---
license: apache-2.0
base_model: Qwen/Qwen3.5-27B
tags:
  - prism-coder
  - qwen3.5
  - function-calling
  - mcp
  - tool-routing
  - gguf
  - qlora
  - DeltaNet
language:
  - en
  - zh
---

# Prism Coder 27B — Qwen3.5-27B Function-Calling Model

Fine-tuned from [Qwen3.5-27B](https://huggingface.co/Qwen/Qwen3.5-27B) for MCP tool-routing.
Part of the [Prism Coder](https://github.com/dcostenco/prism-coder) fleet.

## Performance

| Metric | Value |
|--------|-------|
| **BFCL Accuracy** | **100% × 3 seeds** (345/345 test cases) |
| **Raw accuracy** | 100% (no L3 correction needed) |
| Tokens/sec (Q4_K_M, M5 48GB) | 28.5 |
| GGUF Q4_K_M size | 16 GB |
| Architecture | Hybrid DeltaNet (48/64 layers) + GQA (16/64) |
| Long context | O(n) via recurrent DeltaNet state |

## Training

| Parameter | Value |
|-----------|-------|
| Base model | Qwen/Qwen3.5-27B |
| Method | QLoRA (4-bit NF4) |
| LoRA rank | 128, alpha=256 |
| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| Layers | All 64 (including DeltaNet) |
| Training data | 24,798 examples (AAC 54%, tool-use 25%, safety 8%, abstention 8%) |
| Hardware | NVIDIA H100 PCIe 80GB |
| Duration | 12.5 hours |
| Final loss | 0.25 |
| Token accuracy | 93.2% |
| Cost | ~$29 |

## Fleet

| Tag | Size | BFCL | Role |
|-----|------|------|------|
| `prism-coder:2b` | 2.3 GB | 99.1% | Mobile / iPhone |
| `prism-coder:4b` | 3.4 GB | 100% | Verifier |
| `prism-coder:9b` | 5.8 GB | 100% | Default router |
| **`prism-coder:27b`** | **16 GB** | **100%** | **Quality tier** |

## Usage

```bash
ollama pull dcostenco/prism-coder:27b
ollama run dcostenco/prism-coder:27b "Load context for the analytics project"
```

Or via the Prism MCP server:
```json
{"mcpServers": {"prism": {"command": "npx", "args": ["-y", "prism-mcp-server"]}}}
```

## License

Apache 2.0 (same as base model)