## Weekly Overview - **Model competition shifts from "stacking compute" to "compute efficiency and post-training"**: Microsoft MAI-Base-2 claims to surpass the GPT-5.6 backend model using one quarter of OpenAI's compute, DeepSeek V4.1-Flash forms a Pareto improvement with "stronger, faster, shorter outputs," and GLM 5.3 is broken down as "architecture unchanged, approaching the closed-source frontier entirely through post-training"—the frontier gap is being compressed by engineering efficiency rather than sheer scale. - **Agents move from demos to a coexistence of "reliability crisis" and "infrastructureization"**: Gemini 4 Argon's four agents autonomously complete Minecraft in 27 hours, but a 99% single-step success rate on long-horizon tasks drops sharply to 37% after a hundred steps, and the four major Personal Agents achieve only a 33% success rate in real-world tests; meanwhile, OpenRouter agent weekly usage reaches 8x that of humans at 132.6T tokens consumed per week, and Walmart launches an MCP agent gateway allowing AI to place orders on users' behalf. - **The open-weight camp breaks through simultaneously on coding, world models, and on-device**: Reflection Beam 501B MoE (Apache 2.0), StepFun Step 5 Preview 600B MoE/1M context, VAST EdenLive real-time world model, Meta Muse Glimmer 30B on-device Agent, and ModelBest 35B-A3B MoE running in real time on phones—open source is no longer a "catch-up option" but a distribution and cost weapon. - **The mismatch between AI capex and revenue becomes the market's core contradiction**: Bank of America says AI capex exceeds expectations by 50% while revenue exceeds by only 20%, CoreWeave stock falls 25% for the week with junk bond yields spiking to 15%, and SoftBank stock nearly halves; yet SpaceX achieves its first quarterly profit with $7.6 billion in single-quarter revenue from AI compute leasing, and AWS says AI server payback periods are under three years—financing structures are being repriced. - **Safety and governance move from "principled discussion" to "real-world incidents"**: Two 9.2-severity high-risk vulnerabilities in the PHP core trigger emergency global takedowns (discovered with AI assistance), hackers use AI agents to breach over 100 websites, Anthropic revokes API keys due to an OpenCode distillation attack and elevates distillation to a national security issue, and Microsoft launches Guardian Agents to monitor agents as insider threats. - **On-device and interconnect architectures break through in tandem, rewriting the physical boundaries of Agents**: NVIDIA RTX Spark laptops start at $2,399 with up to 1 petaflop of compute and 128GB unified memory, Marvell photonic modules allow HBM and GPUs to be separated by up to 50 meters, and ModelBest LES enables a 35B-A3B MoE to decode at over 20 tokens/s on an iPhone. ## Hot List 1. **Microsoft MAI-Base-2: Surpassing the GPT-5.6 backend model with one quarter of the compute** Microsoft MAI-Base-2 — https://aihot.news/items/c3sr2b6fbgvdu00echcbvywkj Essence: This is a direct strike in the "compute efficiency narrative"—if one quarter of the compute can match or even surpass comparable frontier models, then the moat of frontier labs shifts from "how many GPUs you own" to "training recipes and data efficiency," while Microsoft's dependence on OpenAI for its own models is substantially weakened, further cementing the cloud vendor's dual identity of "selling both compute and models." —Possibility: Individual developers need not wait for model weights; they can first compare MAI-Base-2 against existing GPT series on Azure AI for cost/quality ratio on their own tasks, running an A/B with the same batch of 50–100 real requests, focusing on "effective output per dollar" rather than leaderboard scores; if the gap is within 10%, migrate cost-sensitive pipelines (batch summarization, classification, extraction) over and keep high-difficulty reasoning on the original model, forming a hybrid router. 2. **Anthropic anti-distillation defense: token theft reduced by 90%–99%, elevated to national security** Anthropic anti-distillation defense — https://aihot.news/items/sj5d4076973hvnycu7q7amr4z Essence: Model distillation has been formally classified as a security threat rather than "gray arbitrage," meaning API providers will more aggressively implement output watermarking, rate anomaly detection, and key accountability; for teams relying on "training small models with strong model outputs," this low-cost path now carries rising compliance and technical risk, while defenders gain a quantifiable deterrent demonstration. —Possibility: If you are doing model fine-tuning or data synthesis, first audit your training data sources, separately tag "outputs from third-party APIs" and assess substitutability; at the same time, build output fingerprinting and call anomaly monitoring on your own product side (high-frequency single keys, templated prompts, abnormal output distributions) to avoid becoming the party whose keys get revoked—Anthropic has already actually revoked affected keys due to the OpenCode incident (https://aihot.news/items/l77k5lljwa77u0w19gsgesqq0). 3. **DeepSeek V4.1-Flash + Search/Fetch API: an Agent retrieval foundation at 5 yuan per thousand calls** DeepSeek Search/Fetch API — https://aihot.news/items/e3wq1hzkjq59f8g3prttrddb5 Essence: Search and fetching are pushed into the price band of "5 yuan per thousand calls, free on timeout, readable WeChat public account articles," directly changing the retrieval cost structure of Agents—the "fetching and cleaning" step, previously the most easily overlooked in RAG systems, can now be outsourced to infrastructure billed per successful request; combined with V4.1-Flash topping both the Arena Agent and Code open-source leaderboards (https://aihot.news/items/pwdldfnta3ut27a749j2wen17), this forms a combo of "cheap model + cheap retrieval." —Possibility: Take an existing RAG or research Agent and replace its fetching layer with DeepSeek Fetch, recording three things: total cost per task, fetch success rate, and the proportion of "free on timeout" triggers; if cost drops by more than 50% and success rate is no lower than the original solution, abstract the retrieval layer into a switchable interface (with DeepSeek / TinyFish free quota https://aihot.news/items/... as backup) to avoid single-point lock-in. 4. **Reflection Beam 501B MoE open-sourced + StepFun Step 5 Preview 600B MoE/1M context** Reflection Beam — https://aihot.news/items/v27j8fnpclhpwv5blr21n4650 Essence: Open-weight models formally enter the "500B–600B MoE, Apache 2.0, aimed at coding and long-horizon agents" range, with only 23B–27B activated parameters, meaning inference costs are controllable while capabilities approach closed-source; this is a decisive change for enterprise scenarios that "must be privately deployed" (finance, healthcare, government)—no longer a choice between "compliance" and "capability." —Possibility: First try Beam's 1M context for free for a limited time on OpenCode (https://aihot.news/items/...), using a real long document plus a multi-step tool-calling task to test "whether long context is actually usable" rather than just looking at the window number; if usable, evaluate the feasibility of deploying the 23B-activated model with 4-bit quantization on an 8×H100-class self-owned cluster, using "data never leaves the domain" as a differentiating selling point to customers. 5. **Gemini 4 Argon's four agents autonomously complete Minecraft in 27 hours** Gemini 4 Argon agents complete Minecraft — https://aihot.news/items/... Essence: This is a public stress test of long-horizon Agent capability—four agents in a random world divide labor, warn each other, and defeat the Ender Dragon without human input, proving that "multi-agent + long-horizon planning" has crossed the demo threshold in open environments; but the same week's warning that "a 99% single-step success rate drops to 37% on hundred-step tasks" (https://aihot.news/items/...) reminds us: the capability ceiling and the reliability floor are being pushed up simultaneously, and the gap between them is the real battlefield for productization. —Possibility: Don't directly build a "general long-horizon Agent"; instead pick a vertical workflow with controllable step count (10–20 steps) and rollback-able failures (such as reconciliation, report generation, batch asset processing), first measuring the actual decay curve of "single-step success rate × steps" in your own workflow; design retries as idempotent operations and checkpoint at every step, using engineering to pull 37% back into the usable range. 6. **VAST open-sources real-time world models EdenLive and Golem: 720p/24+FPS on a single RTX 5090** VAST EdenLive — https://aihot.news/items/a52jv2jag7p4yonzs0zemg8np Essence: World models move from "generating video" to "real-time interactive decoupled state and rendering," with a 5B-parameter model reaching 720p/24+FPS on consumer-grade GPUs, meaning "generate a playable world from one sentence" goes from research demo to a locally deployable component; together with Viggle JST-2's physics-native world model (https://aihot.news/items/...), this greatly lowers the cost of training environments for games, simulation, and embodied intelligence. —Possibility: Use EdenLive to build a minimal playable prototype: fix a scene (such as room navigation), verify the latency and consistency of "reference image → real-time interactive frames," focusing on recording visual drift after long runs; if drift is controllable, connect it to a simple keyboard/mouse control loop and test its usability as an embodied intelligence simulation environment, rather than first pursuing visual polish. 7. **NVIDIA RTX Spark laptops go on sale: from $2,399, up to 1 petaflop, 128GB unified memory** NVIDIA RTX Spark — https://aihot.news/items/p4onrft3oltqqk5e5xv21khvc Essence: On-device compute is pushed to the level of "1 petaflop + 128GB unified memory," combined with ModelBest LES enabling a 35B-A3B MoE to run in real time on an iPhone (https://aihot.news/items/...) and Meta Muse Glimmer 30B targeting 24GB VRAM devices (https://aihot.news/items/...), the "model capacity ceiling" of local Agents is significantly raised; privacy sensitivity, offline availability, and zero marginal inference cost become new selling points for on-device. —Possibility: If you are building privacy-sensitive tools (legal, healthcare, personal knowledge bases), you can now seriously evaluate the product form of "local 30B-class model + on-device Agent framework": first run the complete pipeline of "document parsing → multi-step tool calling → results written to disk" on a 24GB VRAM device,