Instructions to use dcostenco/prism-coder-1.7b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use dcostenco/prism-coder-1.7b with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="dcostenco/prism-coder-1.7b",
	filename="prism-aac-1b7-q4km.gguf",
)

llm.create_chat_completion(
	messages = "No input example has been defined for this model task."
)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use dcostenco/prism-coder-1.7b 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-1.7b:Q8_0
# Run inference directly in the terminal:
llama-cli -hf dcostenco/prism-coder-1.7b:Q8_0

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf dcostenco/prism-coder-1.7b:Q8_0
# Run inference directly in the terminal:
llama-cli -hf dcostenco/prism-coder-1.7b:Q8_0

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-1.7b:Q8_0
# Run inference directly in the terminal:
./llama-cli -hf dcostenco/prism-coder-1.7b:Q8_0

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-1.7b:Q8_0
# Run inference directly in the terminal:
./build/bin/llama-cli -hf dcostenco/prism-coder-1.7b:Q8_0

Use Docker

docker model run hf.co/dcostenco/prism-coder-1.7b:Q8_0

LM Studio
Jan
Ollama
How to use dcostenco/prism-coder-1.7b with Ollama:
```
ollama run hf.co/dcostenco/prism-coder-1.7b:Q8_0
```

Unsloth Studio

How to use dcostenco/prism-coder-1.7b 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-1.7b 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-1.7b 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-1.7b to start chatting

How to use dcostenco/prism-coder-1.7b 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-1.7b:Q8_0

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-1.7b:Q8_0"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use dcostenco/prism-coder-1.7b 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-1.7b:Q8_0

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-1.7b:Q8_0

Run Hermes

hermes

Docker Model Runner
How to use dcostenco/prism-coder-1.7b with Docker Model Runner:
```
docker model run hf.co/dcostenco/prism-coder-1.7b:Q8_0
```

Lemonade

How to use dcostenco/prism-coder-1.7b with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull dcostenco/prism-coder-1.7b:Q8_0

Run and chat with the model

lemonade run user.prism-coder-1.7b-Q8_0

List all available models

lemonade list

dcostenco commited on 21 days ago

Commit

f7a3fe7

verified ·

1 Parent(s): 04be453

docs: updated benchmark scores — v26 system prompt + nothink template (May 14 2026)

Browse files

Files changed (1) hide show

README.md +35 -54

README.md CHANGED Viewed

@@ -19,57 +19,51 @@ On-device MCP tool router based on **Qwen/Qwen3-1.7B**. Runs in 1.6 GB RAM at Q4
 ## Test results — Prism routing 100-case eval (May 14 2026)
-3-run mean across seeds 2027/2028/2029 — variance was essentially zero.
-| Category | Score |
-|---|---|
-| **Overall** | **84.0% ± 0.0%** |
-| session_load_context | 100% |
-| session_save_ledger | 76% |
-| session_search_memory | 78% |
-| session_save_handoff | 74% |
-| session_compact_ledger | 100% |
-| brave_web_search | 100% |
-| knowledge_search | 71% |
-| AAC plain-text | 91% |
-| translate plain-text | 100% |
-| static facts (pred) | 88% |
-| live-info refusal | 83% |
-| info / lookup | 60% |
-| edge (multi-step) | 65% |
-| **avg latency** | 1.6 s |
-| **invented tools** | 2 (across 300 cases) |
-**Below the 90% gate** — this model is published for the **on-device / cost-sensitive** use case, not for accuracy-critical work. For production-quality routing, use the 14B (90%) or 32B (93.7%).
-**What this benchmark measures**: routing precision against the *exact* 7-tool Prism Coder taxonomy. **Not** a general-capability score. Not comparable to public leaderboards. Methodology + runner: [github.com/dcostenco/prism-coder/tree/main/tests/benchmarks/prism-routing-100](https://github.com/dcostenco/prism-coder/tree/main/tests/benchmarks/prism-routing-100).
 **Where this model wins**:
-- **Fits on an iPhone** (1.6 GB Q4_K_M runtime — every other Prism Coder tier requires a Mac with ≥24 GB)
-- **Fastest** of any Prism Coder model — 1.6 s average response
-- **100% on 3 critical categories**: session_load_context, session_compact_ledger, web_search, translate
 - Zero cloud, zero per-token cost, fully offline
-**Where it underperforms** (vs 14B at 90% / 32B at 93.7% / Claude Sonnet at 99%):
-- `info` — 60%. Small models struggle to distinguish "what does X mean" (lookup) from "what is X" (static fact).
-- `edge` — 65%. Multi-step routing decisions are too long-tailed for a 1.7B model.
-- `save` / `smem` / `hand` / `know` — all in the 70s. The exact session-memory tool boundary is subtle at this size.
-- **May at capacity ceiling** — a v25-max retrain attempt regressed gate scores. Small models have inherent precision limits for routing taxonomies this granular.
-For production: this model is the "free tier" fallback. Pair with cloud Claude for accuracy-critical paths. The [Synalux router](https://github.com/dcostenco/prism-coder) handles the cascade automatically.
 ## Use case: iOS deployment
 This is the model the [Prism AAC](https://github.com/dcostenco/prism-aac) iOS / iPadOS / watchOS apps load on-device via **llama.cpp Swift SPM**. CoreML was evaluated and rejected — `coremltools` doesn't support Qwen3 attention ops.
-GGUF available: `prism-aac-1b7-q4km.gguf` (1.0 GB on disk, ~1.6 GB RAM at runtime). Fallback path: Mac Ollama over local WiFi via `OLLAMA_HOST=0.0.0.0`.
 ## Training recipe (v19)
-- **Base**: Qwen/Qwen3-1.7B + v19 system-prompt engineering
-- **Note**: The published `:1b7-v19-q8` Ollama tag uses **system-prompt engineering only** (no LoRA fine-tune). This was found to be more reliable than LoRA fine-tuning at this scale — small models over-fit aggressively (v25-max regressed 100% → 93.8% on a 16-case gate test).
-- **Quantization**: Q4_K_M (1.0 GB) for iOS, Q8 (2.2 GB) for Mac / better quality
-- **Hardware**: Mac M4 Max (MLX-LM) for the prompt-engineering eval cycle
 ## Usage
@@ -80,26 +74,13 @@ ollama pull dcostenco/prism-coder:1b7
 ollama run dcostenco/prism-coder:1b7 "Search my knowledge for OAuth scope reference"
 ```
-### iOS — llama.cpp Swift SPM
-See [prism-aac/ios-native](https://github.com/dcostenco/prism-aac/tree/main/ios-native) for the full integration: `LLMEngine.swift` → `prismNativeBridge.askAI()` → `window.prismNativeAIResult()` token stream.
-### HuggingFace (transformers)
-```python
-from transformers import AutoModelForCausalLM, AutoTokenizer
-model = AutoModelForCausalLM.from_pretrained("dcostenco/prism-coder-1.7b", torch_dtype="auto")
-tok = AutoTokenizer.from_pretrained("dcostenco/prism-coder-1.7b")
-```
 ### System prompt
-Use the [v25 routing prompt](https://github.com/dcostenco/prism-coder/blob/main/tests/benchmarks/prism-routing-100/benchmark.py#L47) verbatim. This model relies *heavily* on the system prompt — see "Training recipe" above.
 ## Hardware requirements
 - **iPhone / iPad**: A14 Bionic+ (iPhone 12+), iOS 15+, ~1.6 GB free RAM
-- **Apple Watch (standalone)**: Series 10+ for on-device, otherwise paired-mode via iPhone
 - **Mac**: any M-series (M1 8 GB unified is enough at Q4_K_M)
 - **Linux**: any CPU works at Q4_K_M; a 4 GB GPU dramatically speeds inference
 - **Inference speed**: ~0.5–1.6 s per 200-token response

 ## Test results — Prism routing 100-case eval (May 14 2026)
+100 prompts (seed=2027), v26 system prompt.
+| Category | Current | Previous (v19 old prompt) | Δ |
+|---|---|---|---|
+| **Overall** | **88%** | 84.0% | **+4.0** |
+| session_load_context | 100% | 100% | = |
+| session_save_ledger | 85% | 76% | +9 |
+| session_search_memory | 100% | 78% | +22 |
+| session_save_handoff | 75% | 74% | +1 |
+| session_compact_ledger | 100% | 100% | = |
+| brave_web_search | 86% | 100% | -14 |
+| knowledge_search | 43% | 71% | -28 |
+| AAC plain-text | **100%** | 91% | **+9** |
+| translate plain-text | 100% | 100% | = |
+| plain text (pred/irrel) | 100% | 88% | +12 |
+| no-tool refusal | 83% | 83% | = |
+| info / lookup | 100% | 60% | +40 |
+| edge (multi-step) | 60% | 65% | -5 |
+| **avg latency** | **1.6s** | 1.6s | = |
+| **invented tools** | **0** | 2 | **-2** |
+**Key improvements**: AAC routing 91% → 100% (life-critical fix), zero invented tool names (was 2), info/lookup 60% → 100%, session_search_memory 78% → 100%.
+**Below the 90% gate** — this model is published for the **on-device / cost-sensitive** use case, not for accuracy-critical work. The bottleneck is `knowledge_search` at 43% (same as 14B).
+**What this benchmark measures**: routing precision against the *exact* 7-tool Prism Coder taxonomy. **Not** a general-capability score. Methodology + runner: [github.com/dcostenco/prism-coder/tree/main/tests/benchmarks/prism-routing-100](https://github.com/dcostenco/prism-coder/tree/main/tests/benchmarks/prism-routing-100).
 **Where this model wins**:
+- **Fits on an iPhone** (1.6 GB Q4_K_M runtime)
+- **Fastest** of any Prism Coder model — 1.6 s average
+- **100% on 6 categories**: load_ctx, srch_mem, compact, translate, AAC, plain_text
+- Zero invented tool names (improved from 2 in previous version)
 - Zero cloud, zero per-token cost, fully offline
 ## Use case: iOS deployment
 This is the model the [Prism AAC](https://github.com/dcostenco/prism-aac) iOS / iPadOS / watchOS apps load on-device via **llama.cpp Swift SPM**. CoreML was evaluated and rejected — `coremltools` doesn't support Qwen3 attention ops.
+GGUF available: `prism-aac-1b7-q4km.gguf` (1.0 GB on disk, ~1.6 GB RAM at runtime).
 ## Training recipe (v19)
+- **Base**: Qwen/Qwen3-1.7B + v26 system-prompt engineering
+- **Note**: Uses **system-prompt engineering only** (no LoRA fine-tune). Small models over-fit aggressively — a v25-max LoRA attempt regressed 100% → 93.8% on a 16-case gate test.
+- **Quantization**: Q4_K_M (1.0 GB) for iOS, Q8 (2.2 GB) for Mac
 ## Usage
 ollama run dcostenco/prism-coder:1b7 "Search my knowledge for OAuth scope reference"
 ```
 ### System prompt
+Use the [v26 routing prompt](https://github.com/dcostenco/prism-coder/blob/main/tests/benchmarks/prism-routing-100/benchmark.py#L47) verbatim. This model relies *heavily* on the system prompt. Key: rules 1-7 must say `-> respond directly (no tool)`.
 ## Hardware requirements
 - **iPhone / iPad**: A14 Bionic+ (iPhone 12+), iOS 15+, ~1.6 GB free RAM
 - **Mac**: any M-series (M1 8 GB unified is enough at Q4_K_M)
 - **Linux**: any CPU works at Q4_K_M; a 4 GB GPU dramatically speeds inference
 - **Inference speed**: ~0.5–1.6 s per 200-token response