AI Daily: Ideogram 4.6 4K Edit, Hy4 Open, GPT-Live-1.1, Edge AI, Valkey Cache
This daily tech brief covers five AI updates: Ideogram 4.6 introduces native 4K generation and precise multi-round editing with masks; Tencent releases Hy4 mixture-of-expert model weights with 770B total parameters; OpenAI launches GPT-Live-1.1 real-time voice API with improved multilingual pronunciation; Microsoft Edge integrates Aion Instruct 2.3B on-device model for web summarization; and Valkey 9.2 adds distributed KV caching for LLM inference prefix sharing.
Ideogram 4.6: Native 4K Generation and Precise Multi-Round Editing
Ideogram released version 4.6 on November 19. The model can natively generate images up to 4K resolution and supports consecutive modifications on the same image. The API splits into Generate for new images and Precise Edit for modifying existing images, with optional masking to isolate the edit region.
Original image: purple gemstone with oblique lighting, metal bezel, and background as reference.
Edited image: gemstone changed to teal while preserving near-original camera angle and lighting.
Multi-round editing interface showing a garment project with multiple revision history.
Official benchmarks compare version 4.5 and 4.6 on the same prompts at High quality tier. Recoloring success rate (matching a specified color value) rose from 46% to 83%; mask-based editing success rose from 82% to 90%. The author notes these figures come from vendor-selected tasks and judgment criteria, and real-world results may vary with fine text, strong reflections, or unusual materials.
Pricing is per image. At High tier: 2K generation costs $0.10 , 2K editing costs $0.22 , and 4K editing costs $0.60 . Each round of multi-round editing incurs a separate API call and adds to the total cost.
Tencent Hy4: Open Weights with MoE Architecture and Custom License
On November 20, Tencent opened the Hy4 instruction model, an FP8 quantized variant, and the pre-training base model. The model card lists approximately 770 billion total parameters in the language backbone, with about 49 billion active parameters per token , and a context window of roughly 1 million tokens . Hy4 handles text, image, and video inputs, and provides deployment support for vLLM and SGLang.
Hy4 weights are split into instruction, quantized, and base model versions.
The author explains the mixture-of-experts (MoE) structure: a router assigns input to compute components, so each forward pass activates fewer parameters, but deployment must still accommodate the full weight set plus inference memory overhead. Quantization reduces storage and runtime burden, yet the scale still targets multi-GPU servers.
Weights are released under the Tencent Hy License 1.0 , based on Apache 2.0 with additional clauses. Large-scale businesses that meet the agreement threshold and offer third-party general model access or fine-tuning services must obtain separate written authorization. The license explicitly distinguishes pure internal use from providing model access externally.
OpenAI GPT-Live-1.1: Real-Time Voice API with Improved Multilingual Support
OpenAI's changelog on November 19 listed GPT-Live-1.1 , followed by a developer forum announcement. This model previously powered part of ChatGPT's voice experience. It can listen while speaking, improving multilingual pronunciation, accent consistency, and transcription. During conversation, tasks requiring deeper reasoning or tool use can be offloaded to a backend model.
Real-time voice pricing distinguishes active and eligible standby minutes; backend usage billed separately.
Voice sessions are billed by duration: $0.05 per minute for active time. WebRTC sessions configured with standby are charged $0.02 per minute during pauses in inference. The author cautions that user silence does not automatically trigger the lower rate across all connection types. Backend model and tool invocations are billed additionally.
Microsoft Edge: On-Device AI via Aion Instruct and Prompt API Origin Trial
On November 19, Microsoft announced Aion Instruct as Edge's built-in small language model, replacing Phi-4-mini. The 2.3B-parameter model's weights are also publicly released. Edge downloads the model on supported devices at first use; users can disable on-device AI in settings.
Illustration of device-cloud split; website access to local model requires browser, trial, and device conditions.
For websites, the change comes from the Prompt API origin trial . Eligible participating sites can invoke the browser's on-device model for text summarization, rewriting, or structured output. Availability depends on trial enrollment, browser version, and device capabilities. A separate Decisions API for fast structured judgments is in developer preview on Edge Canary and Dev channels for prototyping.
Valkey 9.2: Distributed KV Cache for LLM Inference Prefix Sharing
The open-source data store Valkey released version 9.2 on November 19, adding background thread snapshots, replication stream compression, and Path Hash. Alongside, a distributed KV cache capability for multi-GPU inference servers was introduced.
AWS architecture diagram showing multiple servers retrieving computed prefixes from a shared cache.
When an application repeatedly sends the same system instructions, tool descriptions, or long documents, the model re-processes the identical prefix each time. The distributed cache allows servers to share pre-computed intermediate results. When another server receives the same prefix, it fetches the cached result and only processes the new suffix. The cache stores computation outputs; the model still generates the final response for the request.
According to Linux Foundation release materials, under specified workloads inference cost drops up to 40% and time-to-first-token (TTFT) reduces up to 80% . Actual gains depend on prefix repetition rate, cache hit rate, data transfer speed, and cluster configuration. Applications with short, diverse requests are unlikely to reproduce these numbers.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
