cactus-compute/cactus-hybrid

Source: https://github.com/cactus-compute/cactus-hybrid

Cactus Hybrid

A small, on-device model is fast and private, but sometimes wrong. At Cactus we post-train models to know when they are wrong: we ship probes inside the checkpoint that score every answer with a confidence between 0 and 1, returned as structured data (never parsed out of the answer text). Answer on-device when confidence is high; you can re-route to a bigger model when it’s low:

if confidence < 0.85:
    answer = ask_a_bigger_model(prompt)

We start the rollout with Gemma 4 E2B Hybrid, all builds live in the Cactus Hybrid collection on Hugging Face.

Gemma 4 E2B hybrid, the smallest Gemma model, matches Gemini 3.1 Flash-Lite on most benchmarks by routing only 15–35% of queries to the Gemini 3.1 Flash-Lite and running the remnant itself.

• N/B: Quantisation quality is measured on Cactus Quants which performs well at uniform quantization.
• Developers are encouraged to benchmark for Unsloth, GGUF, and MLX quantization independently.

Cactus

# pip install cactus-compute
import json
from cactus.bindings.cactus import cactus_complete, cactus_init
from cactus.cli.download import download_bundle

lm = cactus_init(str(download_bundle("Cactus-Compute/gemma-4-E2B-it")))
result = cactus_complete(
    lm,
    [{"role": "user", "content": "What is the capital of France?"}],
    json.dumps({"max_tokens": 512, "auto_handoff": False}),
    None,
    lambda *_: None,
)
print(result["response"].strip())
print("confidence:", result["confidence"])

MLX

# pip install mlx-lm
import re
from mlx_lm import load, generate

model, tokenizer = load(
    "Cactus-Compute/gemma-4-e2b-it-hybrid-mlx",
    tokenizer_config={"trust_remote_code": True},
)

messages = [{"role": "user", "content": "What is the capital of France?"}]
answer = generate(
    model,
    tokenizer,
    prompt=tokenizer.apply_chat_template(messages, add_generation_prompt=True),
    max_tokens=512,
)
# the checkpoint reasons before answering; keep only the final answer
answer = re.split(r"<\|?channel\|?>", answer)[-1]
answer = re.sub(r"^(thought|final)\b\s*", "", answer).strip()
print(answer)
print("confidence:", model.last_confidence)

Transformers

# pip install "transformers>=5.5.4,<5.6" torch   (5.14+ segfaults on this checkpoint)
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Cactus-Compute/gemma-4-e2b-it-hybrid"
device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True, dtype="auto").to(device)

messages = [{"role": "user", "content": "What is the capital of France?"}]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt", return_dict=True
).to(device)
out = model.generate(**inputs, return_confidence=True, max_new_tokens=512)

print(tokenizer.decode(out.sequences[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
print("confidence:", out.confidence)

Load the model with an explicit .to(device), not device_map="auto": the probe scores generations outside the module forward() path, so weights that accelerate offloads (left on the meta device) crash the confidence read.

llama.cpp

llama.cpp is C++, so the probe is a patch you compile into the engine (see patches/llama.cpp/). Build the patched server once:

git clone https://github.com/cactus-compute/cactus-hybrid && cd cactus-hybrid
./patches/llama.cpp/install.sh && rehash

Then serve and query it like any llama-server — the response carries a top-level confidence field:

llama-server -hf Cactus-Compute/gemma-4-e2b-it-hybrid-GGUF:Q4_K_M --jinja

curl -s http://localhost:8080/v1/chat/completions \
  -d '{"messages":[{"role":"user","content":"What is the capital of France?"}],"max_tokens":512}' \
  | jq '{answer: .choices[0].message.content, confidence}'

Routing Quality (AUROC)

Gemma 4 E2B Hybrid AUROC measures how well the the separates wrong answers from right ones (higher = better, 0.5 is random, 1.0 is perfect):

The strongest result: the probe was trained on zero audio data, yet achieves 0.79–0.88 AUROC on four audio benchmarks (two transcription, one audio MCQ, one out-of-domain transcription).

This rules out surface-level explanations, the probe is reading a modality-independent correctness signal from the hidden state, not memorizing patterns from training data.

MIT-licensed. Gemma model use is subject to the Gemma terms.