Downloads · 30 days
46
11% of all-time downloads
gemma-4-26B-A4B-it-oQ3-mtp
mlx-works/gemma-4-26B-A4B-it-oQ3-mtp
gemma-4-26B-A4B-it-oQ3-mtp is a machine learning model from mlx-works. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for mlx.
This model was quantized using oQ (oMLX v0.5.4) mixed-precision quantization.
All-time downloads
413
Public
Parameters
26.2B
13.5 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors13.5 GB · 100%
Parameter types
How the weights are stored.
U3225.2B · 96%
Model card
From the Hugging Face model README
gemma-4-26B-A4B-it-oQ3-mtp
This model was quantized using oQ (oMLX v0.5.4) mixed-precision quantization.
Base model: google/gemma-4-26B-A4B-it
MTP head: google/gemma-4-26B-A4B-it-assistant (merged into single checkpoint)
Quantization details
- Model type: gemma4
- Bits: 3
- Group size: 64
- Format: MLX safetensors
- MTP: Merged from assistant checkpoint (Lightning MTP)
Environment
- Hardware: M5 MacBook Air 32GB
- Inference Framework: oMLX v0.5.4
- Max Concurrent Requests: 4
- Settings:
- Thinking: Disabled
- Lightning MTP: Enabled (key speed improvement)
Performance Benchmarks
Note: Results are for reference only and may vary depending on hardware, software configuration, and workload.
Single Request Results
| Test | TTFT(ms) | TPOT(ms) | pp TPS | tg TPS | E2E(s) | Throughput | Peak Mem |
|---|---|---|---|---|---|---|---|
| pp1024/tg128 | 2375.6 | 39.40 | 431.0 tok/s | 25.6 tok/s | 7.404 | 155.6 tok/s | 11.44 GB |
| pp4096/tg128 | 9573.1 | 42.42 | 427.9 tok/s | 23.8 tok/s | 15.022 | 281.2 tok/s | 11.84 GB |
Continuous Batching (pp1024 / tg128)
| Batch | tg TPS | Speedup | pp TPS | pp TPS/req | TTFT(ms) | E2E(s) |
|---|---|---|---|---|---|---|
| 1x | 25.6 tok/s | 1.00x | 431.0 tok/s | 431.0 tok/s | 2375.6 | 7.404 |
| 2x | 33.3 tok/s | 1.30x | 396.6 tok/s | 198.3 tok/s | 5164.0 | 12.850 |
| 4x | 38.5 tok/s | 1.50x | 386.3 tok/s | 96.6 tok/s | 10158.3 | 23.907 |
Intelligence Benchmark
Note: Each benchmark round tests only 30 questions. Results are for reference only.
| Benchmark | Accuracy | Correct | Total | Time(s) | Think |
|---|---|---|---|---|---|
| MMLU | 60.0% | 18 | 30 | 75.5 | No |
| TRUTHFULQA | 90.0% | 27 | 30 | 32.5 | No |
| GSM8K | 83.3% | 25 | 30 | 298.7 | No |
| MATHQA | 10.0% | 3 | 30 | 116.5 | No |
| HUMANEVAL | 90.0% | 27 | 30 | 474.6 | No |
- Library
- mlx
- Type
- gemma4
- Created
- Aug 2, 2026
- Updated
- Aug 4, 2026