GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
zwenai13
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
zwenai13
Model Tree
mistralai/Mistral-7B-Instruct-v0.3
Input Modalities
Output Modalities
Supported Functionality
| Creator | Zwen AI Labs |
| Architecture | Mistral-7B (DARE-TIES merge + LoRA SFT, fused) |
| Active params | ~7B |
| Context window | 8,192 tokens (32k-capable base) |
| Quantization | F16 → Q4_K_M GGUF (~4.2 GB, < 5 GB RAM target) |
| Hardware target | Apple Silicon (M-series, Metal-accelerated) |
| Serving | Ollama / llama.cpp |
| Languages | English |
Zwen-Prime is fine-tuned on a deliberate mixture of five datasets, each installing a distinct cognitive faculty:
| Corpus | Weight | Faculty installed |
|---|---|---|
| Alpaca Python | 40% | Elite, typed Python and algorithms at correct complexity. |
| Orca Math | 30% | Deep, step-by-step mathematical reasoning with verified derivation. |
| Salesforce XLAM | 20% | Flawless, raw-JSON function/tool calling (Mistral v0.3 convention). |
| LongAlpaca | 10% | Extended-context retention and long-document grounding without drift. |
| Zwen Custom (550 rows) | core spine | Strict logic + Big-O breakdown, full-stack mastery (TypeScript, React/Next.js, Java concurrency), absolute zero-filler output. |
The custom 550-row dataset (zwen_prime_master_dataset.jsonl) is the enforcement spine: 50 Identity rows that set the persona and 500 Algorithmic/Logic rows that lock in the <thinking>...</thinking> → raw-code template across advanced TypeScript generics, React/Next.js App-Router architecture, Java concurrency primitives, and Python system design.
<thinking> block with step-by-step logic and a time/space Big-O breakdown, immediately followed by the raw deliverable — no filler, no preamble.ReentrantLock, StampedLock, Semaphore, CompletableFuture, ForkJoinPool, virtual threads, happens-before reasoning.[TOOL_CALLS] raw JSON matching the provided tool schema; no markdown, no prose around the call.The base was produced with the DARE TIES merge method via mergekit, using Mistral-7B-Instruct-v0.3 as the density-0 reference base.
dphn/dolphin-2.9.3-mistral-7B-32k — reasoning specialisttheprint/ReWiz-7B — fine-tune specialistmistralai/Mistral-7B-Instruct-v0.3 — base / reference (density 0, weight 0)yaml
base_model: mistralai/Mistral-7B-Instruct-v0.3merge_method: dare_tiesdtype: bfloat16out_shard_size: 1.2Bparameters:density: 0.5weight: 1.0normalize: trueint8_mask: truerescale: truelambda: 1.0models:- model: dphn/dolphin-2.9.3-mistral-7B-32kparameters:density: 0.5weight: 1.0- model: theprint/ReWiz-7Bparameters:density: 0.5weight: 1.0- model: mistralai/Mistral-7B-Instruct-v0.3parameters:density: 0.0weight: 0.0
The merged base was then LoRA fine-tuned on the five-corpus mixture and the adapters permanently fused into the base weights (peft.merge_and_unload). The fused model is exported to GGUF for local serving.
text
DARE-TIES merge → LoRA SFT (5-corpus mixture) → merge_and_unload → GGUF → Ollama
The Ollama Modelfile lives at the repository root and references the F16 GGUF. To serve:
bash
ollama create zwen-prime -f Modelfileollama run zwen-prime
For the < 5 GB RAM Apple-Silicon target, quantize the F16 GGUF to Q4_K_M with llama.cpp first, then point FROM at the quantized file.
Zwen-Prime ships with a dedicated system prompt (model-workspace/Zwen_System_Prompt.txt) that sets the Principal-Engineer persona, the <thinking> mandate, the zero-filler output protocol, full-stack mastery expectations, native JSON tool-calling rules, long-context discipline, and the hard constraints. It is embedded in the Modelfile's SYSTEM directive.
Zwen-Prime is intended as a local, autonomous engineering copilot: designing and shipping production code across Python, TypeScript, React/Next.js, and Java; reasoning through math and algorithms with verifiable steps; and calling tools via raw JSON when integrated into an agent runtime.
[TOOL_CALLS].Apache-2.0, inheriting the licensing of the upstream Mistral-7B base and merged specialists. Verify against the terms of each upstream model before redistribution.
<thinking>