Instructions to use dcostenco/prism-coder-14b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use dcostenco/prism-coder-14b with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B")
model = PeftModel.from_pretrained(base_model, "dcostenco/prism-coder-14b")

llama-cpp-python

How to use dcostenco/prism-coder-14b with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="dcostenco/prism-coder-14b",
	filename="prism-aac-14b-q4km.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps

llama.cpp

How to use dcostenco/prism-coder-14b with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf dcostenco/prism-coder-14b
# Run inference directly in the terminal:
llama-cli -hf dcostenco/prism-coder-14b

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf dcostenco/prism-coder-14b
# Run inference directly in the terminal:
llama-cli -hf dcostenco/prism-coder-14b

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf dcostenco/prism-coder-14b
# Run inference directly in the terminal:
./llama-cli -hf dcostenco/prism-coder-14b

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf dcostenco/prism-coder-14b
# Run inference directly in the terminal:
./build/bin/llama-cli -hf dcostenco/prism-coder-14b

Use Docker

docker model run hf.co/dcostenco/prism-coder-14b

LM Studio
Jan

vLLM

How to use dcostenco/prism-coder-14b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "dcostenco/prism-coder-14b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dcostenco/prism-coder-14b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/dcostenco/prism-coder-14b

Ollama
How to use dcostenco/prism-coder-14b with Ollama:
```
ollama run hf.co/dcostenco/prism-coder-14b
```

Unsloth Studio new

How to use dcostenco/prism-coder-14b with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for dcostenco/prism-coder-14b to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for dcostenco/prism-coder-14b to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for dcostenco/prism-coder-14b to start chatting

Pi new

How to use dcostenco/prism-coder-14b with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf dcostenco/prism-coder-14b

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "dcostenco/prism-coder-14b"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use dcostenco/prism-coder-14b with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf dcostenco/prism-coder-14b

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default dcostenco/prism-coder-14b

Run Hermes

hermes

Docker Model Runner
How to use dcostenco/prism-coder-14b with Docker Model Runner:
```
docker model run hf.co/dcostenco/prism-coder-14b
```

Lemonade

How to use dcostenco/prism-coder-14b with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull dcostenco/prism-coder-14b

Run and chat with the model

lemonade run user.prism-coder-14b-{{QUANT_TAG}}

List all available models

lemonade list

prism-coder:14b (v26-polish) — 98% routing, ties Claude Opus 4.7

LoRA fine-tune of Qwen3-14B for offline MCP tool routing. Ties Claude Opus 4.7 at 98.0% ± 0.0% on the 100-case Prism eval (3-seed verified, zero variance). 3x faster (1.1s vs 3.0s), fully offline, zero cost per request.

Routing accuracy — 100-case Prism eval (May 15 2026, 3-seed mean)

Model	Accuracy	Cost/req	Latency
Claude Sonnet 4	99%	~$0.01	3.2s
prism-coder:14b	98.0% ± 0.0%	$0	1.1s
Claude Opus 4.7	98%	~$0.05	3.0s

Per-category (3-seed mean, zero variance):

Category	Score
Overall	98.0%
session_load_context	100%
session_save_ledger	100%
session_search_memory	100%
session_save_handoff	87%
session_compact_ledger	100%
brave_web_search	100%
knowledge_search	100%
AAC plain-text	100%
translate plain-text	100%
plain text	100%
no-tool refusal	100%
info / lookup	100%
edge (multi-step)	82%
avg latency	1.1s
invented tools	0

How it got to 98% — prompt engineering, zero retraining

The 14B went from 87% to 98% with zero retraining, zero GPU cost — purely system prompt changes:

v26 (+4 pts): -> plain text changed to -> respond directly (no tool). Q4_K_M models misread "plain text" as a tool name.
v27 (+7 pts): Labeled category headers added to routing rules:
```
CONVERSATION RECALL: what did we discuss / previously talked about -> session_search_memory
SAVED KNOWLEDGE: what do I know / stored notes / on file about -> knowledge_search
```
Labels act as semantic anchors stronger than keyword matching at Q4_K_M precision. knowledge_search jumped from 43% to 100%.

Offline cascade architecture

In the Prism AAC app, the 14B is the primary offline router:

Synalux cloud (Claude, 99%) → prism-coder:14b (98%, 1.1s) → prism-coder:1.7b (88%, iPhone fallback)

Training recipe (v26-polish)

Base: Qwen/Qwen3-14B (bf16)
LoRA: r=8, alpha=16, dropout 0.05, QKVO only
Corpus: 576 rows, 56% plain-text + 44% tool
Schedule: 50 iters, LR 1e-6, Mac M4 Max (MLX-LM), ~5 min
Note: the 87% to 98% improvement is from prompt engineering (v25→v27), not weight changes

Usage

ollama pull dcostenco/prism-coder:14b

Use the v27 system prompt with the nothink template. The 98% score requires both.

Hardware

Mac: M2 Pro+ / 24GB+ unified memory
Linux: RTX 3090/4090 (24GB)
VRAM: ~10 GB loaded

All Prism Coder models

Model	Accuracy	Size	Device	HuggingFace
prism-coder:14b	98%	8.4 GB	Mac / iPad Pro 16GB	dcostenco/prism-coder-14b
prism-coder:8b	96%	4.7 GB	iPhone / iPad 8GB	dcostenco/prism-coder-8b
prism-coder:32b	97.3%	19 GB	Mac M2 Ultra+	dcostenco/prism-coder-32b
prism-coder:1.7b	88%	2.2 GB	Any device / iPhone	dcostenco/prism-coder-1.7b

GitHub: dcostenco/prism-coder · AAC app: dcostenco/prism-aac · Portal: synalux.ai

Get the full stack

The model routes tool calls — but needs a backend to route TO:

# Install the memory server (free, local, no API keys)
npm install -g prism-mcp-server

# Pull the model
ollama pull dcostenco/prism-coder:14b

# Done — your AI agent now has persistent memory + 98% tool routing

Free tier: local SQLite, no cloud, no account needed. Synalux portal: cloud sync, HIPAA dashboard, team access, Claude fallback → synalux.ai

Prism Routing Benchmark

This model is evaluated on the Prism Routing Benchmark — a 100-case, 13-category eval for MCP tool routing. Run it yourself:

git clone https://github.com/dcostenco/prism-coder
cd prism-coder
python3 tests/benchmarks/prism-routing-100/benchmark.py --models 14b --seed 2027

Not a general function-calling benchmark (BFCL). This measures routing precision on 7 specific MCP tools — the task these models were built for. The value is offline reliability at zero cost, not competing with frontier models on arbitrary APIs.

License

Apache-2.0.

Downloads last month: 156

Safetensors

Model size

15B params

Tensor type

BF16

Model tree for dcostenco/prism-coder-14b

Base model

Qwen/Qwen3-14B-Base

Finetuned

Qwen/Qwen3-14B

Adapter

(210)

this model

Adapters

2 models