Yandex open-sources Alice AI Search weights

// 1h agoMODEL RELEASE

Yandex open-sources Alice AI Search weights

Yandex has open-sourced Alice AI Search Pretrain under the Apache 2.0 license, publishing the weights and architecture behind its search-integrated AI answers. The model uses a hybrid encoder-decoder framework with Mixture-of-Experts routing, activating only roughly 600 million of its 35 billion parameters per token to cut computational overhead and deliver 40% more answers on equivalent hardware.

// ANALYSIS

While Google and Microsoft treat their AI overview architectures as proprietary trade secrets, Yandex's open release provides an unprecedented public look into the brutal latency and cost trade-offs of production web-scale retrieval. Sparse routing with under 2% active parameters per token illustrates that hybrid encoder-decoders remain far more viable for high-throughput search workloads than dense generative decoders. Giving the community open weights trained specifically on web retrieval summaries offers a purpose-built foundation for low-latency retrieval-augmented generation pipelines. Distributing weights under Apache 2.0 puts pressure on Western search competitors who have faced heavy scrutiny over search quality and non-transparent summarization algorithms.

// TAGS

yandexsearchopen-weightsllmmoeragai-searchapache-2.0

DISCOVERED

1h ago

2026-09-14

PUBLISHED

1h ago

2026-09-14

RELEVANCE

8/ 10

AUTHOR

yandexcom

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

INFRA1h ago

MixRoute outlines shadow traffic for model evaluation

MixRoute outlines shadow traffic mirroring as a superior testing strategy compared to relying on synthetic benchmarks or manual prompt tests when considering a new AI model. By duplicating a portion of live incoming user requests and asynchronously sending them to a candidate model, developers can assess real-world response quality, latency, and failure modes under actual production conditions without returning the shadow responses to users or impacting their experience.

UPDATE1h ago

emulate 0.11.2 boosts fidelity across six emulators

Local API emulator emulate has rolled out version 0.11.2, bringing updates and fidelity improvements across six built-in emulators: GitHub, Google, Resend, AWS S3, Slack, and Microsoft. The GitHub emulator receives notable enhancements, including support for raw contents and README endpoint responses, organization member and administrator seeding, proper checks handling for branch and tag names containing slashes, and App-level authentication for organization installation writes. Maintained by Vercel Labs, the project provides zero-config, stateful local replacements for external cloud services to enable reliable, offline-first development and CI testing.

MODEL1h ago

Mia-AI Lab unlocks 200K context for Qwen3.8-27B

Oluwaphilemon1 highlights an experimental EXL3 quantization release of Qwen3.8-27B developed by Mia-AI Lab, designed to resolve the KV cache memory bottlenecks that typically cripple local inference at extreme context lengths. Targeted at 24GB consumer GPUs such as the RTX 3090, 4090, and 5090, this configuration pairs EXL3 quantization with DFlash2 speculative decoding to sustain over 200K tokens of context without exhausting VRAM. By drastically cutting the combined memory footprint of the weights and active cache, the setup reportedly enables an RTX 3090 to achieve around 63 tokens per second across up to 256K context, making local codebase analysis and long-horizon agent workloads practical on consumer hardware.

Yandex open-sources Alice AI Search weights