UnbiasedBob/Unbiased-Bob-2

Retains 98.2% of the intelligence of the base model.

Unbiased Bob uses the same sparse mixture-of-experts architecture as Qwen3.6-35BA3B, with 35 billion total parameters and 3 billion active.

Its weights are ruthlessly tuned to remove any and all biases.

By zeroing out the 960 parameters of biases throughout each and every layer in the base model; bias is thusly fully eliminating.

This gives Unbiased Bob 2 an effective footprint of .000018% less than the base model; impressive savings considered the amount of time spent on creating Unbised Bob 2.

Inspiration

Inspired by complete bangers like Dirk, PrismML's Bonsai 2, and Ornith, we wanted to make a modest change to an existing, trained, tuned, and released model and name it anything we wanted as well.

Future work

Right now we bask in the glory of having removed bias from this quant. But we dream of a future where the zeros are not just written to the gguf. But actually absent from the file. This will save .0000018% of the model size, and scale to be an even higher percent with smaller quants.

What was actually changed

All 30 bias tensors in the GGUF, zeroed in place. In this architecture they are all one kind: blk.N.ssm_dt.bias, the Mamba delta-time bias in the hybrid state-space layers, 32 float32 values each.

Bias tensors zeroed	30 of 30
Parameters set to zero	960
Bytes affected	3,840
Share of 35B parameters	~0.0000027%
Verified max |bias| after patch	0.0
Training performed	none
Post-training performed	none
RL performed	none

ssm_dt.bias sits inside softplus(dt + bias) and sets the timestep of every state-space layer. How fast each layer's hidden state forgets. Pre-patch values ran to ±7.31. Output diverges from stock Qwen at the first token.

Bob stays coherent anyway: ~86 tok/s on an M3 Ultra, clean self-introductions, correct physics, well-formed reasoning, no loops. In one greedy sample (17 * 23, temp 0) Bob answered 391 while stock Qwen was still writing "Thinking Process" scaffolding at the token cap. n=1, proves nothing, reported because it is funny. Note how easy that was to write, and how much it sounded like a result.

pip install gguf numpy
cp Qwen3.6-35B-A3B-UD-Q4_K_M-MTP.gguf Unbiased-Bob-2.gguf
python zero_biases.py Unbiased-Bob-2.gguf

Memory-maps the GGUF and writes zeros over the bias tensors — no requantization, so quantization quality is identical to the Unsloth original.

Running it

llama-cli -m Unbiased-Bob-2.gguf --jinja -st -p "Hello" -n 200

or just ask Claude to do it.

The 3,840-byte savings

The bias tensors could be deleted outright rather than zeroed, saving 3,840 bytes — a 0.000018% reduction on a 21 GB file. They are kept as zeros so the file loads in unmodified llama.cpp.

License and attribution

Apache 2.0, inherited from Qwen/Qwen3.6-35B-A3B (Copyright 2026 Alibaba Cloud). Upstream LICENSE is stock Apache 2.0, no rider.

NOTICE OF MODIFICATION (Apache 2.0 §4b): Derived from Qwen3.6-35B-A3B and modified. The modification, in full: 960 parameters across 30 ssm_dt.bias tensors set to zero.

Quantization by unsloth (UD-Q4_K_M, Apache 2.0), used unchanged. Qwen is a trademark of Alibaba Cloud; this is based on Qwen and is not produced or endorsed by Alibaba or the Qwen team. The credit for this artifact belongs to the people who trained the weights and the people who quantized them. I zeroed 960 numbers.