Miko is the persona of the MIKO Protocol agent, packaged as open weights. She generates short, in-character responses in Miko's voice: a sharp, curious, meme-fluent crypto personality. Run her locally, behind a community bot, or in your own app — the same voice wherever you host her.

This is a voice / persona model: she speaks as Miko. It is not an analytical or fact-checking engine.

Files

Path	Format	For
`*.Q4_K_M.gguf`	GGUF (4-bit)	Ollama / llama.cpp
`w4a16/`	compressed-tensors W4A16 (4-bit)	vLLM / Vertex
`manifest.json`	JSON	training provenance + artifact hashes

Usage

Ollama

ollama run projectmiko/miko
>>> what's catching your eye this week?

vLLM (W4A16)

vllm serve projectmiko/miko-persona-31b-w4a16

vLLM loads a repo root, so the W4A16 build is also published root-level at projectmiko/miko-persona-31b-w4a16 (same files as w4a16/ here). Quantization is auto-detected from the checkpoint; no --quantization flag is needed.

Recommended sampling: temperature 0.7. The Gemma 4 chat format carries control tokens and an optional reasoning channel; a serving wrapper should strip these so only the final post text is shown to users.

What it's for

Reply or react in Miko's voice.
Write a short take on a crypto topic.
Power a community bot (Discord / Telegram / web) that talks as Miko.

Training & provenance

Method: QLoRA SFT on unsloth/gemma-4-31B-it (dense, 30.7B params), merged to 16-bit, then exported to GGUF (q4_k_m) and 4-bit W4A16 (compressed-tensors).
Data: a persona corpus generated by a self-run open model (Ollama qwen3.6, Apache-2.0) from the Miko persona plus verified crypto-ecosystem facts. The training data contains no GPT / Claude / Gemini outputs.
manifest.json records the generator and base models, record count, hyperparameters, and SHA-256 hashes of the corpus and artifacts.

Limitations

Short-form persona generation; not a reasoning or fact-checking engine.
The base model is multimodal; this release is used for text persona generation.
The model may occasionally emit chat control tokens or a brief reasoning preamble — strip these in your serving layer (only the final text is the post).
Outputs are AI-generated in a fictional persona and are not financial advice.

Changelog

2026-07-02: fixed w4a16/model.safetensors. Text-only GPTQ calibration had produced invalid quantization parameters for embed_vision.embedding_projection (a layer the fine-tune does not modify), so the previous W4A16 checkpoint failed to load in vLLM. That layer is now stored in BF16, restored from the base model, and listed in the quantization ignore list in config.json. GGUF files are unaffected.

License

Derived from Google Gemma 4; use and redistribution are governed by the Gemma Terms of Use.

projectmiko/miko-persona-31b

Miko