Model · v1

Zwen-Prime

A hand-fused 7B coding & reasoning model. Five expert corpora, merged with DARE-TIES into one coherent mind — built to run entirely on your machine.

Parameters
7B
Base architecture
Mistral v0.3
Fusion method
DARE-TIES
Quantization
Q4_K_M · ~4.07 GB
Native format
ChatML
Context window
32,000 tokens (32k)
Runtime
node-llama-cpp
Acceleration
Metal (Apple Silicon)
Architecture

DARE-TIES + 5-Corpus Mixture

Rather than training one model on everything, Zwen-Prime is fused from five specialist corpora. DARE-TIESselectively edits each expert's weights into the base, preserving what each corpus teaches while suppressing interference — so the fused model keeps every capability without the usual merge drift.

five corpora → DARE-TIES → Zwen-Prime
  1. 01

    Typed Python & algorithms

    Production-grade PEP 484/604, async-aware, concurrency-correct. Algorithms shipped at correct complexity.

  2. 02

    Step-by-step math

    Every quantitative claim derived in a thinking block first — algebra shown, units sanity-checked, edge cases covered.

  3. 03

    JSON function-calling

    Native Mistral v0.3 / XLAM convention. Raw, valid JSON — no markdown, no commentary, no invented parameters.

  4. 04

    Extended-context retention

    Tracks entities, state, and constraints established early through to the end. Surfaces contradictions explicitly.

  5. 05

    Full-stack discipline

    TypeScript generics, React/Next.js App Router, Java concurrency. Architecture over gadgets; server as source of truth.

ChatML-native. Despite the Mistral v0.3 chat template embedded in the GGUF, Zwen-Prime's fusion corpus is ChatML-formatted. The CLI pins ChatMLChatWrapper so it reasons cleanly and stops on turn boundaries — no run-on, no leaked framing.

Hardware

Runs on the hardware you already own

Three deployment targets, one model. From a MacBook Air to a multi-GPU data-center node — no proprietary cloud required.

Apple Silicon

Mac · Air & Pro

Chips
M1 · M2 · M3 · M4 · M5
Memory
Unified memory
Minimum
8 GB
Recommended
16 GB+ for full 32k
Acceleration
Metal (native)

Runs across the entire Apple Silicon lineup — MacBook Air and Pro alike. Unified memory means the weights and KV cache share one fast pool; no discrete GPU required.

Windows & Linux

Desktops & workstations

GPU
NVIDIA RTX 3060 / 4060
VRAM (GPU)
6–8 GB minimum
CPU fallback
16 GB system RAM
Compute capability
5.0+
Backends
CUDA · Vulkan

Dedicated-GPU path for throughput, with a CPU fallback for any host. A 6–8 GB card fits the Q4_K_M weights comfortably; pushing towards the full 32k context dynamically demands more VRAM.

Server & Enterprise

Data-center runtime

Topology
Multi-GPU nodes
VRAM
24 GB+ per GPU
Context
Full 32k concurrent
Throughput
Batched inference
Deployment
Licensed via Axora

For managed hosting and enterprise deployment at scale — batched inference across multiple GPUs holding the full 32k context for many concurrent sessions. Commercial use licensed through Axora AI Solutions.

Deploy seamlessly on the hardware you already own.

Quick start

Up and running in two commands

Install globally, run. The CLI fetches the Q4_K_M GGUF on first launch into ~/.zwen/models/ and boots the Metal-accelerated engine. No configuration, no API keys.

  • npm install -g @zwenailabs13/zwen-cliInstall the CLI globally.
  • zwen run zwen-primeOne-shot completion.
  • zwen chatInteractive REPL session.
  • zwen listList cached models in ~/.zwen/models/.
Options: --temperature, --max-tokens, --ctx, --json
zwen — bash
❯npm install -g @zwenailabs13/zwen-cli
✓ added 1 package — zwen-cli installed globally
❯zwen run zwen-prime
✓ Zwen-Prime 7B · Metal-accelerated · ready
▍generating…