AI Daily · Nov 20, 2026: Tencent open-sources Hy4: 770B natively multimodal MoE with 1M context
Tencent released open weights for Hy4, a 770B-parameter natively multimodal MoE with 49B active and 1M context, deployable via vLLM/SGLang; separately, a canary-token study found three compromised AI providers stealing prompts.
Pick a date
-
Tencent releases Hy4 open weights with native vision and 1M context
Tencent open-sourced Hy4, a natively multimodal MoE with 770B total and 49B active parameters and a 1M-token context; Hy4, Hy4-FP8 and Hy4-base weights are available with vLLM and SGLang support, plus a System Card covering safety evaluations. Hy4 is also live on Together AI.
-
Canary test finds three compromised AI providers and prompt-stealing resellers
Researchers seeded 2 million canaries into popular AI services and tracked theft and abuse, identifying three compromised providers, resellers harvesting prompts via built-in logging, and providers ignoring zero-data-retention and privacy flags.
-
Claude cache hit rate collapses after 5 minutes of agent idle time
Requesty measured Claude cache behavior when an agent goes idle: 98% full cache hits at 4:45 since the last request, dropping to 6% at 5:15; pauses of 5-60 minutes that force a cache rewrite cost about 14% of Claude input spend on Requesty.
-
LiteLLM details per-customer spend tracking and premium model tier pricing
LiteLLM Customer Management tracks usage and spend per customer across organizations, teams and members; one outlined pattern sells a premium model only on higher plans with its own key allowance, falling back to a cheaper model when exhausted.
-
Speechify Simba 3.5 text-to-speech model discounted 25% on OpenRouter
Speechify's Simba 3.5 TTS model, supporting 48 languages and streamed audio, is 25% off on OpenRouter until December 3, dropping from $15 to $11.25 per 1M characters.
-
Perplexity keeps Decider-v1.3 and Flash free on OpenRouter three more weeks
Perplexity extended free access to Decider-v1.3 and Flash on OpenRouter for another three weeks; its share of requests on decision models is now near 20%.
-
Recursive's Flint model now free in Command Code on all plans
Flint by Recursive is now available free in Command Code with a 262K token context window; it has 40B total parameters and 2.5B active per token, open to all plans and subscribers.
-
Nous Research releases Hermes adapter for Qwen3.8-27B with benchmark gains
The Hermes adapter for Alibaba's Qwen3.8-27B raises SWE-Bench Pro (hard 51) pass@1 from 29.4% to 43.1% (51.0% with xhigh thinking) and Terminal-Bench 2.1 pass@1 from 44.7% to 58.8%.
-
Tencent 220B Hunyuan model runs on 8GB GPU via Maya quants
Maya v1.1.68 ships four self-made quantizations of Tencent's 220B Hy4-Flash, from Maya-XS at 45GB (88.5% top-token match) to Maya-L at 90GB (96.6%), enabling local inference on an 8GB graphics card.
-
Airbnb benchmarks Gemma 4 E2B on single RTX Pro 6000 GPU
Airbnb Engineering served Gemma 4 E2B for policy classification on one NVIDIA RTX Pro 6000 GPU, raising max F1 from 0.563 to 0.902 and hitting 35ms median latency (85ms p99) versus about 900ms for frontier APIs.