Instructions to use dcostenco/prism-coder-4b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use dcostenco/prism-coder-4b with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="dcostenco/prism-coder-4b",
	filename="prism-coder-4b-v43-Q4_K_M.gguf",
)

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

Notebooks
Google Colab
Kaggle
Local Apps

llama.cpp

How to use dcostenco/prism-coder-4b 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-4b:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf dcostenco/prism-coder-4b:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf dcostenco/prism-coder-4b:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf dcostenco/prism-coder-4b:Q4_K_M

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-4b:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf dcostenco/prism-coder-4b:Q4_K_M

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-4b:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf dcostenco/prism-coder-4b:Q4_K_M

Use Docker

docker model run hf.co/dcostenco/prism-coder-4b:Q4_K_M

LM Studio
Jan
Ollama
How to use dcostenco/prism-coder-4b with Ollama:
```
ollama run hf.co/dcostenco/prism-coder-4b:Q4_K_M
```

Unsloth Studio new

How to use dcostenco/prism-coder-4b 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-4b 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-4b 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-4b to start chatting

Pi new

How to use dcostenco/prism-coder-4b 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-4b:Q4_K_M

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-4b:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use dcostenco/prism-coder-4b 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-4b:Q4_K_M

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-4b:Q4_K_M

Run Hermes

hermes

Docker Model Runner
How to use dcostenco/prism-coder-4b with Docker Model Runner:
```
docker model run hf.co/dcostenco/prism-coder-4b:Q4_K_M
```

Lemonade

How to use dcostenco/prism-coder-4b with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull dcostenco/prism-coder-4b:Q4_K_M

Run and chat with the model

lemonade run user.prism-coder-4b-Q4_K_M

List all available models

lemonade list

prism-coder:4b — Full Prism Memory Router (Mid-Tier)

Fine-tuned Qwen3-4B for 17-tool Prism Memory routing in the Prism AAC system. Primary deployment: Mac / PC / high-memory mobile via Ollama or llama.cpp GGUF — for devices with ≥8 GB free RAM.

BFCL Routing Benchmark — v43 (Current)

100.0% (64/64 strict, 8 categories)

Category	Count	Description	Accuracy
simple	10	Direct single-tool invocations	100%
relevance_detection	10	No-tool abstention for off-topic prompts	100%
hallucination	10	Reject fabricated / nonexistent tools	100%
disambiguation	8	Pick correct tool from near-neighbors	100%
format_sensitivity	5	Varied natural phrasing for same intent	100%
ast_parameter	5	Correct argument extraction	100%
edge_case	8	Boundary and adversarial inputs	100%
multi_turn_chain	8	Two-step tool sequences	100%

Eval: Ollama inference, temperature=0, greedy decode. Gate: ≥90% = deploy.

SWE Bench Blind Eval — v43

100.0% (68/68 strict, 7 categories) — held-out test set, no overlap with training data.

Category	Count	Accuracy
adversarial_trap	15	100%
cascade	10	100%
disambiguation	8	100%
edge_case	8	100%
multi_intent	4	100%
natural_phrasing	15	100%
verifier	8	100%

eval-300 — v43

100.0% (300/300 strict, 5 shuffled runs, 0 flaky tests)

Category	Count	Accuracy
abstention	20	100%
adversarial_trap	70	100%
cascade	25	100%
disambiguation	40	100%
edge_case	25	100%
multi_intent	20	100%
natural_phrasing	50	100%
param_extraction	25	100%
verifier	25	100%

Version History

Version	BFCL	SWE Bench	eval-300	Notes
v43	100%	100%	100%	Qwen3-4B base, 17-tool full router, Layer 3 inference-time remapping, 5 surgical patches

Tools

The model routes to 17 Prism Memory tools:

Tool	Trigger
`session_load_context`	Load / resume / catch me up on project context
`session_save_ledger`	Jot down / log / note / record what we did
`session_save_experience`	Log milestone / achievement / success event
`session_save_handoff`	Save state for next agent / shift change
`session_search_memory`	Recall / remind me / find what we decided
`session_forget_memory`	Delete a specific memory entry by ID
`session_export_memory`	Export session to file (JSON / Markdown)
`session_compact_ledger`	Compact / prune old session entries
`session_health_check`	Check session integrity
`session_synthesize_edges`	Verify / rebuild session link graph
`session_backfill_links`	Reconnect / patch missing session links
`session_task_route`	Route a task to the right agent tier
`knowledge_search`	Search knowledge base / accumulated docs
`knowledge_forget`	Delete knowledge entries / wipe records
`knowledge_upvote`	Upvote / boost / increase rank of entry
`knowledge_downvote`	Downvote / lower rank of entry
`knowledge_set_retention`	Set TTL / auto-expire / retention policy

Plain text (no tool) for: greetings, general questions, math, code help, weather, CS concepts.

Model Details

Base: Qwen/Qwen3-4B
Format: GGUF Q4_K_M (~2.3 GB)
Context: 32,768 tokens
Training: MLX LoRA on Apple Silicon, rank=32, alpha=64, 16/36 layers, LR=1e-4 (full) → 3e-5 (surgical patches), 5 patch rounds
Corpus: ~30K rows — 36% tool-use, 40% AAC/clinical, 12% abstention, 12% safety
Merge: direct safetensors delta merge (delta = (alpha/rank) × B.T @ A.T) — mlx_lm.fuse not used (silently drops LoRA weights)
Quantization: llama.cpp F16 → Q4_K_M

Usage

ollama pull dcostenco/prism-coder:4b-v43
ollama run dcostenco/prism-coder:4b-v43

Or drop the GGUF into any llama.cpp-compatible runtime (LM Studio, Jan, llama-server).

In Prism AAC the app loads this model automatically on devices with ≥8 GB free RAM.

Training Scripts

The training/ folder in this repo contains the full v43 training pipeline:

Script	Purpose
`build_4b_v43_corpus.py`	Full v43 corpus builder (~30K rows)
`build_4b_v43_patch.py`	Patch 1 — initial BFCL failures
`build_4b_v43_patch2.py`	Patch 2 — param extraction + format
`build_4b_v43_patch4.py`	Patch 4 — task_route + casual phrasing
`build_4b_v43_swe_patch.py`	Patch 5 — SWE bench targeted
`combine_4b_swe_corpus.py`	Merge base + SWE patch corpus
`train_4b_v43_local.sh`	MLX LoRA training (Apple Silicon)
`train_4b_v43_swe_patch.sh`	Surgical SWE patch training run
`merge_4b_v43.py`	Safe LoRA merge (delta = scale × B.T @ A.T)
`export_4b_v43_gguf.sh`	HF safetensors → GGUF F16 → Q4_K_M → Ollama
`orchestrate_4b_to_100.sh`	Autonomous patch→train→eval loop
`bfcl_eval.py`	64-test BFCL eval harness with Layer 3
`swe_bench_test.py`	68-test SWE blind eval harness
`eval_300.py`	300-test standard eval (9 categories)
`analyze_swe_failures.py`	Parse failures → patch targets
`TRAINING_DECISIONS_4B_V43.md`	Hyperparams, corpus ratios, lessons learned

Model Family

Model	GGUF	RAM	Tools	Repo
prism-coder:1b7	1.2 GB	≥3 GB	6	dcostenco/prism-coder-1.7b
prism-coder:4b	2.3 GB	≥8 GB	17	this repo
prism-coder:8b	4.9 GB	≥16 GB	6	dcostenco/prism-coder-8b
prism-coder:14b	8.4 GB	≥24 GB	6 + TypeScript	dcostenco/prism-coder-14b
prism-coder:32b	16 GB	≥48 GB	6	dcostenco/prism-coder-32b

Downloads last month: -

GGUF

Model size

4B params

Architecture

qwen3

Hardware compatibility

4-bit

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for dcostenco/prism-coder-4b

Base model

Qwen/Qwen3-4B-Base

Finetuned

Qwen/Qwen3-4B

Quantized

(220)

this model