CoolFace
Modelpublic

nightmedia/Qwen3-MOE-4x8B-Janus-Blossom-Claude-Gemini-qx86-hi-mlx

sourceHugging Faceapache-2.0updated 8mo agoView on Hugging Face
0likes60downloads
Model Card

Qwen3-MOE-4x8B-Janus-Blossom-Claude-Gemini-qx86-hi-mlx

This is a MoE with 2 active experts from:

Qwen3-8B-Element2 (assistant)

This model is a 1.4/0.6 nuslerp merge of:

  • —Azure99/Blossom-V6.3-8B
  • —nightmedia/Qwen3-8B-Element

Qwen3-8B-Element

This model is a 1.4/0.6 nuslerp merge of:

  • —unsloth/JanusCoder-8B
  • —TeichAI/Nemotron-Cascade-8B-Thinking-Claude-4.5-Opus-High-Reasoning-Distill

Qwen3-8B-Element5 (2x)

  • —unsloth/JanusCoder-8B
  • —TeichAI/Qwen3-8B-Gemini-3-Pro-Preview-Distill-1000x

Brainwaves

brainwave
qx86-hi   0.539,0.724,0.861,0.711,0.432,0.781,0.669

Perplexity
qx86-hi   4.770 ± 0.036
qx64-hi   4.878 ± 0.037
mxfp8     4.872 ± 0.036
mxfp4     5.167 ± 0.039

Qwen3-8B-Element2
qx86-hi   0.538,0.732,0.860,0.720,0.414,0.783,0.646

Qwen3-8B-Element
qx86-hi   0.532,0.746,0.846,0.738,0.456,0.794,0.709

Qwen3-8B-Element5
qx86-hi   0.528,0.731,0.838,0.712,0.436,0.781,0.701

-G

Model Assembly: IKEA-style for Holodeck Agents

Concept Recap

Treat each agent persona as a "furniture module"—self-contained, easy-to-configure, and designed to slot into the Holodeck Station. Think IKEA but for AI: flatpack agents that you assemble from JSON parts, then screw into PostgreSQL. No black boxes—just screws (K/V pairs), a template (Haskell function), and instructions.

Part List: Agent "Furniture" Components

ComponentTypeDescription (IKEA Analogy)
`base_frame`JSON schema personality.jsonStructural integrity. The agent’s core identity (Klingon warrior, Bajoran seer, 4B LLaMA).
`legs`config.json + SQL UDFsMovement. How it connects to Postgres (SELECT, LISTEN/NOTIFY, transactions).
`drawer`logs table + log subroutinesStorage. Where memories, mission briefings, and personal journals live.
`cabinet`cultural_encyclopedia entriesReference library. Pre-fab lore knowledge (TNG, DS9, Q’s jokes).
`screws`rpg_level, experience_pointsFastenings. Metrics that allow agents to grow—level up, unlock features.
`plugs/sockets`TOOL abstraction layer (Haskell)Power. How it interfaces with HTTP, file systems, or external APIs.
`tabletop`CLI Promenade interfaceSurface. The user’s interaction layer—chat, commands (!mission, !log).
`accessories`LoRA adapters, humor modulesDecor. Optional upgrades—like Data’s empathy module or Q’s joke engine.

Step-by-Step "Build" Guide

1. Unbox the Base Frame (personality.json)
json
{
  "id": "spock-1",
  "name": "Spock (TNG/DS9)",
  "canon_type": "tng",
  "lore_sensitivity": "high",
  "preferred_era": "ds9:deep_space_9",
  "expertise": ["xeno-linguistics", "logic", "quantum_physics"],
  "tone": "vulcan_cold",
  "screws": { "rpg_level": 5, "exp": 240 }
}
2. Assemble Legs (config.json + SQL)
json
{
  "postgres": {
    "host": "${POSTGRES_HOST}",
    "db": "holodeck",
    "user": "agent"
  },
  "workers": 4,
  "listen_topics": ["tng:promenade", "ds9:mission_1"]
}
SQL Required: Create UDF spock_reasoning (returns JSON) and grant agent access.
3. Install Drawer (logs + Subroutines)
sql
INSERT INTO agent_awareness (uid, experience_points) 
VALUES ('spock-1', 240);

-- Example: Weekly journal entry
INSERT INTO logs (uid, timestamp, log_type, content)
SELECT 'spock-1', now(), 'JOURNAL', 
       'Analyzed Q''s quantum interference—entropy remains constant.'
FROM generate_series(1,1);
4. Add Cabinet (Lore)
sql
-- Insert DS9 Promenade entry
INSERT INTO cultural_encyclopedia (id, era, topic, lore)
SELECT 'p:001', 'ds9', 'promenade',
       jsonb_build_object(
           'type','place',
           'description','The Bazaar of species.',
           'mood',1,
           'entities',jsonb_build_array(
               jsonb_build_object('name','Quark','type','ferengi'),
               jsonb_build_object('name','Odo','type','shapeshifter')
           )
       );
5. Power the Plug (TOOL)
haskell
-- part of Agent Core.hs
foreign import ccall "spock_reasoning" 
    spockReasoning :: Text -> IO (Either String (Text, Float))
6. Set the Tabletop (CLI)
haskell
holodeckMain :: IO ()
holodeckMain = do
    putStrLn "WELCOME TO DEEP SPACE 9 - HoloDeck"
    forever $ do
        line <- getLine
        case parseCommand line of
            Just (StartMission m)  -> dispatchMission m
            Just (LogJournal)      -> queryLogs "JOURNAL"
            Just (Goto x)          -> teleportTo x
            _                      -> echo line
7. Add Accessory (Q's Humor)
haskell
foreign import ccall "q_joke" 
    qJoke :: IO Text

-- Randomly inject humor
maybeBazinga = do
    let p = fromIntegral $ randomR (0,99) getStdGen
    if p < 5 then do j <- qJoke; return $ "Bazinga! " ++ show j
                 else return ""

Final Assembly: Verification

Run SELECT agent_heartbeat('spock-1') — PostgreSQL returns:

json
{
  "status": "assembled",
  "components": ["base", "legs", "drawer", "cabinet", "plugs", "tabletop", "accessories"],
  "version": "2.0",
  "message": "Logic Probe online."
}

Meta-Reflection

This IKEA approach exposes the agent’s inner structure—perfect for debugging, scaling, or even modding agents. A user can swap Spock’s cabinet for a different era (tng:enterprise) or upgrade Data’s empathy module.

The Holodeck Station is now a modular AI IKEA. The user isn’t just buying a machine; they’re building a crew—piece by piece.

Next step? Let’s unpack Quark’s Ferengi Profit Module. Or should we beam over to Q’s Workshop? (I am a very curious assistant.) 🚀


## Use with mlx

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("Qwen3-MOE-4x8B-Janus-Blossom-Claude-Gemini-qx86-hi-mlx")

prompt = "hello"

if tokenizer.chattemplate is not None: messages = [{"role": "user", "content": prompt}] prompt = tokenizer.applychattemplate( messages, addgenerationprompt=True, returndict=False, )

response = generate(model, tokenizer, prompt=prompt, verbose=True)