zwenai13/Zwen-Prime

TEXT GENERATIONConcurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:4kTool Calling:SupportedPublished:Jul 5, 2026License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Zwen-Prime is a 7 billion parameter full-stack engineering copilot developed by Zwen AI Labs, built on a DARE-TIES neural fusion of Mistral-7B-class specialists. It is fine-tuned for strict reasoning, Big-O discipline, native JSON function-calling, and zero-filler full-stack output across Python, TypeScript, React/Next.js, and Java. Optimized for local inference on Apple Silicon, it features a 32,000-token context window for extensive code analysis and generation.

Loading preview...

Zwen-Prime: An Elite Full-Stack Engineering Copilot

Zwen-Prime, developed by Zwen AI Labs, is a 7 billion parameter model engineered as a local, autonomous engineering copilot. It leverages a DARE-TIES neural fusion of three Mistral-7B-class specialists, followed by a supervised fine-tune on a unique five-corpus mixture. This two-stage process hard-installs strict <thinking> reasoning, Big-O discipline, native JSON function-calling, and zero-filler full-stack output.

Key Capabilities

  • Strict Reasoning Core: Every non-trivial answer includes a <thinking> block with step-by-step logic and a time/space Big-O breakdown, followed by raw deliverables without conversational filler.
  • Full-Stack Mastery: Proficient in typed Python, advanced TypeScript (generics, conditional types), React/Next.js (App Router, RSC, Server Actions), and Java concurrency primitives (ReentrantLock, CompletableFuture, virtual threads).
  • Native Function Calling: Emits raw JSON [TOOL_CALLS] matching provided tool schemas, adhering to Mistral v0.3 conventions.
  • Mathematical Reasoning: Derives quantitative claims with verified steps and sanity checks units/magnitudes.
  • Long-Context Discipline: Supports a 32,000-token context window, anchoring claims to source passages and refusing to confabulate.
  • Hardware Optimization: Designed for local inference on Apple Silicon (M-series) and consumer GPUs, requiring less than 5 GB RAM for the Q4_K_M GGUF quantization.

Good For

  • Designing and shipping production code across Python, TypeScript, React/Next.js, and Java.
  • Reasoning through complex math and algorithms with verifiable steps.
  • Integrating into agent runtimes for tool calling via raw JSON.
  • Developers seeking a direct, no-filler engineering assistant for local use.

Finetunes
1 models