Cactus 混合模型
小型端侧模型速度快且保护隐私,但有时会出错。在 Cactus,我们对模型进行后训练,使其能够识别自身何时出错:我们在检查点内部署探针,为每个答案打出 0 到 1 之间的置信度分数,并以结构化数据形式返回(绝不从答案文本中解析)。当置信度高时,在端侧作答;当置信度低时,你可以将请求重新路由到更大的模型:
if confidence < 0.85:
answer = ask_a_bigger_model(prompt)
我们从 Gemma 4 E2B Hybrid 开始推出,所有构建版本都托管在 Hugging Face 上的 Cactus Hybrid 集合中。
Gemma 4 E2B Hybrid 是 Gemma 系列中最小的模型,通过仅将 15–35% 的查询路由到 Gemini 3.1 Flash-Lite,其余部分自行运行,即可在大多数基准测试上媲美 Gemini 3.1 Flash-Lite。
| 基准测试 | 切换至 Flash-Lite(FP16)以匹配其表现 | 4-bit 量化下 | 3-bit 量化下 |
|---|---|---|---|
| ChartQA | 15–20% | 25–30% | 40–50% |
| MMBench | 30–35% | 40–45% | 50–55% |
| LibriSpeech | 25–30% | 35–40% | 55–65% |
| GigaSpeech | 30–35% | 40–45% | 50–55% |
| MMAU | 30–35% | 35–40% | 50–55% |
| MMLU-Pro | 45–55% | ~90% | 不适用 |
- 注:量化质量在 Cactus Quants 上测量,该模型在均匀量化下表现良好。
- 鼓励开发者自行针对 Unsloth、GGUF 和 MLX 量化进行基准测试。
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)
使用显式的 .to(device) 加载模型,而不是 device_map="auto":探针在模块 forward() 路径之外对生成结果进行评分,因此加速卸载的权重(留在 meta 设备上)会导致置信度读取崩溃。
llama.cpp
llama.cpp 是 C++ 编写的,因此探针是你编译进引擎的补丁(参见 patches/llama.cpp/)。构建一次打过补丁的服务器:
git clone https://github.com/cactus-compute/cactus-hybrid && cd cactus-hybrid
./patches/llama.cpp/install.sh && rehash
然后像任何 llama-server 一样启动并查询它——响应会携带一个顶层的 confidence 字段:
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}'
路由质量(AUROC)
Gemma 4 E2B Hybrid 的 AUROC 衡量模型区分错误答案与正确答案的能力(越高越好,0.5 为随机水平,1.0 为完美):
| 留出集 | 模态 | Cactus Hybrid | Token 熵 |
|---|---|---|---|
| MMLU | 文本多选题 | 0.770 | 0.697 |
| MMLU-Pro | 文本多选题 | 0.771 | 0.692 |
| ARC-Easy | 文本多选题 | 0.888 | 0.655 |
| ARC-Challenge | 文本多选题 | 0.834 | 0.646 |
| GSM8K(3-shot) | 文本生成 | 0.782 | 0.731 |
| MMBench-EN-Dev | 视觉多选题 | 0.840 | 0.435 |
| ChartQA | 视觉问答 | 0.779 | 0.615 |
| DocVQA | 视觉问答 | 0.781 | 0.512 |
| MMAU | 音频多选题 | 0.789 | 0.517 |
| GigaSpeech | 音频 | 0.876 | 0.343 |
| Earnings-22 | 音频 | 0.839 | 0.323 |
| LibriSpeech | 音频 | 0.822 | 0.427 |
| 平均值 | 0.814 | 0.549 |
最有力的结果是:该探针在零音频数据上训练,却在四个音频基准上取得了 0.79–0.88 的 AUROC 分数(两个转写任务、一个音频多选题、一个跨域转写任务)。
这排除了表面层面的解释——探针读取的是隐藏状态中与模态无关的正确性信号,而非记忆训练数据中的模式。
采用 MIT 许可证。Gemma 模型的使用需遵守 Gemma 条款。
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.
| Benchmark | Handoff to match Flash-Lite (FP16) | At 4-bit | At 3-bit |
|---|---|---|---|
| ChartQA | 15–20% | 25–30% | 40–50% |
| MMBench | 30–35% | 40–45% | 50–55% |
| LibriSpeech | 25–30% | 35–40% | 55–65% |
| GigaSpeech | 30–35% | 40–45% | 50–55% |
| MMAU | 30–35% | 35–40% | 50–55% |
| MMLU-Pro | 45–55% | ~90% | n/a |
- 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):
| Hold-out | Modality | Cactus Hybrid | Token Entropy |
|---|---|---|---|
| MMLU | text MCQ | 0.770 | 0.697 |
| MMLU-Pro | text MCQ | 0.771 | 0.692 |
| ARC-Easy | text MCQ | 0.888 | 0.655 |
| ARC-Challenge | text MCQ | 0.834 | 0.646 |
| GSM8K (3-shot) | text gen | 0.782 | 0.731 |
| MMBench-EN-Dev | vision MCQ | 0.840 | 0.435 |
| ChartQA | vision QA | 0.779 | 0.615 |
| DocVQA | vision QA | 0.781 | 0.512 |
| MMAU | audio MCQ | 0.789 | 0.517 |
| GigaSpeech | audio | 0.876 | 0.343 |
| Earnings-22 | audio | 0.839 | 0.323 |
| LibriSpeech | audio | 0.822 | 0.427 |
| Mean | 0.814 | 0.549 |
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.