TheMindExpansionNetwork/MindBot-Ultra-27B-v0.1-GGUF

MindBot Ultra 27B v0.1 GGUF

[image]

GGUF export package for `TheMindExpansionNetwork/MindBot-Ultra-27B-v0.1`.

MindBot Ultra: Your Mind. Expanded. Your Vision. Amplified.

This repo packages the practical Ollama/llama.cpp builds for the newer Qwen3.6-derived MindBot Ultra 27B line, including a high-fidelity BF16 GGUF and a deployment-friendly Q4KM GGUF for agent-swarm serving.

Files

Recommended Ollama use

Use the Q4KM build for practical local/server use:

ollama create mindbot-ultra-27b:q4_k_m -f Modelfile.Q4_K_M
ollama run mindbot-ultra-27b:q4_k_m

BF16 is included as a high-fidelity archival/export GGUF and requires a very high-memory machine:

ollama create mindbot-ultra-27b:bf16 -f Modelfile.BF16

Modal + Ollama server endpoint pattern

[image]

The live Modal/Ollama deployment created for this package is locked to:

mindbot-ultra-27b:q4_k_m

Endpoint base URL:

https://m1ndb0t-2045--hermes-mindbot-ultra-ollama-fastapi-app.modal.run

Routes:

GET  /health
POST /v1/chat/completions
POST /api/chat

The chat routes require a bearer token and reject requests for any model name other than mindbot-ultra-27b:q4_k_m.

This deployment follows Modal's Ollama examples:

• —Modal guide: <https://modal.com/blog/howtorunollamaarticle>
• —Modal docs example: <https://modal.com/docs/examples/ollama>

Pattern:

1. 1.Build a Modal image with the latest Ollama install script.
2. 2.Mount a persistent Modal volume for OLLAMA_MODELS.
3. 3.Create exactly one locked Ollama model tag from this repo's Q4_K_M GGUF.
4. 4.Expose OpenAI-compatible /v1/chat/completions through a FastAPI guard.
5. 5.Lock requests so only mindbot-ultra-27b:q4_k_m is accepted.
6. 6.Let Modal scale the endpoint to zero when idle.

Agent-swarm routing note

For autonomous agents, point your OpenAI-compatible client at the deployed Modal Ollama endpoint and use only:

model = mindbot-ultra-27b:q4_k_m

Do not allow arbitrary model names unless you intentionally expand the allowed model list.

Evaluation reports

The repo includes a standard cross-model stress test folder:

evals/mindbot-ultra-27b-11q-self-training-eval.md
evals/mindbot-ultra-27b-11q-self-training-eval.json

This 11-question check asks every model the same identity, training-lineage, deployment, safety, synthetic-data, instruction-following, and wordplay questions so results can be compared across the Mindbotz model family.

Smoke test

A live endpoint smoke test completed successfully with a poem prompt about “how many hours are in the word strawberry.” Warm request result:

completion_tokens: 172
elapsed_seconds: 13.947
tokens_per_second: ~12.33

Sample output:

In the strawberry field of my mind, Where red neurons glow and sweet data unwind, I count the hours, one by one, In the seed of a word, a digital sun. ... The model is alive enough to rhyme, not alive enough to lie.

Conversion notes

• —Source model: TheMindExpansionNetwork/MindBot-Ultra-27B-v0.1
• —Base model: unsloth/Qwen3.6-27B
• —Architecture family: Qwen3.5 / Qwen3.6-derived causal language model
• —GGUF conversion: llama.cpp convert_hf_to_gguf.py
• —Q4KM quantization: llama.cpp llama-quantize
• —Quantization detail observed for Q4KM: ~4.92 BPW
• —Ollama templates included as Modelfile.Q4_K_M and Modelfile.BF16
• —License follows the source model card: Apache 2.0

Safety and scope

MindBot Ultra is a text-generation model. It can be wired into tools and agents, but tool execution should be separately permissioned, logged, and sandboxed. Keep high-impact actions approval-gated unless your deployment has its own safety layer.