AbteeX AI Labs conversational CLI, model downloader, and router for the LumynaX Hugging Face model family. One install. 98 curated LumynaX model entries. Pick a model, pull files, then run it locally.

License

Python

What is MaramaRoute?

MaramaRoute is the AbteeX AI Labs command line surface for LumynaX models:

chat opens a conversational model picker and prompt loop. The picker starts with /help guidance before model results.
catalog shows the bundled Hugging Face-backed LumynaX registry.
pull downloads model files into the local MaramaRoute cache.
verify can hash pulled artifacts and write local SHA256 manifests.
run behaves like chat when no prompt is supplied, or returns one answer for one prompt.
route, compare, matrix, and analytics explain which LumynaX model fits a request and why.
serve starts the local browser console and route API for governed applications.
agent and hpe commands prepare coding-agent and HPE/HPC operator workflows.

The package does not bundle large model weights. It bundles the registry and download logic, then pulls files from the AbteeXAILab Hugging Face repos selected by model id. Every bundled registry entry is selectable in the CLI: GGUF models use direct local chat, Transformers text-generation models use the local Transformers runtime, and task models open an offline task shell with their local files.

Runtime compatibility at a glance

MaramaRoute is explicit about what can run directly, what can be served through production backends, and what needs validation or conversion:

Category	Count	What it means
Bundled registry entries	98	Every entry can be searched, selected, pulled, inspected, and routed.
Direct local chat-capable entries	68	GGUF/llama.cpp entries that the CLI treats as direct chat models after pull.
GGUF artifacts	65	Quantized local artifacts; direct chat uses `llama-cpp-python`.
llama.cpp runtime entries	65	Primary direct local runtime path for GGUF chat models.
Transformers runtime entries	21	Selectable and pullable local runtime entries; useful chat depends on tokenizer support and model task.
vLLM-compatible/candidate entries	69	Production backend path; GGUF entries are marked experimental/candidate where tokenizer/config validation is required.
NVIDIA NIM-compatible/candidate entries	73	NIM deployment path; validate architecture, tokenizer, config, and folder layout before production.
NVIDIA NeMo/NEM entries	73	5 direct-compatible plus 68 conversion-path entries.

GGUF entries are precise in the registry: they are direct local-chat packages for llama-cpp-python. A GGUF row may also carry vLLM, NIM, or NeMo/NEM metadata, but those labels mean deployment path, candidate, experimental support, or conversion path; operators should run MaramaRoute compat <model-id> before treating a backend as production-ready.

Use these commands before production deployment:

bash

MaramaRoute categories
MaramaRoute compat --target vllm --status usable
MaramaRoute compat --target nim --status usable
MaramaRoute compat --target nemo --status pathway
MaramaRoute compat <model-id>

Install

pip install lumynax-marama-route

The same CLI is available through npm for Node-first teams:

npm install -g lumynax-marama-route

Six command aliases are installed:

bash

MaramaRoute --help
LumynaXRoute --help
marama-route --help
maramaroute --help
lumynax-route --help
lumynaxroute --help

30-second Quickstart

bash

pip install lumynax-marama-route

# Start the conversational picker. It opens with options; use /help for commands.
MaramaRoute chat

# Or jump straight to a model by id or search fragment.
MaramaRoute chat qwen25-05b

# Create local config, default alias, coding-agent bridge files, and HPE scaffold.
MaramaRoute setup --all-targets --hpe
MaramaRoute agent doctor --model qwen25-7b

# Inspect registry categories before choosing a model.
MaramaRoute categories

# Install local runtimes when you want direct generation.
python -m pip install llama-cpp-python
python -m pip install torch
python -m pip install sentencepiece tiktoken tokenizers

# Estimate, pull, verify, and chat.
MaramaRoute pull qwen25-05b --estimate --remote-sizes
MaramaRoute pull lumynax-coder-qwen25-05b-instruct-gguf
MaramaRoute verify --deep --write-hashes
MaramaRoute run lumynax-coder-qwen25-05b-instruct-gguf

After MaramaRoute pull, chat and run load local files only. GGUF models use llama-cpp-python; Transformers text-generation models use the bundled tokenizer support plus torch; embedding, reranker, OCR, speech, and other task entries stay offline and expose their task-model selection plus local file paths.

For a small conversational model, prefer lumynax-tiny-qwen25-05b-gguf. The lumynax-tiny Transformers seed is runnable, but it is labelled as smoke-test because it is meant for install/runtime checks rather than useful chat.

Every bundled model also has a runtime compatibility assessment:

bash

MaramaRoute compat
MaramaRoute compat --target vllm --format json
MaramaRoute compat --target nim --status candidate
MaramaRoute compat lumynax-tiny-qwen25-05b-gguf

The matrix covers llama_cpp, vllm, nvidia_nim, and nvidia_nemo. It uses supported, candidate, experimental, convert_required, and unsupported statuses so production users can distinguish direct local runtimes from backend-specific validation work.

If you only want to inspect what would be downloaded:

bash

MaramaRoute pull lumynax-coder-qwen25-05b-instruct-gguf --dry-run
MaramaRoute local

Inside chat mode:

text

/models      show direct local GGUF chat-capable LumynaX models
/hardware    show models suitable for this machine
/recommended show recommended local chat models
/all         show all 98 bundled AbteeXAILab Hugging Face registry entries
/search qwen search model id, repo, family, or tags
/categories  show family/runtime/tag/modality/capability counts
/families    alias for /categories
/family qwen filter the full registry to a family or category
/next        next page of model results
/prev        previous page of model results
/menu        return to the picker menu
/switch      change model
/switch qwen switch directly by search text
/pull        download the selected model
/pull qwen25 download another matching model and switch to it
/local       show pulled models
/settings    show current runtime settings
/clear       clear chat history
/history     show current chat history
/save work   save current chat history
/load work   load saved chat history
/export work work.md export saved chat as markdown
/info        show the selected model card metadata
/exit        quit

Model download commands

bash

# Download the primary GGUF artifact listed in the registry.
MaramaRoute pull lumynax-coder-qwen25-05b-instruct-gguf

# Non-GGUF task and Transformers entries pull a full local repo snapshot.
MaramaRoute pull lumynax-embed-bge-m3

# Download every file in the Hugging Face repo snapshot.
MaramaRoute pull lumynax-coder-qwen25-05b-instruct-gguf --all-files

# Use a custom cache directory.
MaramaRoute pull lumynax-coder-qwen25-05b-instruct-gguf --cache-dir ./models

# Batch-plan downloads by family/search/runtime before committing.
MaramaRoute pull --search qwen --limit 3 --dry-run
MaramaRoute pull --search embed --limit 3 --dry-run
MaramaRoute pull --search qwen --chat-only --limit 3 --dry-run
MaramaRoute pull --family qwen --limit 3 --yes
MaramaRoute pull qwen25-05b --estimate
MaramaRoute pull qwen25-05b --estimate --remote-sizes

# Hash pulled files and write a local verification manifest.
MaramaRoute verify --deep --write-hashes

# Run locally after pull.
MaramaRoute run lumynax-coder-qwen25-05b-instruct-gguf --stream "Write a tiny Python function."

# Conversational loop; omit the prompt.
MaramaRoute run lumynax-coder-qwen25-05b-instruct-gguf

Production operator checks

Use these commands before putting a machine or workspace into regular use:

bash

# Install, registry, cache, HF tooling, and local runtime readiness.
MaramaRoute doctor --hardware

# One-shot production bootstrap: local config, aliases, agent bridge files, and optional HPE bundle.
MaramaRoute setup qwen25-05b --all-targets --hpe
MaramaRoute setup qwen25-7b --target claude-code,codex,continue,opencode,litellm,tabby --hpe --backend vllm

# Workspace bridge config and optional gateway health probe.
MaramaRoute agent doctor --target claude-code --model qwen25-7b
MaramaRoute agent doctor --target codex --model qwen25-7b
MaramaRoute agent doctor --target continue --model qwen25-7b
MaramaRoute agent doctor --target litellm --model qwen25-7b
MaramaRoute agent doctor --target hpe --model qwen25-7b --base-url http://127.0.0.1:8787/v1

# Exact remote size planning, local hash verification, and registry drift check.
MaramaRoute pull qwen25-05b --estimate --remote-sizes
MaramaRoute verify --deep --write-hashes
MaramaRoute update-registry --dry-run --diff

# HPE/HPC scaffold: Slurm, live gateway config, backend launch, and run notes.
MaramaRoute hpe plan qwen25-7b --backend vllm
MaramaRoute hpe init qwen25-7b --backend vllm --gpus 1
MaramaRoute hpe init qwen25-7b --backend nim --backend-base-url http://127.0.0.1:8000/v1
MaramaRoute hpe init qwen25-7b --backend nemo --backend-command ./start-nemo-backend.sh

The CLI never bundles large model weights into the package. It records the selected artifact paths, local cache path, and optional SHA256 manifest after operators pull models from the AbteeXAILab Hugging Face repos.

Registry coverage

Distribution across the LumynaX family:

Runtimes: llama_cpp (65) | llama_cpp_multimodal (3) | python_embedding (4) | transformers (21) | transformers_multimodal (5)

Every model in the bundled registry carries:

model_id
repo_id
runtime
modalities
context_tokens
residency
license_id
sovereignty_tier
primary_artifact

Common catalog commands:

bash

MaramaRoute models
MaramaRoute catalog --task code --limit 10
MaramaRoute catalog --task reasoning --requires-tools --jurisdiction NZ
MaramaRoute catalog --search qwen --family qwen --limit 20
MaramaRoute analytics
MaramaRoute categories
MaramaRoute recommend --task code --sensitivity restricted --prompt-text "Refactor a private Python service"
MaramaRoute hardware --recommend
MaramaRoute doctor --hardware
MaramaRoute ls
MaramaRoute disk
MaramaRoute verify
MaramaRoute verify --deep --write-hashes
MaramaRoute alias set code qwen25-7b
MaramaRoute favorite qwen25-05b
MaramaRoute bench qwen25-05b --dry-run
MaramaRoute eval

Full Hugging Face model list

Model id	Hugging Face repo	Runtime	Tier	Primary artifact
`lumynax-chat-hermes-3-llama31-8b-gguf`	`AbteeXAILab/lumynax-chat-hermes-3-llama31-8b-gguf`	`llama_cpp`	3	`Hermes-3-Llama-3.1-8B-Q4_K_M.gguf`
`lumynax-chat-yi-15-34b-gguf`	`AbteeXAILab/lumynax-chat-yi-15-34b-gguf`	`llama_cpp`	3	`Yi-1.5-34B-Chat-Q4_K_M.gguf`
`lumynax-coder-codellama-70b-instruct-gguf`	`AbteeXAILab/lumynax-coder-codellama-70b-instruct-gguf`	`llama_cpp`	3	`codellama-70b-instruct.Q4_K_M.gguf`
`lumynax-coder-codeqwen15-7b-chat-gguf`	`AbteeXAILab/lumynax-coder-codeqwen15-7b-chat-gguf`	`llama_cpp`	3	`codeqwen-1_5-7b-chat-q4_k_m.gguf`
`lumynax-coder-deepseek-coder-33b-gguf`	`AbteeXAILab/lumynax-coder-deepseek-coder-33b-gguf`	`llama_cpp`	3	`deepseek-coder-33b-instruct.Q4_K_M.gguf`
`lumynax-coder-deepseek-v2-lite-16b-gguf`	`AbteeXAILab/lumynax-coder-deepseek-v2-lite-16b-gguf`	`llama_cpp`	3	`DeepSeek-Coder-V2-Lite-Instruct-Q4_K_M.gguf`
`lumynax-coder-qwen25-05b-instruct-gguf`	`AbteeXAILab/lumynax-coder-qwen25-05b-instruct-gguf`	`llama_cpp`	3	`qwen2.5-coder-0.5b-instruct-q4_k_m.gguf`
`lumynax-coder-qwen25-14b-instruct-gguf`	`AbteeXAILab/lumynax-coder-qwen25-14b-instruct-gguf`	`llama_cpp`	3	`qwen2.5-coder-14b-instruct-q4_k_m.gguf`
`lumynax-coder-qwen25-15b-instruct-gguf`	`AbteeXAILab/lumynax-coder-qwen25-15b-instruct-gguf`	`llama_cpp`	3	`qwen2.5-coder-1.5b-instruct-q4_k_m.gguf`
`lumynax-coder-qwen25-7b-instruct-gguf`	`AbteeXAILab/lumynax-coder-qwen25-7b-instruct-gguf`	`llama_cpp`	3	`qwen2.5-coder-7b-instruct-q4_k_m.gguf`
`lumynax-coder-qwen25-coder-32b-gguf`	`AbteeXAILab/lumynax-coder-qwen25-coder-32b-gguf`	`llama_cpp`	3	`Qwen2.5-Coder-32B-Instruct-Q4_K_M.gguf`
`lumynax-coder-starcoder2-15b-gguf`	`AbteeXAILab/lumynax-coder-starcoder2-15b-gguf`	`llama_cpp`	3	`starcoder2-15b-instruct-v0.1-Q4_K_M.gguf`
`lumynax-coder-yi-coder-9b-gguf`	`AbteeXAILab/lumynax-coder-yi-coder-9b-gguf`	`llama_cpp`	3	`Yi-Coder-9B-Chat-Q4_K_M.gguf`
`lumynax-doc-donut-base`	`AbteeXAILab/lumynax-doc-donut-base`	`transformers`	3	`pytorch_model.bin`
`lumynax-doc-layoutlmv3-base`	`AbteeXAILab/lumynax-doc-layoutlmv3-base`	`transformers`	3	`pytorch_model.bin`
`lumynax-doc-nougat-base`	`AbteeXAILab/lumynax-doc-nougat-base`	`transformers`	3	`pytorch_model.bin`
`lumynax-doc-table-transformer-detection`	`AbteeXAILab/lumynax-doc-table-transformer-detection`	`transformers`	3	`pytorch_model.bin`
`lumynax-embed-bge-m3`	`AbteeXAILab/lumynax-embed-bge-m3`	`python_embedding`	2	`merged_model/pytorch_model.bin`
`lumynax-embed-e5-mistral-7b`	`AbteeXAILab/lumynax-embed-e5-mistral-7b`	`python_embedding`	2	`merged_model/model-00001-of-00002.safetensors`
`lumynax-embed-granite-278m-multilingual`	`AbteeXAILab/lumynax-embed-granite-278m-multilingual`	`python_embedding`	3	`pytorch_model.bin`
`lumynax-embed-nomic-v2-moe`	`AbteeXAILab/lumynax-embed-nomic-v2-moe`	`python_embedding`	3	`model.safetensors`
`lumynax-frontier-coder-deepseek-v25-1210-gguf`	`AbteeXAILab/lumynax-frontier-coder-deepseek-v25-1210-gguf`	`llama_cpp`	2	`DeepSeek-V2.5-1210-Q4_K_M/DeepSeek-V2.5-1210-Q4_K_M-00002-of-00004.gguf`
`lumynax-frontier-coder-qwen3-480b-a35b-gguf`	`AbteeXAILab/lumynax-frontier-coder-qwen3-480b-a35b-gguf`	`llama_cpp`	2	`Q4_K_M/Qwen3-Coder-480B-A35B-Instruct-Q4_K_M-00001-of-00006.gguf`
`lumynax-frontier-dbrx-instruct-132b-gguf`	`AbteeXAILab/lumynax-frontier-dbrx-instruct-132b-gguf`	`llama_cpp`	3	`dbrx-instruct.Q2_K.gguf`
`lumynax-frontier-minimax-m2-230b`	`AbteeXAILab/lumynax-frontier-minimax-m2-230b`	`transformers`	2	`Q4_K_M/MiniMax-M2-Q4_K_M-00002-of-00003.gguf`
`lumynax-frontier-minimax-m25-unsloth`	`AbteeXAILab/lumynax-frontier-minimax-m25-unsloth`	`llama_cpp`	3	`MiniMax-M2.5-UD-TQ1_0.gguf`
`lumynax-frontier-mixtral-8x22b-instruct-gguf`	`AbteeXAILab/lumynax-frontier-mixtral-8x22b-instruct-gguf`	`llama_cpp`	3	`Mixtral-8x22B-Instruct-v0.1.Q4_K_M-00001-of-00002.gguf`
`lumynax-frontier-olmo2-32b-instruct`	`AbteeXAILab/lumynax-frontier-olmo2-32b-instruct`	`transformers`	2	`model-00001-of-00014.safetensors`
`lumynax-frontier-phi-35-moe-instruct-gguf`	`AbteeXAILab/lumynax-frontier-phi-35-moe-instruct-gguf`	`llama_cpp`	3	`Phi-3.5-MoE-instruct-Q4_K_M.gguf`
`lumynax-frontier-phi-4-14b-gguf`	`AbteeXAILab/lumynax-frontier-phi-4-14b-gguf`	`llama_cpp`	3	`phi-4-Q4_K_M.gguf`
`lumynax-frontier-qwen25-72b-instruct-gguf`	`AbteeXAILab/lumynax-frontier-qwen25-72b-instruct-gguf`	`llama_cpp`	3	`Qwen2.5-72B-Instruct-Q4_K_M.gguf`
`lumynax-frontier-qwen3-235b-a22b-instruct`	`AbteeXAILab/lumynax-frontier-qwen3-235b-a22b-instruct`	`transformers`	2	`Q4_K_M/Qwen3-235B-A22B-Instruct-2507-Q4_K_M-00001-of-00003.gguf`
`lumynax-guard-text-moderation`	`AbteeXAILab/lumynax-guard-text-moderation`	`transformers`	3	`pytorch_model.bin`
`lumynax-infused-gemma-e4b`	`AbteeXAILab/lumynax-infused-gemma-e4b`	`transformers_multimodal`	2	`merged_model/model.safetensors`
`lumynax-infused-gemma-e4b-text-gguf`	`AbteeXAILab/lumynax-infused-gemma-e4b-text-gguf`	`llama_cpp`	3	`lumynax-infused-gemma-e4b-text-gguf-q4_k_m.gguf`
`lumynax-infused-gemma4-26b-a4b-gguf`	`AbteeXAILab/lumynax-infused-gemma4-26b-a4b-gguf`	`llama_cpp`	3	`lumynax-infused-gemma4-26b-a4b-ud-q4_k_m.gguf`
`lumynax-infused-granite31-1b-a400m-gguf`	`AbteeXAILab/lumynax-infused-granite31-1b-a400m-gguf`	`llama_cpp`	3	`granite-3.1-1b-a400m-instruct-Q4_K_M.gguf`
`lumynax-infused-granite33-2b-gguf`	`AbteeXAILab/lumynax-infused-granite33-2b-gguf`	`llama_cpp`	3	`granite-3.3-2b-instruct-Q4_K_M.gguf`
`lumynax-infused-granite33-8b-gguf`	`AbteeXAILab/lumynax-infused-granite33-8b-gguf`	`llama_cpp`	3	`granite-3.3-8b-instruct-Q4_K_M.gguf`
`lumynax-infused-mistral-7b-v03-gguf`	`AbteeXAILab/lumynax-infused-mistral-7b-v03-gguf`	`llama_cpp`	3	`Mistral-7B-Instruct-v0.3-Q4_K_M.gguf`
`lumynax-infused-mistral-small-text-gguf`	`AbteeXAILab/lumynax-infused-mistral-small-text-gguf`	`llama_cpp`	3	`lumynax-infused-mistral-small-text-gguf-f16.gguf`
`lumynax-infused-olmo2-1b-0425-gguf`	`AbteeXAILab/lumynax-infused-olmo2-1b-0425-gguf`	`llama_cpp`	3	`OLMo-2-0425-1B-Instruct-Q4_K_M.gguf`
`lumynax-infused-olmo2-7b-1124-gguf`	`AbteeXAILab/lumynax-infused-olmo2-7b-1124-gguf`	`llama_cpp`	3	`olmo-2-1124-7B-instruct-Q4_K_M.gguf`
`lumynax-infused-phi-4-text-gguf`	`AbteeXAILab/lumynax-infused-phi-4-text-gguf`	`llama_cpp`	3	`lumynax-infused-phi-4-text-gguf-f16.gguf`
`lumynax-infused-phi3-mini-4k-gguf`	`AbteeXAILab/lumynax-infused-phi3-mini-4k-gguf`	`llama_cpp`	3	`Phi-3-mini-4k-instruct-q4.gguf`
`lumynax-infused-phi4-mini-instruct-gguf`	`AbteeXAILab/lumynax-infused-phi4-mini-instruct-gguf`	`llama_cpp`	3	`Phi-4-mini-instruct-Q4_K_M.gguf`
`lumynax-infused-qwen2-audio-7b`	`AbteeXAILab/lumynax-infused-qwen2-audio-7b`	`transformers`	2	`merged_model/model-00003-of-00005.safetensors`
`lumynax-infused-qwen25-15b-instruct-gguf`	`AbteeXAILab/lumynax-infused-qwen25-15b-instruct-gguf`	`llama_cpp`	3	`qwen2.5-1.5b-instruct-q4_k_m.gguf`
`lumynax-infused-qwen25-3b-instruct-gguf`	`AbteeXAILab/lumynax-infused-qwen25-3b-instruct-gguf`	`llama_cpp`	3	`qwen2.5-3b-instruct-q4_k_m.gguf`
`lumynax-infused-qwen25-7b-instruct-gguf`	`AbteeXAILab/lumynax-infused-qwen25-7b-instruct-gguf`	`llama_cpp`	3	`qwen2.5-7b-instruct-q3_k_m.gguf`
`lumynax-infused-qwen25-omni-7b-voice`	`AbteeXAILab/lumynax-infused-qwen25-omni-7b-voice`	`transformers`	2	`merged_model/model-00003-of-00005.safetensors`
`lumynax-infused-qwen3-06b-gguf`	`AbteeXAILab/lumynax-infused-qwen3-06b-gguf`	`llama_cpp`	3	`Qwen3-0.6B-Q8_0.gguf`
`lumynax-infused-qwen3-14b-gguf`	`AbteeXAILab/lumynax-infused-qwen3-14b-gguf`	`llama_cpp`	3	`Qwen3-14B-Q4_K_M.gguf`
`lumynax-infused-qwen3-17b-gguf`	`AbteeXAILab/lumynax-infused-qwen3-17b-gguf`	`llama_cpp`	3	`Qwen3-1.7B-Q8_0.gguf`
`lumynax-infused-qwen3-30b-a3b-gguf`	`AbteeXAILab/lumynax-infused-qwen3-30b-a3b-gguf`	`llama_cpp`	3	`lumynax-infused-qwen3-30b-a3b-q4_k_m.gguf`
`lumynax-infused-qwen3-8b-gguf`	`AbteeXAILab/lumynax-infused-qwen3-8b-gguf`	`llama_cpp`	3	`lumynax-infused-qwen3-8b-q4_k_m.gguf`
`lumynax-infused-qwen3-coder-30b-a3b-gguf`	`AbteeXAILab/lumynax-infused-qwen3-coder-30b-a3b-gguf`	`llama_cpp`	3	`lumynax-infused-qwen3-coder-30b-a3b-q4_k_m.gguf`
`lumynax-infused-qwen3-text-gguf`	`AbteeXAILab/lumynax-infused-qwen3-text-gguf`	`llama_cpp`	3	`lumynax-infused-qwen3-text-gguf-f16.gguf`
`lumynax-infused-smollm-135m-gguf`	`AbteeXAILab/lumynax-infused-smollm-135m-gguf`	`llama_cpp`	3	`SmolLM-135M-Instruct.Q4_K_M.gguf`
`lumynax-infused-smollm2-17b-gguf`	`AbteeXAILab/lumynax-infused-smollm2-17b-gguf`	`llama_cpp`	3	`smollm2-1.7b-instruct-q4_k_m.gguf`
`lumynax-infused-smollm2-360m-gguf`	`AbteeXAILab/lumynax-infused-smollm2-360m-gguf`	`llama_cpp`	3	`smollm2-360m-instruct-q8_0.gguf`
`lumynax-infused-smollm3-3b-gguf`	`AbteeXAILab/lumynax-infused-smollm3-3b-gguf`	`llama_cpp`	3	`SmolLM3-Q4_K_M.gguf`
`lumynax-infused-zephyr-7b-beta-gguf`	`AbteeXAILab/lumynax-infused-zephyr-7b-beta-gguf`	`llama_cpp`	3	`zephyr-7b-beta.Q4_K_M.gguf`
`lumynax-longctx-glm4-9b-chat-1m-gguf`	`AbteeXAILab/lumynax-longctx-glm4-9b-chat-1m-gguf`	`llama_cpp`	3	`glm-4-9b-chat-1m-Q4_K_M.gguf`
`lumynax-longctx-prolong-512k-instruct`	`AbteeXAILab/lumynax-longctx-prolong-512k-instruct`	`transformers`	3	`model-00004-of-00007.safetensors`
`lumynax-longctx-qwen25-7b-1m-gguf`	`AbteeXAILab/lumynax-longctx-qwen25-7b-1m-gguf`	`llama_cpp`	3	`Qwen2.5-7B-Instruct-1M-Q4_K_M.gguf`
`lumynax-longctx-yi-9b-200k`	`AbteeXAILab/lumynax-longctx-yi-9b-200k`	`transformers`	3	`model-00002-of-00004.safetensors`
`lumynax-math-qwen25-math-7b-gguf`	`AbteeXAILab/lumynax-math-qwen25-math-7b-gguf`	`llama_cpp`	3	`Qwen2.5-Math-7B-Instruct-Q4_K_M.gguf`
`lumynax-moe-moonlight-16b-a3b-gguf`	`AbteeXAILab/lumynax-moe-moonlight-16b-a3b-gguf`	`llama_cpp`	3	`lumynax-moe-moonlight-16b-a3b-iq4_xs.gguf`
`lumynax-moe-olmoe-1b-7b-0924-instruct-gguf`	`AbteeXAILab/lumynax-moe-olmoe-1b-7b-0924-instruct-gguf`	`llama_cpp`	3	`OLMoE-1B-7B-0924-Instruct-Q4_K_M.gguf`
`lumynax-moe-olmoe-1b-7b-gguf`	`AbteeXAILab/lumynax-moe-olmoe-1b-7b-gguf`	`llama_cpp`	3	`olmoe-1b-7b-0924-instruct-q4_k_m.gguf`
`lumynax-multimodal-aria-25b-moe`	`AbteeXAILab/lumynax-multimodal-aria-25b-moe`	`transformers_multimodal`	2	`model-00001-of-00012.safetensors`
`lumynax-multimodal-glm46v-flash`	`AbteeXAILab/lumynax-multimodal-glm46v-flash`	`llama_cpp_multimodal`	2	`lumynax-multimodal-glm46v-flash-ud-iq2_m.gguf`
`lumynax-multimodal-internvl3-78b-instruct`	`AbteeXAILab/lumynax-multimodal-internvl3-78b-instruct`	`transformers_multimodal`	2	`model-00001-of-00033.safetensors`
`lumynax-multimodal-kimi-vl-a3b-thinking`	`AbteeXAILab/lumynax-multimodal-kimi-vl-a3b-thinking`	`llama_cpp_multimodal`	2	`lumynax-multimodal-kimi-vl-a3b-thinking-q4_k_m.gguf`
`lumynax-multimodal-llava-next-34b`	`AbteeXAILab/lumynax-multimodal-llava-next-34b`	`transformers_multimodal`	2	`model-00006-of-00015.safetensors`
`lumynax-multimodal-pixtral-large-124b`	`AbteeXAILab/lumynax-multimodal-pixtral-large-124b`	`transformers_multimodal`	2	`consolidated-00004-of-00052.safetensors`
`lumynax-multimodal-qwen25-vl-72b-instruct-gguf`	`AbteeXAILab/lumynax-multimodal-qwen25-vl-72b-instruct-gguf`	`llama_cpp_multimodal`	3	`Qwen2.5-VL-72B-Instruct-Q4_K_M.gguf`
`lumynax-nz-3b`	`AbteeXAILab/lumynax-nz-3b`	`transformers`	2	`merged_model/model-00001-of-00055.safetensors`
`lumynax-nz-qwen25-coder-3b-gguf`	`AbteeXAILab/lumynax-nz-qwen25-coder-3b-gguf`	`llama_cpp`	3	`lumynax-nz-qwen25-coder-3b-q4_k_m.gguf`
`lumynax-ocr-trocr-large-handwritten`	`AbteeXAILab/lumynax-ocr-trocr-large-handwritten`	`transformers`	3	`pytorch_model.bin`
`lumynax-ocr-trocr-large-printed`	`AbteeXAILab/lumynax-ocr-trocr-large-printed`	`transformers`	3	`pytorch_model.bin`
`lumynax-reasoning-deepseek-distill-text-gguf`	`AbteeXAILab/lumynax-reasoning-deepseek-distill-text-gguf`	`llama_cpp`	3	`lumynax-reasoning-deepseek-distill-text-gguf-f16.gguf`
`lumynax-reasoning-deepseek-prover-v2-671b-gguf`	`AbteeXAILab/lumynax-reasoning-deepseek-prover-v2-671b-gguf`	`llama_cpp`	2	`Q4_K_M/DeepSeek-Prover-V2-671B-Q4_K_M-00003-of-00009.gguf`
`lumynax-reasoning-deepseek-r1-distill-llama-70b-gguf`	`AbteeXAILab/lumynax-reasoning-deepseek-r1-distill-llama-70b-gguf`	`llama_cpp`	3	`DeepSeek-R1-Distill-Llama-70B-Q4_K_M.gguf`
`lumynax-reasoning-deepseek-r1-qwen-15b-gguf`	`AbteeXAILab/lumynax-reasoning-deepseek-r1-qwen-15b-gguf`	`llama_cpp`	3	`DeepSeek-R1-Distill-Qwen-1.5B-Q4_K_M.gguf`
`lumynax-reasoning-deepseek-r1-qwen-7b-gguf`	`AbteeXAILab/lumynax-reasoning-deepseek-r1-qwen-7b-gguf`	`llama_cpp`	3	`DeepSeek-R1-Distill-Qwen-7B-Q4_K_M.gguf`
`lumynax-reasoning-glm46-355b-moe`	`AbteeXAILab/lumynax-reasoning-glm46-355b-moe`	`transformers`	2	`Q4_K_M/GLM-4.6-Q4_K_M-00001-of-00005.gguf`
`lumynax-reasoning-gpt-oss-20b-gguf`	`AbteeXAILab/lumynax-reasoning-gpt-oss-20b-gguf`	`llama_cpp`	3	`lumynax-reasoning-gpt-oss-20b-mxfp4.gguf`
`lumynax-reasoning-internlm3-8b-gguf`	`AbteeXAILab/lumynax-reasoning-internlm3-8b-gguf`	`llama_cpp`	3	`internlm3-8b-instruct-Q4_K_M.gguf`
`lumynax-reasoning-phi4-mini-gguf`	`AbteeXAILab/lumynax-reasoning-phi4-mini-gguf`	`llama_cpp`	3	`Phi-4-mini-reasoning-Q4_K_M.gguf`
`lumynax-reasoning-qwq-32b-gguf`	`AbteeXAILab/lumynax-reasoning-qwq-32b-gguf`	`llama_cpp`	3	`qwq-32b-q4_k_m.gguf`
`lumynax-reranker-bge-v2-m3`	`AbteeXAILab/lumynax-reranker-bge-v2-m3`	`transformers`	3	`model.safetensors`
`lumynax-speech-kokoro-82m-tts`	`AbteeXAILab/lumynax-speech-kokoro-82m-tts`	`transformers`	3	`kokoro-v1_0.pth`
`lumynax-speech-whisper-large-v3-turbo`	`AbteeXAILab/lumynax-speech-whisper-large-v3-turbo`	`transformers`	3	`model.safetensors`
`lumynax-tiny`	`AbteeXAILab/lumynax-tiny`	`transformers`	2	`merged_model/model.safetensors`
`lumynax-tiny-qwen25-05b-gguf`	`AbteeXAILab/lumynax-tiny-qwen25-05b-gguf`	`llama_cpp`	3	`lumynax-tiny-qwen25-05b-q4_k_m.gguf`
`lumynax-translate-nllb-200-3b`	`AbteeXAILab/lumynax-translate-nllb-200-3b`	`transformers`	3	`pytorch_model-00002-of-00003.bin`

Routing behavior

Every routing decision passes through ordered gates. Models that fail any gate are rejected with a documented reason.

Gate	Rejects when
Modality match	Requested modalities are not a subset of model modalities
Context length	Model `context_tokens` is below `min_context_tokens`
Tool support	`requires_tools=true` but the model lacks tool support
JSON support	`requires_json=true` but the model lacks JSON support
License allowlist	License ID is not in the caller allowlist
Jurisdictional residency	`requires_local=true` and jurisdiction is not in model residency
Sovereignty tier	Data sensitivity requires a higher sovereignty tier

Surviving candidates are scored on:

Jurisdiction fit (+8)
Task-type tag match (+7, with +10 for coder specialization, +9 for reasoning)
Sovereignty bonus for iwi / data sovereignty keywords (+3 x tier)
Runtime preference (GGUF/llama.cpp gets +2.5)
Quality rank vs cost rank tradeoff

The router returns the winner plus the full rejection log so operators can see why each candidate did or did not qualify.

More commands

Local browser console and route API

bash

MaramaRoute serve --port 8787 --open

# Serve routed requests from pulled local GGUF models.
MaramaRoute serve --port 8787 --live-local --cache-dir ./models

The local server exposes:

GET /health
GET /v1/models
GET /v1/local/cache
GET /v1/local/health
POST /v1/route
POST /v1/chat/completions

Ask the router which model fits

MaramaRoute route --request examples/request.code-restricted.json

Inspect a single model

bash

MaramaRoute catalog --search starcoder --limit 5
MaramaRoute compare --model lumynax-coder-starcoder2-15b-gguf --model lumynax-coder-qwen25-coder-32b-gguf

Run the built-in route scenario matrix

MaramaRoute matrix

Generate agent and HPE/HPC helper configs

bash

# One command writes MaramaRoute config, aliases, agent bridge files, and HPE scaffold.
MaramaRoute setup qwen25-7b --all-targets --hpe --backend vllm

# Command bridge JSON for coding-agent workspaces.
MaramaRoute agent-config --target claude-code --model qwen25-7b
MaramaRoute agent-config --target codex --model qwen25-7b
MaramaRoute agent-config --target continue --model qwen25-7b
MaramaRoute agent-config --target opencode --model qwen25-7b
MaramaRoute agent-config --target litellm --model qwen25-7b
MaramaRoute agent-config --target tabby --model qwen25-7b
MaramaRoute agent-init --target claude-code --model qwen25-7b
MaramaRoute agent-init --target codex --model qwen25-7b
MaramaRoute agent init --target claude-code --model qwen25-7b
MaramaRoute agent doctor --target claude-code --model qwen25-7b

# HPE/HPC Slurm job script, live gateway config, backend launch, and run notes.
MaramaRoute hpe plan qwen25-7b --backend vllm
MaramaRoute hpe-job qwen25-05b --mode serve > marama-route.slurm
MaramaRoute hpe-job qwen25-7b --backend vllm --gpus 1 > marama-route.slurm
MaramaRoute hpe init qwen25-7b --backend vllm --gpus 1
MaramaRoute hpe init qwen25-7b --backend nim --backend-base-url http://127.0.0.1:8000/v1
MaramaRoute hpe init qwen25-7b --backend nemo --backend-command ./start-nemo-backend.sh
MaramaRoute hpe tunnel

# Generic local command bridge config.
MaramaRoute agent-config --target generic --base-url http://127.0.0.1:8787/v1

Audit receipts and registry maintenance

bash

MaramaRoute audit record --request examples/request.code-restricted.json
MaramaRoute audit list
MaramaRoute audit export --output marama-route-audit.json
MaramaRoute update-registry --dry-run
MaramaRoute update-registry --dry-run --diff

Emit an OpenCode provider config (drop into `~/.opencode/providers/`)

MaramaRoute opencode-config > ~/.opencode/providers/lumynax.json

Drive it from Python

python

from marama_route import (
    SovereignModelRouter,
    RoutingRequest,
    load_model_registry,
)
from pathlib import Path

models = load_model_registry(Path("./my_registry.json"))
router = SovereignModelRouter(models)

decision = router.route(
    RoutingRequest(
        prompt="Translate this paragraph to te reo Maori",
        task_type="general",
        jurisdiction="NZ",
        data_sensitivity="personal",   # routes only to sovereignty_tier >= 2
        requires_local=True,
    )
)

print(decision.selected_model.model_id)   # e.g. lumynax-translate-nllb-200-3b
print(decision.reasons)                    # rationale
print(decision.scores)                     # full scorecard

Why this exists

LumynaX is built by AbteeX AI Labs in Auckland, Aotearoa New Zealand. Three principles drive the design:

Sovereignty over convenience. Every routing decision can be justified to a Maori data-governance reviewer, a privacy officer, or an iwi advisory board. The registry, the routing log, and the policy gates exist for that conversation.
Local-first by default. Tier-3+ models run on machines the data owner controls. The router never silently escalates a sensitive request to a remote frontier model.
Open weights, open license, open evals. Apache-2.0 on this routing layer. Upstream model licenses surfaced honestly per entry. No vendor lock-in.

Every model card states its provenance. Every routing decision is auditable. Every sovereignty constraint is testable.

Companion products

abteex-sovereigncode - Policy API and audit ledger for coding agents. Pairs with MaramaRoute when you need per-request policy enforcement and tamper-evident logs.
LumynaX model family - 98 sovereign-tagged model repos on Hugging Face, all routable through MaramaRoute out of the box.
LumynaX release monorepo - the public release repo for MaramaRoute, SovereignCode, model scaffolds, Spaces, and publishing tooling.

Download stats

Live package counters:

The npm public API exposes package downloads by day/month and per-version downloads for the previous 7 days. It does not publish historical per-version-by-month downloads, so this diagram combines monthly package totals since first publish with the current all-version split.