Vision: these weights carry the Qwen3.5 vision tower (333 of 775 tensors). Until 2026-08-14 the repo omitted preprocessor_config.json, video_preprocessor_config.json and chat_template.jinja, so no converter could wire the tower up and every downstream build came out text-only. Those files are now published. Images are accepted; there is no audio encoder in this model family, and video is processed as frames through the vision tower.

prism-coder:9b — Prism Memory Tool Router (Default Tier)

QLoRA fine-tuned Qwen3.5-9B for MCP tool routing in the Prism Coder system. Replaces the previous 14B tier with 36% smaller footprint and higher accuracy.

Quick Start

ollama pull dcostenco/prism-coder:9b

BFCL Benchmark — 100% × 3 seeds

64/64 × 3 shuffled runs = 100.0%, 0 hallucinations

vs Previous 14B (Qwen 3)

Training

• Method: QLoRA (4-bit base + bf16 adapters) on Apple Silicon M5 48GB
• Base: mlx-community/Qwen3.5-9B-MLX-4bit for training, Qwen/Qwen3.5-9B bf16 for merge
• LoRA: rank=32, alpha=64, 16 layers
• Corpus: 26K rows (40% AAC / 12% abstention / 12% safety / 36% tool-use)
• Iterations: 2000 at LR 1e-4
• Layer 3: Inference-time rules for tool remapping, param normalization, multi-turn chain parsing

Architecture

Qwen3.5-9B uses a hybrid attention architecture:

• Linear attention layers (Gated DeltaNet) — O(n) inference, pattern matching
• Full attention layers (standard softmax) — precise retrieval and reasoning

Fleet Position

Links

• Ollama model page — pull and run
• Prism MCP Server — the MCP server
• Qwen3.5-9B base — upstream model