SHELFCRITTER AI Digest

MODEL

Tencent Hy4

modeltopic-noteopen-sourcechinatencentmoe

Overview

Tencent Hy4 Preview is Tencent‘s August 2026 frontier Mixture-of-Experts release — 770B total parameters with 49B active per token, a native 1M-token context window, and an Apache 2.0 license. Distributed on both Hugging Face and OpenRouter at $0.83/M input / $2.50/M output — roughly one-fifth of comparable frontier-tier US closed models. Tencent’s own launch framing compares Hy4 against agentic-parallel-research setups running Codex rather than the Bloomberg-headline claim of “outperforming Z.AI and Moonshot”; the vendor-cited comparator is the load-bearing detail. Extends the Chinese open-weights cadence Hy3 opened in July 2026 into the frontier-scale MoE tier at a permissive-license price band that puts pressure on the API-only mid-tier.

Timeline

  • 2026-08-29-AI-Digest — Tencent releases Hy4 Preview: 770B / 49B-active MoE, native 1M-token context, Apache 2.0, on Hugging Face and OpenRouter at $0.83/M in / $2.50/M out. Tencent’s own benchmark framing compares Hy4 against agentic-parallel-research setups running Codex rather than the Bloomberg headline “outperforming Z.AI and Moonshot” — the vendor comparator is load-bearing. Narrow read the digest carries: specs (770B/49B/1M) are confirmed on the HF card and OpenRouter listing; pricing is roughly one-fifth of comparable frontier-tier US closed models; trust Tencent’s stated Codex comparator, not Bloomberg’s editorial Z.AI/Moonshot framing. Structural read: do NOT extend the running “death zone for mid-tier US model makers” narrative from Bloomberg without pressure-testing — Databricks just posted >80% YoY at $7B run-rate and Cohere is trending toward IPO at $240M ARR, contradicting the sector-squeeze framing on August numbers. Disciplined framing is price pressure on API-only mid-tier plays (where DeepSeek and Qwen already sit), not a sector-wide squeeze.

  • 2026-08-30-AI-Digest — Hy4 Preview surfaces as reference-only flagship-lite comparator in the Alibaba / Qwen 3.8-Flash Bloomberg-launch story — the digest’s Structural read names Hy4 Preview’s $0.83 / $2.50 flagship-lite tier as one of three credible open-weight-adjacent options crowding the same pricing band alongside DeepSeek V4-Flash ($0.14 / $0.28) and Qwen 3.8-Flash ($0.16 / $0.47). Narrow read the digest carries: no fresh Hy4 product action today; log as flagship-lite comparator anchor in the crowded-Flash-tier pricing-band frame — the differentiator for practitioners choosing an API in the Flash tier is now capability profile and licence, not price. Structural read: Hy4’s Apache-2.0 permissive-license positioning (2026-08-29-AI-Digest) means the three-option Flash-tier crowd is measured on licence-and-capability, not on the last basis point of per-token spend.

Key Developments

  1. Frontier-Scale MoE at Apache 2.0 with 1M-Token Native Context (August 29, 2026): Hy4 Preview lands at 770B total / 49B active per token with a native 1M-token context window under Apache 2.0, distributed on both Hugging Face and OpenRouter at $0.83/M input / $2.50/M output — roughly one-fifth of comparable frontier-tier US closed models. Tencent’s own comparator is agentic-parallel-research setups running Codex; Bloomberg’s “outperforming Z.AI and Moonshot” is editorial, not the lab’s. Extends Tencent‘s Chinese open-weights cadence from Hy3 (July 2026, 295B / 21B-active) into the frontier-scale MoE tier at a permissive-license price band. Load-bearing structural read: price pressure on API-only mid-tier plays, not a sector-wide squeeze — Databricks and Cohere August numbers contradict the “death zone” framing.

See also: Tencent, Hugging Face, OpenRouter, Codex, Z.ai, Moonshot AI, MOC - Open Source Models, MOC - Major Companies.