The mysterious Hy3 LLM is topping OpenRouter Model Rankings by a large margin

Pricing, Caching, and OpenRouter Economics

• Hy3 is popular partly because it’s very cheap on OpenRouter and “good enough” in quality for many workloads.
• DeepSeek V4 Flash can be even cheaper if used via DeepSeek’s own API due to stronger caching discounts; using it through OpenRouter can lose significant savings.
• Within OpenRouter, effective price varies by provider due to different input/output/cache pricing, so “cheapest” depends on workload mix.
• There’s confusion over why some providers are cheaper than DeepSeek itself; one explanation is that cheaper tiers use lower-precision quantization (fp4/fp8) while DeepSeek’s own endpoint may be unquantized.
• OpenRouter also offers a “Response Caching” feature for identical requests at no cost, though some question how often identical requests occur in practice.

Usage Rankings and Market Signal Quality

• OpenRouter leaderboards reflect only traffic through OpenRouter, so they underrepresent models where many users go direct (e.g., Anthropic).
• Some argue the rankings are still a valid market signal; others emphasize that a single large app (“whale”) can dominate tokens, making it hard to infer broad user preference.
• Requests for better metrics include unique-user indicators, spend-based charts, and token volume over time.
• It’s noted that a single SaaS product can burn billions of tokens per day, reinforcing that token volume ≠ number of users.

Hy3 Popularity, Performance, and “Mystery”

• Hy3 is from Tencent, so several commenters question why it’s labeled “mysterious”; others suggest the mystery is more about its sudden leaderboard dominance and deployment details.
• One commenter reports multi-benchmark testing where Hy3 ranks mid-pack among ~25–80 models, suggesting it’s not top-tier but competitive.
• Another user found it poor for coding/agentic tasks compared to newer frontier models and even strong local models, citing thought loops and undesired edits.
• Others find it attractive as a post-training base: good factual knowledge, with issues more on agent behavior than core capabilities.
• A link notes Hy3 started as a >400B parameter model later reduced to ~295B as an “optimal zone.”

Free Access and Usage Spikes

• Debate over whether Hy3’s ranking is inflated by free access.
• Some insist usage stayed high after the free period; others point out it’s still free in various apps and via a “free” OpenRouter endpoint, so the true paid-only demand is unclear.

Data Privacy, Trust, and Jurisdiction Concerns

• Some are wary of sending sensitive business data to an LLM whose developers or jurisdiction they don’t fully trust, especially with state-linked entities.
• Others argue what matters is the inference provider’s data-handling policies (e.g., zero data retention, opt-out of training), which OpenRouter claims to control via contracts.
• Skeptics question how such contracts can be verified or enforced and note that both model developers and hosts are potential threat vectors.
• There is broader unease about major US and Chinese AI companies’ relationships with their respective governments and the opacity around data use.