Same weights as the safetensors repo, repackaged as a GGUF for toolchains that can't consume HF transformers directly. Quant label in the GGUF general.file_type header is Q4_1 (the rocmfp4-llama fork's universal default — not to be trusted), but the actual weight tensors are F16. Match the 2.0 GB file size to confirm.

How hal0 uses this file

This is the brain slot's recommended load in installer/etc-hal0/slots/brain.toml:

toml

[model]
default = "hal0-brain-sft"        # operator assigns one of the hal0ai/ GGUF variants
context_size = 65536

After hal0 model assign hal0-brain-sft-GGUF --slot brain && hal0 slot load brain, the GGUF is loaded by hal0-slot@brain.service (port 8089, GPU-Vulkan llama-server container, profile chat) and fronts the hal0/brain virtual model.

The brain is provisioned by hal0 API lifespan (ensure_brain_provisioned in src/hal0/brain/provision.py) and drives the shared toolloop.engine.run_tool_loop to serve /api/brain/chat (and the legacy /api/board/chat alias). Tool turns are routed off-slot to hal0/code or hal0/agent per [brain_chat].tool_model.

Usage — llama.cpp directly

bash

# CLI
llama-cli -m model.gguf -p "What does the hal0 'agent' slot default to?" -n 256 --jinja

# Server (OpenAI-compatible, what hal0-slot wraps)
llama-server -m model.gguf --jinja -fa --port 8080

Usage — Ollama

bash

# From a Modelfile
FROM ./model.gguf
PARAMETER stop "<|im_end|>"

ollama run hal0-brain-sft "What does the hal0 'agent' slot default to?"

Usage — hal0-slot (the official wiring)

bash

# 1. Pull into hal0's local model store
hal0 model pull Hal0ai/hal0-brain-sft-GGUF

# 2. Assign to the brain slot (default port 8089, GPU-Vulkan)
hal0 model assign hal0-brain-sft --slot brain

# 3. Start the slot container
hal0 slot load brain

# 4. Verify it's serving
curl http://127.0.0.1:8089/v1/models

Then in hal0.toml:

toml

[brain_chat]
enabled = true
model = "hal0/brain"
tool_model = "hal0/agent"          # OR "hal0/code" — routes tool turns
max_rounds = 8
completion_timeout_s = 300.0

Model details


Base model	`Hal0ai/hal0-brain-sft` (BF16 safetensors)
Relation	quantized (same weights, GGUF container)
Architecture	LlamaForCausalLM, 1.5B params
Dtype / quant	F16 (tensors)
SHA-256	`ed9d28c4eac1d7c291bc80d9410c243a3d28e655921ccaf90f2b6619aa24d2c3`
File size	2.0 GB
License	Apache-2.0

Hal0ai/hal0-brain-sft — BF16 safetensors (HF transformers)
Hal0ai/hal0-brain-sft-fpx8-agent — Q8_0 GGUF (agent slot)
Hal0ai/hal0-brain-sft-fpx4-agent — ROCmFP4 GGUF (AMD Strix)
Collection: hal0-brain

License

Apache-2.0.

Provenance

Repo: Hal0ai/hal0-brain-sft-GGUF
Org: Hal0ai
Project: https://github.com/Hal0ai/hal0
Brain spec: docs/rework/hal0-specs/spec-p3-brain.final.md

Hal0ai/hal0-brain-sft-GGUF

hal0-brain-sft (GGUF F16)

How hal0 uses this file

Usage — llama.cpp directly

Usage — Ollama

Usage — hal0-slot (the official wiring)

Model details

Related

License

Provenance

Hal0ai/
hal0-brain-sft-GGUF