Sakhanda Wire
NVDA $154.00 -1.37% MSFT $506.28 -2.26% GOOGL $280.87 -1.00% META $620.34 -0.97% AMZN $256.88 -0.34%

MarkTechPost

396 stories · all sources

MarkTechPost

Buoyant Releases Kinference Control Plane for vLLM on Kubernetes

Buoyant has released Kinference, an Apache 2.0-licensed open-source control plane for vLLM deployments on Kubernetes. The Rust-based project sits ahead of vLLM workers and uses declarative policies to route requests, apply token-rate limits, balance traffic using KV-cache data, and provide per-tenant metrics. The current release is v0.1.0-alpha.7 and is configured through five Kubernetes custom resources.

MarkTechPost

Zyphra releases open-source FP8 attention kernels for AMD MI355X

Zyphra has released ZYFA, an open-source library of FP8 and BF16 prefill attention kernels for AMD Instinct MI355X GPUs. The Apache 2.0 project targets causal forward attention for long-context workloads, with support for FP8 e4m3, BF16, head dimension 128 and MLA shapes of 192/128. It runs on MI355X hardware with ROCm 7.2.3 and PyTorch 2.12, and provides a drop-in flash_attn_func interface.

MarkTechPost

Thinking Machines Lab launches 594B-parameter Inkling 2 Small

Thinking Machines Lab has released Inkling 2 Small, the first model in its second Inkling generation. The sparse Mixture-of-Experts system has 594B total parameters, with 17B active per operation, and supports text and image inputs, variable reasoning effort and a 1M-token context. Audio is not supported.

MarkTechPost

Google launches Gemini Omni 1.2 Flash for 15-second video generation

Google has released Gemini Omni 1.2 Flash, a multimodal model for generating and editing video. It accepts text, images, short video clips and audio, and produces video with audio. The model is generally available through the Gemini API and Google Cloud; its weights are closed and it cannot be deployed locally.

MarkTechPost

Baidu releases 280B-parameter ERNIE 5.5 Flash Base model

Baidu has released ERNIE 5.5 Flash Base, a text-only sparse Mixture-of-Experts checkpoint containing 280B parameters and activating 14B for each token. The weights represent pre-training only, with no mid-training or post-training. Baidu says the model was trained from scratch on 33.5T tokens using the Muon optimizer. Its training context is 128K tokens, while fine-tuning and inference support up to 1M tokens.

MarkTechPost

Together AI launches compiler for continuously optimized model inference

Together AI has introduced Together Compiler (TKCC), an agent-driven system that repeatedly optimizes a model’s inference forward pass as production traffic changes. The system accepts a server launch command, traffic profile and hardware configuration. Coding agents propose alternatives, while a persistent serving engine measures them and can deploy the strongest result. Together AI says the compiler already operates on several of its serverless endpoints.

MarkTechPost

OpenAI releases GPT-Image-2.6 Flare and Sunburst models

OpenAI has released Images 2.6 for ChatGPT, Codex and its API. The developer release, GPT-Image-2.6, is available through two models: Flare, aimed at rapid iteration and interactive use, and Sunburst, designed for detailed compositions and more consistent editing. OpenAI says image generation is 50% faster than in the previous release, while Flare can produce an image in as little as six seconds when low quality is selected.

MarkTechPost

Google Research introduces Agile Deliberation for subjective image classification

Google Research has released Agile Deliberation, a human-in-the-loop framework for defining visual categories whose boundaries are subjective, such as “healthy food,” “inspiring artwork” or clickbait. Rather than requiring a complete definition upfront, the system helps users refine one by examining borderline images. The resulting text serves both as a readable policy and as the prompt for a vision-language classifier.

MarkTechPost

webAI Releases TwIL-8B Reasoning Model for Math and Code

webAI has released TwIL-8B, a dense 8.79B-parameter reasoning model post-trained from IBM Granite 4.2 8B. It supports English text input and output, uses a 131,072-token context window, and is available under the webAI Non-Commercial 1.0 license. A Q4_K_M GGUF version occupies 4.98 GiB, allowing local inference on a CPU or a GPU with modest VRAM; full bf16 weights require 16.38 GiB.

MarkTechPost

Ricursive releases 65B Ripple model for agentic coding

Ricursive Intelligence has released Ripple-65B Preview, a 65-billion-parameter dense open-weight language model trained from scratch for agentic coding. The model uses 12 trillion tokens, supports context lengths of up to 1 million tokens and is distributed under the Apache 2.0 license. The release includes instruction-tuned BF16, experimental NVFP4 and BF16 base checkpoints.

MarkTechPost

Tencent releases WeVisOne-VL for unified visual perception tasks

Tencent’s WeChat Vision team has released WeVisOne-VL, an Apache 2.0 vision-language model available in 2B and 4B versions. The models combine detection, referring, counting, pointing, OCR, segmentation and alpha matting. A 9.6M-parameter pixel decoder produces masks and mattes without requiring a separate SAM model. When a requested object is not present, the system can return None instead of generating a box.

MarkTechPost

Anthropic releases FlashProve for verifiable model inference

Anthropic’s Frontier Red Team has released FlashProve, an open-source research prototype using zkSNARKs to prove that a response was generated by a specific committed model. The proof does not disclose the model’s weights. The system is designed to address a verification gap in API services, where users currently cannot independently confirm which model produced an answer or whether token output was altered.