smtek/ BigBang-v1:IQ4_XS

56 19 hours ago

A 35B-A3B agentic model based on Qwen3.6-35B-A3B, designed for long‑horizon search, software engineering, scientific research, and AI research. 

ollama run smtek/BigBang-v1:IQ4_XS

Details

23 hours ago

04ad8977df4d · 19GB ·

model
qwen35moe
·
35.5B
·
IQ4_XS
template
{{ .Prompt }}
params
{ "num_ctx": 196608, "stop": [ "<|im_start|>", "<|im_end|>" ] }

Readme

Model Description

BigBang-v1 is a 35B‑A3B Mixture‑of‑Experts model built on Qwen3.6‑35B‑A3B. It has 35B total parameters but only 3B activated per inference, striking a great balance between capability and efficiency.

In 8 representative benchmarks covering long‑horizon search, software engineering, scientific reasoning, and AI research, BigBang‑v1 achieved the highest average score among selected 35B‑class models. It even outperformed DeepSeek V4 Pro Preview (1.6T) on FrontierScience Research, Humanity’s Last Exam, PaperBench(Code‑Dev), and BioMysteryBench‑HD.

Feature Details
Base model endless‑frontier/BigBang‑v1 (Qwen3.6‑35B‑A3B)
GGUF source bartowski/endless‑frontier_BigBang‑v1‑GGUF
Q4_K_M Good quality compared to full BF16 model, 64k context
Q4_K_S Recommended, almost as good as Q4_K_M, 96k context
IQ4_XS Smaller than Q4_K_S but worse at tool calling, 192k context
Hardware Runs comfortably on 24GB VRAM cards (e.g., RTX 3090⁄4090)

Recommended Optimization

By default, Ollama allocates the KV-Cache in f16, which will exceed 24GB VRAM over long contexts. To run stably, set these environment variables in your system (/etc/systemd/system/ollama.service.d/override.conf on Linux):

[Service]
Environment="OLLAMA_FLASH_ATTENTION=1"
Environment="OLLAMA_KV_CACHE_TYPE=q8_0"
  • Flash Attention: Dramatically reduces context memory footprint.
  • q8_0 Cache: Compresses context tensors without the severe logic degradation caused by 4-bit (q4_0) cache quantization.

Prompt Format

The model uses the ChatML format:

<|im_start|>system
{system_prompt}<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

Benchmark Results

Benchmark BigBang-v1 Description
BrowseComp 76.5% Long‑horizon search
xbench 58.4% Deep search
SWE‑Bench Pro 54.2% Software engineering
SciCode‑V‑Main 50.0% Scientific coding
FrontierScience‑Research 46.2% Scientific reasoning
Humanity’s Last Exam 50.3% General reasoning
BioMysteryBench‑HD 15.7% Biomedical reasoning
MLE‑Bench(Lite) 59.1% AI research
PaperBench(Code‑Dev) 53.6% AI research (code development)

Note: Evaluations used strict decontamination (e.g. filtering Hugging Face data) to ensure trustworthy results.

Links