Size: 9.7 GB (smaller than the 31B dense at 10.4 GB)
BPW: 3.29
Architecture: Gemma 4 MoE — 128 experts, 8+1 active per token, 25.2B total / 3.8B active

Benchmark results (reasoning OFF)

Note on BFCL scores: All BFCL scores reported here use our internal simplified evaluation (single-function-call subset with custom prompt/scoring), NOT the official Berkeley Function Calling Leaderboard methodology. Our scores are not directly comparable to the official leaderboard. We are working on running the official BFCL evaluation for comparable numbers.

Model	Size	Active	HumanEval+	MBPP+	BFCL v3	NL2Bash F1
Gemma 4 31B IQ2_M	10.4 GB	31B	88.41%	82.01%	92.25%	84.71%
This model (26B MoE IQ2_M)	9.7 GB	3.8B	80.49%	77.78%	92.75%	79.64%
Qwen3.6 IQ2_M	11.1 GB	3B	80.49%	78.31%	94.75%	81.63%
Gemma 4 E4B Q8_0	7.8 GB	4.5B	73.78%	73.28%	93.75%	79.75%

Key findings

BFCL 92.75% — beats the 31B dense (92.25%) with 0.7 GB less and 8x fewer active params
HumanEval+ 80.49% — ties Qwen3.6 IQ2_M at comparable size
MoE expert isolation provides natural quantization resilience, as predicted by our specialization boundary study
All benchmarks run with reasoning OFF; thinking mode expected to add +5-15pt on code tasks

Quickstart

bash

huggingface-cli download KikoCis/gemma-4-26B-A4B-it-IQ2_M-GGUF gemma4-26b-a4b-IQ2_M.gguf --local-dir .

# Without thinking
llama-cli -m gemma4-26b-a4b-IQ2_M.gguf -ngl 99 --ctx-size 8192 --temp 0.1 \
  -p "Write a Python function to merge two sorted lists"

# With thinking (recommended for complex tasks)
llama-cli -m gemma4-26b-a4b-IQ2_M.gguf -cnv -ngl 99 --ctx-size 8192 \
  --reasoning on --reasoning-budget 1024

Files

gemma4-26b-a4b-IQ2_M.gguf — quantized weights (9.7 GB)
gemma4-26b-a4b-domain.imatrix — importance matrix

License

Apache 2.0.

Real-World Agent Test Warning (April 2026)

Benchmark scores do not predict agent capability. In Docker-based autonomous testing, fine-tuned E4B models (95% BFCL) scored 0/10 while the unfine-tuned base scored 6/10. Fine-tuning for BFCL destroyed general reasoning (error recovery, strategy adaptation, anti-repetition). Fine-tuned E4B models have been withdrawn.

For autonomous agent tasks, use the base Gemma 4 model or a larger model at higher BPW. See: The Benchmark Trap — Full Study

KikoCis/gemma-4-26B-A4B-it-IQ2_M-GGUF

Gemma 4 26B-A4B IT — IQ2_M (domain imatrix)

Benchmark results (reasoning OFF)

Key findings

Quickstart

Files

Related

License

Real-World Agent Test Warning (April 2026)

KikoCis/
gemma-4-26B-A4B-it-IQ2_M-GGUF