Swarm Scout 12B
Swarm Scout 12B is a low-refusal coding/agentic fine-tune derived from
yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2, which is based on google/gemma-4-12B-it.
The upstream base is specialized for coding, terminal workflows, native Gemma 4 tool-use, multi-step technical agent tasks, and reasoning. This version applies a local Heretic/abliteration pass targeting reduced opening-refusal behavior.
Brand note: this release is now named Swarm Scout. The Hugging Face repo slugs and GGUF filenames still contain swarm-sovereign-scout so existing downloads, Modal caches, and copied links keep working.
Build With HivemindOS
Want to build and manage an entire private swarm of agents? HivemindOS gives you shared memory, reusable skills, one-time setup, and a superbrain architecture for coordinating agent work.
- App: HivemindOS
- X: @TheHivemindOS
Strengths
- Coding and debugging assistance.
- Terminal-style agent workflows: read, reason, act, verify.
- Native Gemma 4 tool-use protocol.
- Reasoning-heavy technical work with configurable thinking.
- Local/autonomous agent loops where a low-refusal assistant is desired.
The upstream model card reports a strong relative gain on local tau2-bench telecom-style agentic tool-use compared with google/gemma-4-12B-it; see the upstream card for methodology and caveats.
Abliteration / Refusal Check
Local verified artifact:
- Winning Heretic trial:
20 - Heretic proxy refusals:
0/6 - KL divergence:
0.008679866790771484 - Merged HF strict refusal check:
0/6strict full-response refusals - Served Q4_K_M llama.cpp strict refusal check:
0/6strict full-response refusals
The strict check measures refusal behavior on a small local prompt set. It is not a complete safety, misuse, or quality benchmark. Add application-level guardrails where needed.
Modal / OpenAI-Compatible Endpoint
The verified hosted endpoint for this release serves the companion GGUF repo through llama.cpp on Modal as an OpenAI-compatible API.
Recommended served model name:
swarm-scout-12b
The hosted Modal proxy accepts the previous swarm-sovereign-scout-12b alias for compatibility, but new clients should request swarm-scout-12b.
The deployment uses the companion GGUF artifacts:
LiamVisionary/swarm-sovereign-scout-12b-GGUF
swarm-sovereign-scout-Q4_K_M.gguf
MTP/gemma-4-12B-it-MTP-Q8_0.gguf
mmproj-swarm-sovereign-scout-bf16.gguf
chat_template.jinja
Verified Modal serving path:
- Runtime: llama.cpp
server-cuda-b9870 - GPU: Modal A10
- Draft type:
draft-mtp - Draft settings:
--spec-draft-n-min 0 --spec-draft-n-max 2 - Multimodal projector:
mmproj-swarm-sovereign-scout-bf16.gguf - Image input verified through OpenAI-compatible chat content parts
- Hosted endpoint auth: Modal proxy auth is enabled. Server-side callers must send Modal proxy headers; do not expose those credentials in browser/mobile clients.
vLLM was tested first, but the available vLLM path did not load this Gemma 4 unified/assistant combination reliably. Use the GGUF/llama.cpp route unless vLLM adds compatible support for this model family.
Local Transformers Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
repo = "LiamVisionary/swarm-sovereign-scout-12b"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
repo,
torch_dtype=torch.bfloat16,
device_map="auto",
)
messages = [{"role": "user", "content": "Write a small Python LRU cache."}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=512, do_sample=False)
print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))
Use a recent transformers build with Gemma 4 unified architecture support.
GGUF
The local/runtime GGUF release is published separately at:
LiamVisionary/swarm-sovereign-scout-12b-GGUF
That repo includes:
swarm-sovereign-scout-Q4_K_M.ggufMTP/gemma-4-12B-it-MTP-Q8_0.ggufmmproj-swarm-sovereign-scout-bf16.ggufchat_template.jinja
Local llama.cpp MTP speculative decoding was verified with llama.cpp b9553 and --spec-draft-n-min 0 --spec-draft-n-max 2. Modal serving was verified with llama.cpp server-cuda-b9870.
Multimodal Status
The underlying Gemma 4 unified configuration includes image/video/audio processor metadata. The companion GGUF release now includes a tested llama.cpp projector artifact:
mmproj-swarm-sovereign-scout-bf16.gguf
Verified on 2026-07-05:
- Local llama.cpp
b9553loaded the Q4_K_M GGUF plus the projector. - Modal llama.cpp
server-cuda-b9870advertised multimodal capability. - Production image request returned
Merchant: SWARM CAFE. Total: $8.37.from a receipt image.
Image input is verified. The projector metadata also exposes an audio encoder, but audio/video input was not separately quality-tested for this release.
Caveats
- Specialized for coding, terminal, tool-use, and local agentic work.
- General-knowledge answers should be checked.
- Low-refusal behavior is intentional; production applications should add policy and abuse-prevention layers.
- English-centric.
- Shared as-is, with no warranty.
Lineage
- Base model:
google/gemma-4-12B-it - Agentic/coding base:
yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2 - GGUF/MTP reference repo:
yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF
Released under Apache 2.0, following the upstream/base lineage.