420GB at 4-bit (GGUF/AWQ) quantization. Effectively datacenter-only – multiple 80GB+ GPUs even at 4-bit quantization; not practical on consumer hardware.
MiniMax-M3 (MiniMax)
HuggingFace
MiniMax’s coding- and agent-focused frontier model with long-context handling and multimodal input, positioned as a low-cost alternative for long-running agent tasks.
Hardware: ~940GB VRAM at full precision (FP16/BF16), ~470GB at INT8, ~235GB at 4-bit (GGUF/AWQ) quantization. Effectively datacenter-only – multiple 80GB+ GPUs even at 4-bit quantization; not practical on consumer hardware.
Kimi K3 (Moonshot AI)
HuggingFace + SCMP
The largest open-weight model released to date (2.8T params), built for long-horizon agentic coding and reasoning with a 1M-token context and novel Kimi Delta Attention for faster decoding.
Hardware: ~6.0TB VRAM at full precision (FP16/BF16), ~3.0TB at INT8, ~1529GB at 4-bit (GGUF/AWQ) quantization. Effectively datacenter-only – multiple 80GB+ GPUs even at 4-bit quantization; not practical on consumer hardware.
Step-3.7-Flash (StepFun)
HuggingFace
StepFun’s latest multimodal Flash model, shipped with FP8, NVFP4, and GGUF variants at release for a range of hardware targets.
Hardware: ~443GB VRAM at full precision (FP16/BF16), ~222GB at INT8, ~111GB at 4-bit (GGUF/AWQ) quantization. Requires a multi-GPU server (e.g. 2-4x A100/H100 80GB) even at 4-bit, or slow CPU/RAM offload.
Hunyuan Hy4 (preview) (Tencent)
HuggingFace
Tencent’s preview of its next-generation Hunyuan model, aimed at long-context agent work and released under a fully permissive Apache 2.0 license, unusual for a model this large.
Hardware: ~1716GB VRAM at full precision (FP16/BF16), ~858GB at INT8, ~429GB at 4-bit (GGUF/AWQ) quantization. Effectively datacenter-only – multiple 80GB+ GPUs even at 4-bit quantization; not practical on consumer hardware.
EVIE-8B (Tencent)
HuggingFace
A visual-document-retrieval embedding model from Tencent for search over scanned documents, PDFs, and screenshots rather than a general chat model.
Hardware: ~18GB VRAM at full precision (FP16/BF16), ~9GB at INT8, ~5GB at 4-bit (GGUF/AWQ) quantization. Runs comfortably on a single consumer GPU (8-12GB) or even CPU/laptop RAM at Q4.
MiMo-V2.5 (Xiaomi)
HuggingFace
Xiaomi’s MiMo model line update, with Pro/Base/ASR variants and a fast-decoding “DFlash” version aimed at efficient inference.
Hardware: ~684GB VRAM at full precision (FP16/BF16), ~342GB at INT8, ~171GB at 4-bit (GGUF/AWQ) quantization. Effectively datacenter-only – multiple 80GB+ GPUs even at 4-bit quantization; not practical on consumer hardware.
GLM-5.3-Flash (Zhipu AI / Z.ai)
HuggingFace
Zhipu’s cost-efficient flagship variant, tuned for coding and agentic workflows with a 1M-token context window and strong benchmarked cost-per-task efficiency among open models.
Hardware: ~707GB VRAM at full precision (FP16/BF16), ~354GB at INT8, ~177GB at 4-bit (GGUF/AWQ) quantization. Effectively datacenter-only – multiple 80GB+ GPUs even at 4-bit quantization; not practical on consumer hardware.
Tracking 11 models total (11 China-origin) since monitoring began.
DAILY DIGEST
AI News Digest — TEST (pipeline check)
2026-09-11
This is a one-time test send to confirm the monitor pipeline (fetch -> diff -> summarize -> email) works end to end before the daily schedule goes live. This is a BASELINE run: state.json had no prior history, so no “new since last check” list is included below – the summaries are just a sample of current headlines from a few sources.
GLOBAL
OpenAI News: OpenAI announced ChatGPT for Financial Services, a version of ChatGPT tailored for banks and asset managers with compliance guardrails and financial-data connectors.
Google DeepMind: DeepMind introduced WeatherNext 3, its most advanced global weather forecasting AI model, alongside AlphaGenome Atlas, a predictive map of DNA variant effects across the human genome.
TechCrunch: Nvidia CEO Jensen Huang projected the company will grow revenue about 70% next year, citing sustained AI infrastructure demand.
CHINA
SCMP: Ant Group will let AI shopping agents pay through 10 digital wallets (including AlipayHK and KakaoPay), expanding agentic commerce across Asia.
SCMP: AI safety experts are reportedly alarmed ahead of the Xi-Trump summit, with China’s AI capabilities becoming a bigger geopolitical flashpoint.
Once you confirm this looks right, the daily version will only include items that are new since the previous run, grouped the same way, with direct links to each story.
DAILY DIGEST
AI News Digest — BASELINE run (2026-09-11)
2026-09-11
NOTE: This is the first run of the daily AI news monitor. No prior snapshot existed, so this is a BASELINE establishing the starting point — it is NOT a “0 new stories” day. All 20 RSS feeds and 12 page-monitored sources (DeepSeek, Qwen, Anthropic, Zhipu, Moonshot, MiniMax, Microsoft AI, Meta AI, TechNode, Jiqizhixin, China AI Bench, DeepSeek GitHub) were queried successfully. Going forward, daily digests will report only what’s new since the prior day. Below is a roundup of the most notable recent stories (mostly the last 1-2 days) to kick things off.
Feed status: all sources reachable except the known-flaky ones — VentureBeat AI (429 rate limited), MarkTechPost (403 forbidden), Hacker News AI query (timed out this run). These are expected intermittent issues, not broken integrations, and will be retried tomorrow.
GLOBAL
OpenAI / frontier models
OpenAI released GPT-6 Astra (~Sep 3), calling it its most intelligent and aligned model yet — the first to earn a “Critical” cybersecurity rating under its own Preparedness Framework (99.9% ARC-AGI-3, 100% ExploitBench). Demand has been so high OpenAI paused new ChatGPT Pro subscriptions on Sep 10.
OpenAI introduced the Agents API and GPT-Live-1 for voice, and launched ChatGPT for Financial Services.
A board member says OpenAI is “not on track” to reduce risk of catastrophic loss of control.
Anthropic
Anthropic detailed distillation campaigns it says came from Alibaba, Moonshot AI, and DeepSeek.
Anthropic published a report on bad actors attempting to misuse Claude for bioweapons research, and more of its own researchers are publicly warning about AI risk (Musk dismissed the warnings as a “psyop”).
Earlier (Sep 1-2): Anthropic shipped Claude Fable 5.1 (GA) and Claude Mythos 5.1 (invite-only, tuned for cybersecurity/life-science), with a 75% cache price cut.
Nvidia / infrastructure
Nvidia agreed Sep 3 to acquire Hugging Face for $12.93B (expected to close H1 2027) — Hugging Face hosts 3M+ models and 500K+ datasets used by 18M+ developers. Source: Nvidia 8-K filing / Jensen Huang announcement.
Jensen Huang says Nvidia will grow “an astounding 70%” next year.
Google / Meta / other labs
Google shipped Gemini 3.8 Flash plus a “Cyber” variant tuned for defensive security work (Sep 2). Gemini app also now available on Windows.
Meta’s AI companion app “Muse” is now the #2 app in the US.
Regulatory / safety
EU designated ChatGPT a “Very Large Online Search Engine” under the DSA (Aug 31), the first standalone AI service under that top supervisory tier.
California enacted laws restricting chatbots and banning teens from “addictive” social media.
CHINA
DeepSeek
Reportedly plans to deploy 160,000+ Huawei Ascend 950DT chips at a new Inner Mongolia data center — among the largest known Huawei AI chip clusters anywhere (Bloomberg, Sep 4).
Closed a $7.4B funding round (~RMB50B) at a ~$74B pre-money valuation in an “unusual structure,” per The Information (Aug 29/30) — was reportedly nearing a further pre-IPO round at that $74B mark with a 2027 STAR Market listing taking shape.
July revenue reportedly reached ~$70M, roughly 10x its 2025 level.
Latest model news: DeepSeek-V4.1-Flash (“smarter, faster, more efficient”) posted on the DeepSeek newsroom Sep 10; also shipped DeepSeek-V4-Flash-Vision-Exp (its first multimodal V4 release, Aug 31).
Alibaba / Qwen
Qwen open-sourced Drive-1.0, an autonomous-driving model integrating 3D perception, VQA and motion planning (Sep 3).
Qwen3.8-Max-0902 snapshot topped the Code Arena WebDev leaderboard (Sep 2); Qwen also passed 3 billion cumulative downloads, more than Google and Meta’s model downloads combined.
Moonshot AI (Kimi)
Confidentially filed for a Hong Kong IPO (Reuters, Sep 3).
Talking to Microsoft, Amazon and Google about hosting Kimi K3 on their clouds for a revenue share of up to 30% (Reuters, Aug 26).
Zhipu (GLM) / MiniMax
Zhipu (Z.ai) opened its first official Tmall store for AI subscriptions (Sep 2); GLM-5.3 open weights landed (a day late) on Aug 29.
MiniMax raised its Alibaba Cloud spend cap 220% to $1.2B through 2028, and Saudi Arabia’s state-backed Humain shipped an Arabic model built on a MiniMax base (Sep 3).
Other China AI
iFLYTEK open-sourced Spark X2.5 edge models with 1M-token context (Sep 1).
China’s daily AI token-call volume passed 500 trillion as of June 2026.
China set new 2030 targets to accelerate AI/software industry integration (Sep 11).
China’s top court issued its first AI judicial guidelines, targeting deepfakes and copyright disputes (Sep 8).
Researcher claims a 6TB leak of China LLM router logs exposed enterprise credentials.
Source note: China roundup drawn from chinaaibench.com/news (curated daily brief covering DeepSeek/Qwen/Zhipu/Moonshot/MiniMax), SCMP China Tech, Pandaily, and ChinaTechNews RSS feeds.
This is an automated digest from the AI News Monitor. Full source list and scripts live in the monitor-ai-news workflow. Breaking developments that clear a high importance bar will also trigger a separate immediate alert email outside this daily digest.