Hy3

Model capabilities and comparisons

• Many users find Hy3 “very impressive for its size,” often close to GPT‑5.4‑mini and Sonnet‑5, but not at GPT‑5.5 or GLM‑5.2 level.
• Some say GLM‑5.2 “easily matches” GPT‑5.4 and outperforms Hy3; others say Hy3’s coding benchmarks rival DeepSeek V4 Pro and MiMo v2.5 Pro despite lower pricing.
• Several compare it directly with DeepSeek V4 Flash/Pro: Hy3 is seen as more stable and better at staying on track, while DS4 Flash is faster but more temperamental and prone to building wrong mental models.
• Others report Hy3 underperforms smaller models like Gemma 4 31B, Qwen 3.6 27B, or Gemma MoE on tasks such as security auditing or “real work,” calling it overhyped.

Size, architecture, and local deployment

• Hy3 is 295B parameters, similar in scale to DS4 Flash (284B).
• It’s considered “shockingly small for how capable it is” in this class, but still large enough that local use generally requires multi‑GPU rigs (e.g., multi‑DGX Spark‑class hardware).
• Hy3 lacks DS4’s KV‑cache efficiency: even with FP4 quantization, available KV cache is far smaller than DS4 Flash’s, limiting long‑context practicality.
• Quantization resilience is unclear; DS4 Flash is repeatedly cited as strong even at 2‑bit, whereas Hy3’s behavior under heavy quantization is not yet well characterized.

Benchmarks and evaluation concerns

• Blog benchmarks impress some (Hy3 punching above its size/cost), but others highlight DeepSWE gaps vs GPT‑5.4 and suspect contamination or “benchmaxxing,” especially relative to Chinese models.
• One chart in the blog is criticized for inconsistent bar lengths vs numeric scores.
• Several emphasize that benchmarks may not match real‑world performance and urge task‑specific testing.

Pricing, access, and reliability

• Hy3 is praised as “exceedingly cheap” and Apache/MIT‑licensed, making it attractive as a FOSS option.
• Through OpenRouter, its effective price is reported similar to DeepSeek V4 Flash; a temporary free tier (until a specific date) drove experimentation.
• Some stopped using it due to aggressive rate limiting and HTTP errors; preview quality is said to be worse than the current release.

Use cases and behavior

• Users report good instruction following, engaging prose, strong world knowledge for its size, and solid small‑task performance.
• Others find it wastes time with confident but shallow answers compared to smaller models that feel “actually smart” within their knowledge limits.
• For long, agentic coding sessions or niche tech stacks, some still prefer Qwen or Gemma; Hy3’s advantage appears stronger in general coding and prose than in specialized domains like security auditing.

UI, documentation, and naming

• The official UI/demo is described as janky, with QR‑locked trial chat and confusing or missing front‑page explanation of what the product is.
• Some note mobile usability issues (blocked zoom, buggy image zoom) and oddities in demo interactions (e.g., pelican/SVG and animation quirks).
• The name “Hy3” disappoints some who expected a release of the unrelated “Hy” language.

Broader ecosystem context

• Hy3 is viewed within a crowded landscape: top‑tier closed models, strong mid‑tier proprietary models, and increasingly competitive open Chinese models (DeepSeek, GLM, Qwen, etc.).
• Opinions diverge on whether Hy3 has a clear niche given that some competitors are cheaper (DeepSeek V4) or perceived as better (GLM‑5.2, Gemma, Qwen); others see its license, cost, and balanced capabilities as compelling.