GPT-5.6 improves AI efficiency across models, inference, and agentic workflows
GPT-5.6 enhances artificial intelligence efficiency across multiple dimensions, including model architecture, inference processes, and agentic workflows.
OpenAI proposes 'Useful Intelligence per Dollar' scorecard for AI value
OpenAI outlines a framework called "Useful Intelligence per Dollar" to help businesses measure the economic value of their AI investments beyond simple cost-per-token metrics. The article argues that success should be measured by work accomplished, reliability, and scalability rather than just adoption rates.
OpenAI lowers prices for GPT-5.6 Luna and Terra, introduces Fast mode for Sol
OpenAI has reduced the cost of its GPT-5.6 Luna and Terra models while introducing a new "Fast mode" for GPT-5.6 Sol to improve performance per dollar across enterprise workloads.
Clustered Codebook Quantization for 2D Gaussian-based Image Compression
Researchers present Cluster-Guided Vector Quantization (CGVQ), a method designed to improve the rate-distortion performance of Gaussian primitive-based image compression. The approach partitions Gaussian parameters into homogeneous groups prior to quantization, addressing the inefficiency caused by storing large numbers of floating-point parameters per primitive.
OpenAI cuts GPT-5.6 prices by up to 80% via self-optimization
OpenAI has announced significant price reductions for its GPT-5.6 models, driven by systemic efficiency improvements and recursive self-optimization techniques. The company reports that the cost of GPT-5.4 level intelligence has dropped approximately 13x in four months, with GPT-5.6 Luna prices slashed by 80% and Terra by 20%.
OpenAI cuts GPT-5.6 Luna price by 80% using Sol to optimize inference
OpenAI has significantly reduced the pricing for its GPT-5.6 model family, introducing a massive 80% price drop for GPT-5.6 Luna and a 20% reduction for GPT-5.6 Terra. These changes are enabled by GPT-5.6 Sol, which optimizes the model's forward pass and production kernels to improve efficiency.
Tencent open-sources Hy-MT2 multilingual translation models and IFMTBench
Tencent has released the Hy-MT2 family of "fast-thinking" multilingual translation models, including 1.8B, 7B, and 30B-A3B (MoE) sizes, along with the IFMTBench benchmark for evaluating instruction-following capabilities.
vLLM v0.25.1 fixes mixed-dtype allreduce RMSNorm quant fusions
The vLLM project released version 0.25.1, which includes a bugfix for mixed-dtype allreduce RMSNorm quant fusions.
vLLM Launcher provides a Windows desktop workbench for local LLM inference
A new project called vLLM Launcher has been released as a Windows desktop application designed to manage local large language model inference through WSL2. This tool allows users to control multiple engine backends, including vLLM, SGLang, and llama.cpp, from a single graphical interface.
llama.cpp b10331 fixes server get_info to report correct isolate working directory
The llama.cpp project released build b10331, which includes a fix for the server's `get_info` endpoint. Previously, when no explicit current working directory was provided, the function incorrectly fell back to the server process's working directory, even if a tools runtime was configured.
llama.cpp b10330 fuses rms_norm, mul, and rope on CUDA
The llama.cpp project has released version b10330, introducing a CUDA optimization that fuses the rms_norm, multiplication, and RoPE operations into a single kernel. This change is accompanied by memory range checks for the fused operations and new test cases for broadcast weights.
llama.cpp b10328 adds initial server tool isolation via Docker
The llama.cpp project has released build b10328, which introduces initial support for isolating tools within the server component using Docker. This update includes documentation and code adaptations to separate tool I/O sandboxes and rename relevant configuration flags.
llama.cpp b10327 fixes CUDA thread/block count in quantized cpy kernel
The llama.cpp project released version b10327, which includes a fix for the thread and block count in quantized copy kernel launches on CUDA. This update addresses an issue identified in pull request #26731.
llama.cpp b10326 accounts for vocoder pass in timings
The llama.cpp project released build b10326, which modifies how timing measurements are calculated by including the vocoder pass.
llama.cpp b10321 fixes NORM/RMS_NORM for partial simdgroup row lengths
The llama.cpp project released version b10321, which includes a critical fix for GGML_OP_NORM and GGML_OP_RMS_NORM operations on Metal. The update resolves an issue where threadgroups sized to leave a partial SIMD group resulted in incorrect mean and variance calculations due to dropped partial sums.
llama.cpp b10322 coalesces ssm_conv window loads for ~1.85x speedup
The llama.cpp project released build b10322, which includes a performance optimization that coalesces SSM_CONV window loads in the SYCL backend.
llama.cpp b10313 adds LRU scheduler to the server
The llama.cpp project has released version b10313, which introduces a Least Recently Used (LRU) scheduler for its server component. This update includes handling for request coalescing and fixes for streaming cases.
llama.cpp b10311 fixes Qwen3-TTS text stream duplication
The llama.cpp project released build b10311, which includes a fix for the Qwen3-TTS pipeline. The update resolves an issue where the model read the input utterance twice during generation.
llama.cpp b10312 release with server fix for busy models
The llama.cpp project has released version b10312, which includes a specific update to the server component. The key change addresses an issue where busy models were being evicted from the router.
llama.cpp b10307 fixes SYCL NVFP4 parsing
The llama.cpp project released version b10307, which includes a fix for the SYCL GPU code to correctly parse UE4M3 scaling factors used in NVFP4 quantization.