Zacoda-Lite-1.1

Zacoda Lite 1.1

A small-footprint model built to genuinely reason, not just format nice-looking answers — a 4B model meant to run comfortably on modest hardware while still holding its own on real reasoning tasks.

Download

A 4-bit build didn't clear our own quality bar (garbled tokens and wrong arithmetic on repeated testing) — rather than ship something we don't trust, we held it back. The next series bakes in a lot more quantization-aware training from the start specifically so smaller quants hold up properly — expect a real 4-bit (and smaller) release then. No GGUF this round either — this checkpoint's per-layer embedding table isn't sized the way current llama.cpp Gemma4 support expects, so a converted file wouldn't actually load; not worth shipping something broken.

Benchmark — ZACBENCH-600

Our internal 6-category, 600-question benchmark (General Knowledge, Logic & Reasoning, Math, Coding, Instruction-Following, English). Full runs, thinking mode on, shown here against the rest of the current family for context.

*Plus 1.1 wasn't run through the full suite — Coding hit a non-terminating generation loop under greedy decoding that we haven't fixed yet, and English wasn't benchmarked at all, so its column isn't a fair apples-to-apples total. Shown anyway since the sections that do exist are genuinely good numbers.

Weaknesses

• Math is the clear weak spot at 31%. The most common pattern is the model working through the right method in its reasoning but slipping on the final arithmetic step — an execution error more than a conceptual one.
• Instruction-Following (58%) lags General Knowledge and English — multi-constraint prompts (exact formats, strict word counts) are where it trips up most.
• Logic (68%) is solid on its own but not a strength the way GK and English are.
• Worth being upfront about: Logic, Math, and Instruction-Following all sit meaningfully below Lite 1.0's numbers, even though Coding jumped a lot. Not hiding that — it's a real step back in a few areas alongside the real step forward in others, not an unqualified upgrade.

What's next

On Lite 1.1's regressions above trace back to a training data mix quality issue on our end for this release. We're sorry about that; it's not the bar we hold ourselves to, and we know it.

Lite 1.2 is being built on a reworked, quality-controlled data mix specifically to fix this — it's coming clean. Targeted for August 7. Zacagent Lite 1.2 (built for tool use and agentic tasks specifically) follows on August 10.

Usage

MLX

pip install mlx-lm

from huggingface_hub import snapshot_download
from mlx_lm import load, generate

# mlx_lm.load() doesn't take a `subfolder=` argument -- fetch the specific
# precision folder first, then load from the resulting local path.
path = snapshot_download("The-AI-makers88/Zacoda-Lite-1.1", allow_patterns=["mlx/8bit/*"])
model, tokenizer = load(f"{path}/mlx/8bit")
prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Hello!"}], add_generation_prompt=True
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=512))

PyTorch / transformers (8-bit, recommended)

pip install torchao

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("The-AI-makers88/Zacoda-Lite-1.1", subfolder="pytorch_8bit")
tokenizer = AutoTokenizer.from_pretrained("The-AI-makers88/Zacoda-Lite-1.1", subfolder="pytorch_8bit")

PyTorch / transformers (full bf16)

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("The-AI-makers88/Zacoda-Lite-1.1", subfolder="pytorch")
tokenizer = AutoTokenizer.from_pretrained("The-AI-makers88/Zacoda-Lite-1.1", subfolder="pytorch")

Help spread the word

If this model's useful to you, sharing it around — a repost, a mention, a link in a Discord or forum — genuinely helps more than almost anything else at this stage. I'm a solo dev without any budget, so word of mouth is most of what we've got.

---

Made by Sean Zhang

---

On benchmarks

Large labs sometimes train on the questions their models get benchmarked with — deliberately or through data that quietly overlaps. We can't audit anyone else's pipeline, so we won't make that claim about them. What we can do is show ours.

ZACBENCH-700 is our own evaluation set (700 questions across math, coding, general knowledge, logic, instruction-following, English, and agentic tool-use), and we checked — not assumed — that Zacoda was not trained on it: zero of its questions appear anywhere in our training corpus or distillation prompts. We also report GSM8K and HumanEval. GSM8K's train split is in our SFT data (standard practice — many labs do this); its test split is not, verified. HumanEval appears nowhere in training and is our cleanest independent signal.

Every model in any comparison here gets the identical token budget per section and the identical prompt — no model gets a bigger reasoning allowance than another. The full scoring code, every question, and the exact token limits are public: ZACBENCH-700. Run it yourself, against this model or any other.