Sakhanda Wire
NVDA $153.26 -1.85% MSFT $506.01 -2.31% GOOGL $280.39 -1.17% META $620.55 -0.94% AMZN $257.19 -0.22%

MarkTechPost

396 stories · all sources

MarkTechPost

You.com Introduces Inference Router for 12 Open-Weight Models

You.com has launched Inference Router, a fully managed API for running 12 open-weight models in production. The service is compatible with OpenAI APIs and also provides Anthropic-compatible routes, allowing developers to use existing SDK code by changing the base URL and key. It charges per use and has no platform fee.

MarkTechPost

Liquid AI releases RIPL for low-latency LFM inference

Liquid AI has introduced RIPL, or Real-time Inference Pipeline, an inference engine designed to run Liquid Foundation Models in live request paths. The company says the system can return complete generated responses in under 50 milliseconds and maintain that latency under continuous peak traffic. RIPL supports LFM2.5 models ranging from 230M to 2.6B parameters on enterprise deployments using NVIDIA H100, H200, B200 and RTX PRO 6000 GPUs.

MarkTechPost

Gradium open-sources 31M-parameter Phonon speech model

Gradium has released Phonon, a family of compact text-to-speech models designed to generate audio locally on phones, computers, Raspberry Pi 5 boards and browsers. The English Phonon 31m has 31 million parameters, produces 24 kHz mono audio and is distributed under Apache 2.0. Its quantized GGUF file is 64.5 MB, while the Phonon 40m file is 74.6 MB.

MarkTechPost

Cohere releases Rerank 5 with passage extraction and calibrated scores

Cohere has introduced Rerank 5, a retrieval-model family designed to reorder search results before they are sent to a large language model. It is offered in Pro and Fast versions, both with a 32K-token context window and support for more than 100 languages. The models can process documents as well as semi-structured data, including JSON, tables and YAML, without requiring re-indexing.

MarkTechPost

OmniParseBench audit finds tied top parsers and changing slice leaders

A developer-led audit rebuilt OmniParseBench, Datalab’s open benchmark for document parsers, using the verdict files released for 27 parsers. The benchmark evaluates parser output through 16,288 pass-or-fail tests covering 2,937 pages. Its material ranges from scanned forms and handwriting to arXiv mathematics and dense tables.

MarkTechPost

IBM releases Granite Speech 5.0 2B for multilingual speech processing

IBM has released Granite Speech 5.0 2B, an open speech-language model for transcription and translation in six languages. The 2B-parameter model is distributed under Apache 2.0 and can run on GPUs through Transformers or vLLM, on Apple Silicon with mlx-audio, and locally through llama.cpp using GGUF. IBM recommends audio segments of roughly two minutes or less.

MarkTechPost

JOLT обучает одну языковую модель в ролях учителя и ученика

Perplexity и исследователи NVIDIA представили JOLT — метод постобучения языковых моделей, название которого расшифровывается как Joint On-policy Learning and Teaching. Один экземпляр модели одновременно выполняет роли учителя и ученика: учитель получает дополнительную информацию, например проверенное решение, а ученик видит только исходную задачу. В отличие от прежнего подхода OPSD, JOLT также обучает учителя, используя награду за результат и KL-штраф, удерживающий его поведение рядом с политикой ученика.

MarkTechPost

Whistle speech model tested against Whisper on a CPU

A coding guide evaluates Whistle, Cactus Compute’s 16.9 MB speech-to-text model, on a CPU. The model transcribes seven languages and provides word-level timestamps and probabilities. Its model card claims an advantage over Whisper base on LibriSpeech, which the guide tests using identical clips from the seeded test-clean and test-other samples.

MarkTechPost

NVIDIA Nsight Python brings GPU kernel profiling into Python

NVIDIA Nsight Python is an open-source interface for profiling GPU kernels directly from Python. Developers mark selected code with nsight.annotate and decorate benchmark functions with nsight.analyze.kernel. The resulting ProfileResults object can be converted into a pandas DataFrame containing hardware metrics for each annotated region, configuration and run, avoiding a separate command-line profiling step and later report inspection.

MarkTechPost

Recursive releases Flint, a 40B local coding model for agents

Recursive Superintelligence has released Flint, its first model, as an open-weight coding system for local agentic use. Flint is a text-only, decoder-only mixture-of-experts Transformer with 39.8B total parameters and about 2.5B active for each token. Its weights are available under the Apache 2.0 license. The model card documents a 262,144-token context, while Recursive’s blog mentions support of up to 512k tokens.

MarkTechPost

Vercel releases open-source durable-harness for crash-resilient coding agents

Vercel has released durable-harness, an open-source npm package designed to preserve coding-agent sessions when functions crash, time out or are redeployed. The package wraps existing harnesses rather than modifying them and provides adapters for Pi, OpenCode, fx, Claude Code and Codex. It is distributed under Apache-2.0 and remains an experimental 0.0.x API.

MarkTechPost

Mistral AI launches Voxtral 3 speech models and voice agent SDK

Mistral AI has released Voxtral 3, a speech-model family consisting of three transcription models, one text-to-speech model and a voice agent SDK built on LiveKit. The four models are in public preview through the Mistral API, with European, US and global endpoints. They can also be deployed on a customer’s own infrastructure through Mistral sales.