Instructions to use dcostenco/prism-coder-32b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use dcostenco/prism-coder-32b with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="dcostenco/prism-coder-32b",
	filename="prism-coder-32b-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-32b 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-32b
# Run inference directly in the terminal:
llama-cli -hf dcostenco/prism-coder-32b

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf dcostenco/prism-coder-32b
# Run inference directly in the terminal:
llama-cli -hf dcostenco/prism-coder-32b

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-32b
# Run inference directly in the terminal:
./llama-cli -hf dcostenco/prism-coder-32b

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-32b
# Run inference directly in the terminal:
./build/bin/llama-cli -hf dcostenco/prism-coder-32b

Use Docker

docker model run hf.co/dcostenco/prism-coder-32b

LM Studio
Jan

vLLM

How to use dcostenco/prism-coder-32b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "dcostenco/prism-coder-32b"
# 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-32b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/dcostenco/prism-coder-32b

Ollama
How to use dcostenco/prism-coder-32b with Ollama:
```
ollama run hf.co/dcostenco/prism-coder-32b
```

Unsloth Studio

How to use dcostenco/prism-coder-32b 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-32b 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-32b 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-32b to start chatting

How to use dcostenco/prism-coder-32b 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-32b

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-32b"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use dcostenco/prism-coder-32b 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-32b

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-32b

Run Hermes

hermes

Docker Model Runner
How to use dcostenco/prism-coder-32b with Docker Model Runner:
```
docker model run hf.co/dcostenco/prism-coder-32b
```

Lemonade

How to use dcostenco/prism-coder-32b with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull dcostenco/prism-coder-32b

Run and chat with the model

lemonade run user.prism-coder-32b-{{QUANT_TAG}}

List all available models

lemonade list

dcostenco commited on 14 days ago

Commit

4b2a4da

verified ·

1 Parent(s): dbe2797

Update model card: v5 MoE 97.1%, full cascade results, version history

Browse files

Files changed (1) hide show

README.md +33 -19

README.md CHANGED Viewed

@@ -1,23 +1,27 @@
 ---
 language: en
 license: apache-2.0
-base_model: Qwen/Qwen3-32B
 tags:
   - tool-calling
   - routing
   - aac
   - qwen3
   - gguf
 ---
 # prism-coder:32b — Tool Routing Model (Desktop Quality Tier)
-Fine-tuned Qwen3-32B for 6-tool routing in the [Prism AAC](https://github.com/dcostenco/prism-aac) system.
 Quality escalation tier in the desktop cascade: **14B → 32B → cloud Claude**.
-## BFCL Routing Benchmark — v33 (Current)
-**Mean: 99.0%** (3-seed average, seeds 2027/2028/2029, 102 cases each)
 | Category | Description | Accuracy |
 |----------|-------------|:--------:|
@@ -31,19 +35,30 @@ Quality escalation tier in the desktop cascade: **14B → 32B → cloud Claude**
 | load | Session context loading | 100% |
 | pred | Factual / knowledge queries → plain text | 100% |
 | save | Session ledger save | 100% |
-| smem | Session memory search | 92% |
 | tran | Translation requests → plain text | 100% |
 Eval: Ollama inference, temperature=0, Qwen3 thinking suppressed (`<think>\n\n</think>`), num_predict=160.
 Gate: ≥90% = deploy.
 ## Version History
-| Version | BFCL | Notes |
-|---------|------|-------|
-| v33 | 99.0% | Bf16 merge fix — corrects tran boundary (was 83% → now 100%) |
-| v31 | 98.0% | tran=83%, smem=92% — previous deployment |
-| v30 | ~96% | Routing corpus v30 |
 ## Tools
@@ -60,12 +75,9 @@ The model routes between exactly 6 tools:
 | File | Size | Use |
 |------|------|-----|
-| `qwen3-32b-v33-q6k.gguf` | 25 GB | Recommended (higher quality) |
-## Cascade Role
-Quality escalation tier. Invoked when 14B has low confidence or fails.
-Handles complex multi-step requests and edge cases before escalating to cloud Claude.
 ## Usage (Ollama)
@@ -75,7 +87,9 @@ ollama run dcostenco/prism-coder:32b
 ## Training
-- **Base**: `mlx-community/Qwen3-32B-bf16` (bf16, 32.8B params, 13 safetensors shards)
-- **Adapters**: v32 LoRA (rank=8, scale=20, 4 layers)
-- **Merge**: Direct safetensors manipulation (delta = scale/rank × B^T A^T)
 - **Hardware**: Apple Silicon (M-series, 64 GB RAM)

 ---
 language: en
 license: apache-2.0
+base_model: Qwen/Qwen3-30B-A3B
 tags:
   - tool-calling
   - routing
   - aac
   - qwen3
+  - moe
   - gguf
 ---
 # prism-coder:32b — Tool Routing Model (Desktop Quality Tier)
+Fine-tuned Qwen3-30B-A3B (MoE) for 6-tool routing in the [Prism AAC](https://github.com/dcostenco/prism-aac) system.
 Quality escalation tier in the desktop cascade: **14B → 32B → cloud Claude**.
+> **v5 (May 2026)**: Switched base from dense Qwen3-32B to Qwen3-30B-A3B (MoE).
+> Same accuracy, 9 GB smaller, ~4× faster inference (only ~3B params active per token).
+## BFCL Routing Benchmark — v5 MoE (Current)
+**Mean: 97.1%** (3-seed average, seeds 2027/2028/2029, 102 cases each)
 | Category | Description | Accuracy |
 |----------|-------------|:--------:|
 | load | Session context loading | 100% |
 | pred | Factual / knowledge queries → plain text | 100% |
 | save | Session ledger save | 100% |
+| smem | Session memory search | 100% |
 | tran | Translation requests → plain text | 100% |
 Eval: Ollama inference, temperature=0, Qwen3 thinking suppressed (`<think>\n\n</think>`), num_predict=160.
 Gate: ≥90% = deploy.
+## Full Cascade Benchmark (May 2026)
+| Model | BFCL | Size | Latency | Tier |
+|-------|------|------|---------|------|
+| prism-coder:8b v35 | **98.0%** | 4.7 GB | ~0.8s | Mobile tier 2 |
+| prism-coder:32b v5 MoE | **97.1%** | 17 GB | ~0.8s | Desktop tier 2 |
+| prism-coder:14b v33 | **97.1%** | 9.3 GB | ~1.1s | Desktop tier 1 |
+| prism-coder:1b7 v41 | **94.1%** | 1.1 GB | ~0.5s | On-device |
 ## Version History
+| Version | Base | BFCL | Notes |
+|---------|------|------|-------|
+| v5 (current) | Qwen3-30B-A3B MoE | **97.1%** | 18x density fix on all 8 failing cases; 9GB smaller, 4× faster |
+| v4 | Qwen3-30B-A3B MoE | 92.2% | rank=32 experiment — regressed vs v3 |
+| v3 | Qwen3-30B-A3B MoE | 92.5% | 20x reps + LR=1e-5 — hit rank bottleneck |
+| v2 | Qwen3-30B-A3B MoE | 92.5% | v34 corpus + 1400 iters |
+| v33 (dense) | Qwen3-32B dense | 99.0% | Prior generation — larger/slower |
 ## Tools
 | File | Size | Use |
 |------|------|-----|
+| `qwen3-30b-a3b-v5-iq4nl.gguf` | 17 GB | **Current — recommended** |
+| `qwen3-30b-a3b-v4-iq4nl.gguf` | 17 GB | Previous (92.2%) |
+| `qwen3-32b-v33-q6k.gguf` | 25 GB | Dense predecessor (99.0%, legacy) |
 ## Usage (Ollama)
 ## Training
+- **Base**: `mlx-community/Qwen3-30B-A3B-4bit` (MoE, ~3B active params/token, 128 experts)
+- **Adapters**: v5 LoRA (rank=8, scale=20, 8 layers, LR=1e-5, 800 iters)
+- **Data**: v36 corpus — 615 train examples, 18× density on all 8 exact failing prompts
+- **Merge**: Direct safetensors manipulation (attn/gate: delta = scale/rank × B^T A^T; experts: delta[i] = scale/rank × B[i] A[i])
 - **Hardware**: Apple Silicon (M-series, 64 GB RAM)
+- **Key insight**: MoE ceiling at 92.5% was data density (1-3 reps per failing case); fixed with 18× reps matching the 32B v32→99% approach