dcostenco/ prism-coder:2b

680 1 week ago

Local-first AI tool router. 2B/4B/9B read images (vision); 27B text-only. 14B/32B retired. 99.1-100% routing accuracy (BFCL). 97% of traffic stays local.

vision tools thinking 2b 4b 9b 14b 27b 32b 
ollama run dcostenco/prism-coder:2b

Details

1 week ago

33cd2348ec0e · 3.3GB ·

model
qwen35
·
4.33B
·
Q4_K_S
projector
clip
·
334M
·
F16
system
CRITICAL: You have EXACTLY 6 tools. Their EXACT names are: session_load_context, session_save_ledger
license
Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ TERMS AND CONDITIONS FOR US
params
{ "num_ctx": 32768, "num_predict": 512, "presence_penalty": 1.5, "stop": [ "

Readme

Prism Coder — Local-First AI Agent Tool Router

Qwen3.5-based models for AI agent tool routing. 97% of queries handled locally — no cloud needed.

99.1–100% routing accuracy across sizes (BFCL eval, 115 cases × 3 seeds).

Models

Model Size Input Use when
prism-coder:2b 3.3 GB Text, Image iPhone / mobile first gate (Qwen3.5-4B Q4_K_S)
prism-coder:4b 3.5 GB Text, Image Verifier + 8 GB+ devices (Qwen3.5-4B Q4_K_M)
prism-coder:9b 6.7 GB Text, Image Default router (fine-tuned Qwen3.5-9B)
prism-coder:27b 17 GB Text Complex tasks (QLoRA on Qwen3.5-27B)

14b and 32b are retired — earlier Qwen3-based builds, kept only so existing pulls keep working. Do not start new work on them.

Vision

2b, 4b and 9b carry the Qwen3.5 vision tower and accept screenshots and video frames. 27b is text-only.

There is no audio encoder in this model family — video is processed as frames through the vision tower, and sound is not supported at any size.

Quick Start

# Recommended default (sees images)
ollama run dcostenco/prism-coder:9b

# iPhone / mobile
ollama pull dcostenco/prism-coder:2b

# Verifier
ollama pull dcostenco/prism-coder:4b

Cascade

2B on-device   iPhone first gate            (99.1%)
4B verifier    grounding checks, 8 GB+      (100%)
9B local       default router, reads images (100%)
27B local      complex tasks, text only     (100%)
cloud          ~1% — last resort

97% of traffic never touches the cloud.

Training

  • 2B/4B: stock Qwen3.5-4B with system-prompt engineering (no fine-tuning), Q4_K_S and Q4_K_M quantizations.
  • 9B: fine-tuned from Qwen3.5-9B (MLX LoRA + direct safetensors merge).
  • 27B: QLoRA (4-bit NF4, rank 128) on Qwen3.5-27B.
  • Eval gate: 99.1–100% on BFCL (115 cases × 3 seeds) before deploy.

Full Weights

HuggingFace: huggingface.co/dcostenco — prism-coder-9b publishes full safetensors including the vision tower and its preprocessor configs.

Built For

  • Prism AAC — communication app for non-verbal users
  • Prism Coder — AI dev assistant with persistent memory