{"object":"list","count":5,"data":[{"id":"zyr-1","name":"ZYR 1","tagline":"Flash-fast everyday answers","speed":"Fast","parameters":"~7B (Gemini 4.7 Flash)","powered_by":"Google Gemini 4.7 Flash via Z.ai SDK","context_window":"32k tokens","pricing":"Free","best_for":["Quick factual questions","Simple math","Casual conversation","Greeting and small talk","Definitions and lookups"],"pipeline_stages":["Single direct call to Gemini 4.7 Flash"],"capabilities":["Sub-1-second responses","No thinking overhead","Low token usage","Streaming responses"],"limitations":["Not suitable for complex reasoning","Limited coding capability","No multi-step planning","No deep thinking"],"how_it_works":"ZYR 1 sends the user's message directly to Gemini 4.7 Flash with a minimal system prompt. No pipeline, no synthesis, no validation — just a fast single-shot response. Best for trivial queries where latency matters more than depth.","info_url":"https://zyrai.space-z.ai/zyr/zyr-1"},{"id":"zyr-2","name":"ZYR 2","tagline":"Balanced reasoning with GLM 5","speed":"Balanced","parameters":"~100B+ (GLM 5)","powered_by":"Z.ai GLM 5","context_window":"32k tokens","pricing":"Free","best_for":["Work and study questions","Longer conversations","Structured answers","General coding help","Writing assistance"],"pipeline_stages":["1. Planner — breaks the query into sub-tasks","2. Executor — generates the full response","3. Quality Evaluator — checks for gaps and polishes"],"capabilities":["3-stage reasoning pipeline","Better structured than ZYR 1","Good for medium-complexity tasks","Supports all chat modes (@agent, @plan, @code, etc.)"],"limitations":["Slower than ZYR 1 (~2-4s)","Not as deep as ZYR 3 for code","Not as powerful as ZYR 4 for hard problems"],"how_it_works":"ZYR 2 runs a 3-stage pipeline on GLM 5: Plan → Execute → Evaluate. The Planner decomposes the query, the Executor generates a complete answer, and the Quality Evaluator checks for gaps before returning. This makes responses more structured than ZYR 1 without the latency of ZYR 3.","info_url":"https://zyrai.space-z.ai/zyr/zyr-2"},{"id":"zyr-3","name":"ZYR 3","tagline":"Coding beast — 10-stage pipeline tuned for code","speed":"Max","parameters":"Multi-model ensemble (GLM 5.1 + GPT 4.4o + GPT 4o)","powered_by":"GLM 5.1 Max Deep Thinking + GPT 4.4o Mini + GPT 4o via Z.ai SDK","context_window":"32k tokens","pricing":"Free","best_for":["Production code generation","Debugging complex issues","Code refactoring","Architecture design","Algorithm implementation","Code review"],"pipeline_stages":["1. Router — classifies query complexity","2. Researcher — identifies relevant technologies","3. Planner — breaks down the implementation","4. Executor — generates the code","5. Synthesizer — merges outputs cleanly","6. Critic — reviews for bugs and security","7. Optimizer — improves readability and performance","8. Validator — fact-checks APIs and syntax","9. Quality Evaluator — final polish","10. Reviewer — last-pass review"],"capabilities":["10-stage pipeline tuned specifically for code","Multi-model ensemble (GLM + GPT 4.4o + GPT 4o)","Deep thinking enabled on critical stages","Auto-collapses to 3-stage fast path for simple queries","Production-ready output with type hints and error handling","Supports all coding modes (@code, @website, @blender, @web-design)"],"limitations":["Slower than ZYR 1/2 (~5-15s for complex code)","Can over-engineer simple tasks","Pipeline visible to user (can be noisy)"],"how_it_works":"ZYR 3 runs a 10-stage pipeline specifically tuned for code. Multi-model ensemble: GLM 5.1 Max Deep Thinking for reasoning, GPT 4.4o Mini for speed, GPT 4o for quality. Each stage has a specialized persona. The Router collapses to a 3-stage fast path for simple queries (under 80 chars, no complex keywords). Every code block passes through Critic + Optimizer + Validator for production-readiness.","info_url":"https://zyrai.space-z.ai/zyr/zyr-3"},{"id":"zyr-4","name":"ZYR 4 BEAST","tagline":"ZYR AI CORE v2 · 1T+ params · SOTA-tuned","speed":"Beast","parameters":"1T+ (10-model ensemble)","powered_by":"10 models in parallel: Nemotron Ultra 550B, Nemotron Super 120B, Nemotron Nano 30B, GPT-OSS 20B, Gemma 4 26B, Tencent Hy3, GLM, OpenZen Large, OpenZen Fast, OpenZen Reasoning","context_window":"200k+ effective (via compression)","pricing":"Free (all models on free tier)","best_for":["Hardest reasoning problems","Multi-step math (self-consistency voting)","Long-context analysis (200k+ effective)","Vision/image understanding","Writing that needs to beat GPT-4/Claude quality","Knowledge questions requiring web search","Any task where you want the best possible answer"],"pipeline_stages":["1. Smart Router — 4-tier dispatch (instant/fast/standard/deep) via regex","2. Context Compression — summarize old turns when history > 20k chars","3. Web Search — auto-triggered for current-events/knowledge queries","4. ZCG Execute — parallel draft generation across 10 models","5. GLM Synthesis — deep-thinking merge of all drafts","6. TCG Validation — code testing (only for code tasks)","7. Polish Pass — writer-editor persona for prose quality","8. Self-Consistency — 3-sample majority vote for math/reasoning"],"capabilities":["10-model ensemble with 1T+ total parameters","4-tier smart routing (instant: 458ms, fast: 248ms, standard: 6-17s, deep: 4s)","Self-consistency voting for math (3 GLM samples, majority vote)","Context compression for 200k+ effective window","Vision pass-through via GLM vision API","Web search integration via Z.ai SDK","Writer-quality polish pass (beats GPT-4 on prose)","TCG code validation only when needed (faster)","Backend-only ZYR AI CORE SYSTEM (ZCG + ACN + SL + CG + TCG)"],"limitations":["Slower than ZYR 1 for trivial queries (but router avoids this)","Can hit rate limits on OpenRouter free tier under heavy load","Vision limited to GLM's image capabilities (no video)","Web search adds 1-3s latency when triggered"],"how_it_works":"ZYR 4 BEAST v2 runs the ZYR AI CORE SYSTEM — a backend-only AI operating system. A regex-based Smart Router classifies the query into 4 tiers: instant (greetings, 1 model, <1s), fast (short Q&A, 1 model, <1s), standard (code/creative/knowledge, 10-model ensemble + polish, 6-17s), deep (math/reasoning, 3 GLM samples with self-consistency voting, ~4s). When the user asks about current events or facts, a web search runs via Z.ai SDK and results are injected into LLM context for grounded answers with citations. When conversation history exceeds 20k chars, old turns are summarized into a compact context note (effective context window: 200k+ tokens). For code tasks, the TCG (Testing Code Agent) validates generated code before returning. For all tasks, a final GLM deep-thinking polish pass with a writer-editor persona ensures GPT-4/Claude-quality prose. The 10-model ensemble has 1T+ total parameters: Nemotron Ultra 550B + Nemotron Super 120B + Nemotron Nano 30B + GPT-OSS 20B + Gemma 4 26B + Tencent Hy3 + GLM + 3 OpenZen models.","info_url":"https://zyrai.space-z.ai/zyr/zyr-4"},{"id":"zyr-voice","name":"ZYR Voice","tagline":"Natural TTS · powered by Z.ai","speed":"Balanced","parameters":"GLM 5 + Z.ai TTS (tongtong voice)","powered_by":"Z.ai GLM 5 + Z.ai TTS (24 kHz WAV)","context_window":"32k tokens","pricing":"Free","best_for":["Audio playback of responses","Accessibility (visually impaired users)","Hands-free consumption","Podcast-style content","Voice-first interfaces"],"pipeline_stages":["1. ZYR 2's 3-stage pipeline (Plan → Execute → Evaluate)","2. Markdown stripped for natural speech","3. Z.ai TTS synthesis (tongtong voice, 24 kHz WAV)"],"capabilities":["Natural-sounding Z.ai TTS voice (tongtong)","24 kHz WAV output at 16-bit mono PCM","Markdown stripping (no 'backtick backtick python' in audio)","Stop button during playback","Voice button on every AI message","Long-text chunking with rate-limit retry"],"limitations":["TTS adds 1-3s latency on first play","Voice quality limited to Z.ai's available voices","Code blocks read as '(code block)' instead of full content","Rate limits on Z.ai TTS API (sequential chunking with backoff)"],"how_it_works":"ZYR Voice uses ZYR 2's balanced pipeline to generate the text response, then passes it through the Z.ai TTS API (voice: tongtong, format: WAV, 24 kHz). Markdown formatting is stripped before synthesis (code fences become '(code block)', bold/italic markers removed, links reduced to text). Long text is split into 1024-char chunks and synthesized sequentially with exponential backoff on rate limits. WAV chunks are concatenated with patched headers into a single audio file. The voice button appears on every AI message across all ZYR models, not just ZYR Voice.","info_url":"https://zyrai.space-z.ai/zyr/zyr-voice"}]}