AIGC Daily (Nov 13, 2026): Long-Context LLMs, Video Generation, and AI Compute Progress
This AIGC Daily roundup covers a lightweight ERNIE model with long-context features, a video generation model focused on facial micro-expressions and its API pricing, AI chip and supernode roadmaps, billing multipliers on multi-model platforms, and an embodied AI model that jointly predicts video frames and robot actions.
1. Long-Context Capabilities Advance Through Sparse Attention and History Compression
According to official materials, the new lightweight ERNIE model, ERNIE-Light-DT-1.0, introduces two architectural features worth noting: BSA sparse attention and feedback-driven history compression. BSA selects key tokens during attention computation and skips low-relevance segments. This reduces computational complexity from quadratic to near-linear in sequence length, keeping memory use and inference latency manageable when the model processes very long contexts. Feedback-driven history compression does not place the full history into the context window. Instead, it summarizes, filters, and reorganizes earlier content based on task feedback, so the model reasons over the relevant history rather than all of it.
Combined, these two mechanisms allow a maximum output length of 128K tokens. That is enough to generate a complete technical document chapter, a long codebase, or a full analytical report. For developers, another practical feature is a thinking mode that can be switched on or off. Enable it for tasks that require deep reasoning, such as debugging complex code or working through multi-step math derivations, to get higher-quality output. Disable it for high-frequency, low-latency calls, such as customer service replies or simple rewrites, to cut response time and token costs. This “one model, two modes” design is becoming a mainstream approach for productizing reasoning models. Its engineering value is that businesses do not need to maintain two separate model integrations. A single parameter lets them balance quality and cost.
# Illustrative pseudocode: how to call the thinking mode switchresp = llm.chat(messages=msgs,enable_thinking=True, # Enable for complex tasks; set to False for simple onesmax_output_tokens=128000)
2. Human Video Generation Moves Toward Commercial Use, with Micro-Expressions as a Differentiator
The video generation market is highly competitive. The official ERNIE-Wizshot-ME model takes a narrower approach by focusing on close-up human facial expressions. The model accepts three input modalities: images, text, and audio. It can generate 4 to 15 second person videos at 1080p resolution, with an emphasis on facial micro-expressions and mouth movements.
From a technical standpoint, multimodal conditioning serves specific purposes. The image input anchors the person’s identity and appearance. The text description controls the scene and the action semantics. The audio input drives lip movement, facial expression, and speech rhythm so they stay in sync. Combining these three inputs addresses what digital human and virtual host applications care about most. Many existing solutions produce realistic distant shots, but stiff micro-expressions and lip-sync errors become visible in close-ups. Making close-up facial expressions the optimization target shifts the evaluation standard from “does the video look good” to “does the emotion come across convincingly.”
On the business side, Baidu AI Cloud lists the API price for this micro-expression model at 0.7 yuan per second. At that rate, a 15-second video costs about 10.5 yuan to generate. For use cases such as rapid marketing asset production, short-video pre-screening, and virtual hosting, this cost is in a range that teams can compare with the cost of manual production. When selecting a provider, businesses should evaluate expression naturalness, lip-sync error, multi-frame consistency, and the concurrency and rate-limiting policies for batch calls.
3. AI Compute Foundation: Chip Roadmap and Thousand-Card All-Optical Interconnect Supernodes
In AI infrastructure, the conference presented a clear medium- to long-term roadmap. The Kunlunxin M100 is already on the market. The M300 is planned for release in 2027, and several chips based on the N architecture are planned to launch gradually in 2028. The M300 and N-architecture models are roadmap items announced at the time of the event. They are not yet shipping products. Businesses planning procurement or architecture should distinguish between capabilities they can deploy now and long-term roadmap items.
The Tianchi 1024-card supernode was announced at the same time and is planned for release in 2027. According to on-site reports, the product uses an all-optical interconnect and supports 1024 accelerator cards working together. The official statement says its card-to-card interconnect bandwidth has improved significantly over the previous generation, which supports training of large-parameter models. All-optical interconnect addresses the distance and power limits of electrical signal transmission. In clusters with thousands of cards, communication between nodes is often the real bottleneck in large model training. Optical interconnect can significantly reduce communication latency and the share of energy spent on communication, so the efficiency of collective operations such as AllReduce degrades more slowly as the cluster grows.
Typical bottleneck evolution in cluster scaling:Single machine with 8 cards → card-to-card NVLink / on-chip interconnectMulti-machine scaling → inter-node network (electrical interconnect, limited by bandwidth and latency)Thousand-card supernode → all-optical interconnect, reducing the perceived network gap to single-machine level
For teams building training clusters, the significance of products like this supernode is in where the “communication wall” moves. When card-to-card bandwidth is high enough, model parallelism and data parallelism can be partitioned more aggressively, and training throughput scales more linearly.
4. Credit Multipliers on Multi-Model Platforms: A New Billing Perspective for Business Teams
Intelligent programming tools and AI agent products are increasingly adopting multi-model access. These platforms aggregate several models for users to choose from, and they display the credit consumption multiplier for each model on the selection page. Recently, one product updated its list of available models, clarified how multipliers are displayed, and temporarily lowered prices on some lightweight models. These features were previously available mainly to team and enterprise subscribers. This update extends access to more users and adds clearer credit usage indicators. Available models and promotions are shown on the product page and may change.
Developers should take this trend seriously from a cost engineering perspective:
- A multiplier is a unified billing currency. The platform converts the costs of different models into a common credit unit. The multiplier is each model’s cost coefficient relative to a baseline model.
- Routing becomes valuable. In agent workflows, sending simple steps to low-multiplier models and reserving high-multiplier models for key reasoning steps is the most direct way to control costs.
- Transparent display is an industry improvement. Clear multipliers let teams estimate costs during architecture review, instead of reacting to bills after deployment.
When integrating these platforms, teams should build a “task tier to model mapping” table, review consumption data regularly, and adjust routing rules over time.
5. A New Paradigm for Embodied AI Modeling: Joint Prediction of Video and Actions
In robotics, a mixture-of-experts (MoE) embodied model with about 27 billion parameters demonstrates a new modeling paradigm. For robot tasks, the model jointly predicts video frames and action sequences. The core idea is that the model acts as both a world model, which predicts how the environment will change, and a policy model, which decides the next action. Visual prediction gives the action decision a forward-looking basis, and the action output in turn constrains whether the visual prediction is plausible. The two are calibrated against each other in a unified representation space.
For developers, the adoption path is fairly clear. Teams can perform full fine-tuning, or use parameter-efficient methods such as LoRA to adapt the pretrained capabilities to specific tasks. Official code repositories typically provide checkpoints and usage instructions for typical setups, such as single-arm and dual-arm robots, which lowers the barrier to reproduction and secondary development. The advantage of the MoE structure in this kind of model is that robot perception, decision-making, and execution involve different capabilities, including visual understanding, physical common sense, and motion control. Expert routing activates the appropriate sub-networks for each input. This increases the model’s total capacity without increasing inference cost proportionally.
6. Summary and Observations
Taken together, the day’s updates point to three main trends. First, long-context work is no longer just a race for larger windows. Sparse attention, history compression, and the thinking mode switch together form an engineering answer to the trade-off between quality, cost, and latency. Second, video generation is scaling up through specialized scenarios, such as human micro-expressions, and clear per-second pricing. Third, competition in compute is shifting from single-chip performance to interconnect architecture and the delivery pace of cluster-level systems.
For technical teams, a practical approach is to select and cost models based on task characteristics, and to factor long-term compute roadmaps into architecture planning. Waiting for a perfect solution before starting is not practical. All information here comes from official disclosures and on-site reports. For details on future models and launch schedules, refer to the final official releases.
