File size: 3,179 Bytes
1a4770d 443f43a 1a4770d 395cd09 443f43a 395cd09 1a4770d 229da84 1a4770d 229da84 1a4770d e445ce1 229da84 9ae524d 229da84 e445ce1 229da84 e445ce1 1a4770d 395cd09 1a4770d e445ce1 229da84 e445ce1 1a4770d e445ce1 229da84 e445ce1 229da84 e445ce1 ab8afb3 e445ce1 ab8afb3 395cd09 e445ce1 395cd09 ab8afb3 e445ce1 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 | ---
language: en
license: apache-2.0
tags:
- tool-routing
- function-calling
- prism-aac
- qwen3
- gguf
base_model: Qwen/Qwen3-8B
---
# prism-coder:8b β Tool Routing Model (iOS / Edge Tier)
Fine-tuned Qwen3-8B for 6-tool routing in the [Prism AAC](https://github.com/dcostenco/prism-aac) system.
Primary deployment: **iOS and edge devices** via llama.cpp GGUF.
## BFCL Routing Benchmark β v36 (Current)
**Mean: 100.0%** (3-seed average, seeds 2027/2028/2029, 102 cases each)
| Category | Count | Description | Accuracy |
|----------|------:|-------------|:--------:|
| aac | 12 | AAC phrase requests β plain text | 100% |
| cmpct | 6 | Ledger compaction | 100% |
| edge | 6 | Multi-step / compound requests | 100% |
| hand | 8 | Agent handoff / relay | 100% |
| info | 5 | General facts β plain text | 100% |
| irrel | 10 | Irrelevant / live queries β plain text | 100% |
| know | 7 | Knowledge base search | 100% |
| load | 9 | Session context loading | 100% |
| pred | 8 | Factual / knowledge queries β plain text | 100% |
| save | 13 | Session ledger save | 100% |
| smem | 12 | Session memory search | 100% |
| tran | 6 | Translation requests β plain text | 100% |
Eval: MLX inference + thinking, temperature=0, 3-seed mean.
Gate: β₯90% = deploy.
## Cascade Benchmark (May 2026)
Full desktop cascade: **14b β 32b β Claude Opus** (102 cases Γ 3 seeds)
| Metric | Result |
|--------|--------|
| Cascade accuracy | **100.0%** (mean, 3 seeds) |
| Opus-solo etalon | 98.3% |
| Ξ vs Opus | **+1.7%** |
| Traffic served by 14b | **99%** (101/102 cases avg) |
| Traffic escalated to 32b | 1% (1/102 avg) |
| Traffic reaching Opus API | **0%** |
Fine-tuned cascade outperforms Claude Opus on `edge` (+16.7%) and `know` (+14.3%).
## Version History
| Version | BFCL | Notes |
|---------|------|-------|
| v36 | **100.0%** | Fixed: smem "BFCL v4 notes" and "training loss" β session_search_memory |
| v35 | 98.0% | Proper safetensors merge β fixes mlx_lm.fuse LoRA loss |
| v32 | 98.0% | Routing corpus v32_8b, direct safetensors merge |
| v31 | 95.1% | Surgical smem/know boundary fix |
| v30 | ~93% | Baseline 8B routing |
## Tools
The model routes to exactly 6 tools:
| Tool | Trigger |
|------|---------|
| `session_load_context` | Load/resume project context |
| `session_save_ledger` | Note/log/record/remember something |
| `session_save_handoff` | Pass state to next agent/session |
| `session_compact_ledger` | Shrink/prune ledger (no relay) |
| `session_search_memory` | Recall prior session discussions |
| `knowledge_search` | Search stored knowledge base |
Plain text (no tool) for: AAC phrases, translations, weather, general facts, code, math.
## Model Details
- **Base**: Qwen/Qwen3-8B
- **Format**: GGUF Q4_K_M (~4.9 GB)
- **Context**: 32,768 tokens
- **Training**: MLX LoRA, rank=16, 16 layers, 1000 iters, LR=2e-6, v36 corpus (806 examples)
- **Merge**: mlx_lm.fuse β llama.cpp convert β Q4_K_M quantization
## Usage
```bash
ollama pull dcostenco/prism-coder-8b
ollama run prism-coder:8b
```
Or in the [Prism Coder IDE](https://github.com/dcostenco/prism-aac) β set model to `prism-coder:8b` in Settings.
|