Simple-Attention-Sparsification on local hardware
Estimated comfort for repository coding with Aider on a MacBook Pro M4 Pro 48GB.
Simple-Attention-Sparsification
tencent · See model card · 0 downloads
Reasoning / thinking mode
- Parameters
- 4.0B estimated
- Context window
- 40K tokens
- Attention
- Standard (full attention)
- Layers · KV heads · head dim
- 36 · 8 · 128
- KV cache
- 144 KB/token · 4.5 GB at 32K
- Tool calling
- good (from chat template)
- Thinking
- good (from chat template)
- Vision
- basic (inferred — no supporting signal)
- Commercial use
- check cardCheck the model card
- Import confidence
- low · capabilities estimated
Weights by quantization
Q3_K_M
1.9 GB
computed · low confidence
Q4_K_M
2.3 GB
computed · low confidence
Q5_K_M
2.7 GB
computed · low confidence
Q6_K
3.2 GB
computed · low confidence
Q8_0
4.1 GB
computed · low confidence
F16 / BF16
7.7 GB
computed · low confidence
- • Architecture read from Qwen/Qwen3-4B.
- • Parameter count estimated from the architecture (the Hub reports none for this repo).
- • Assumes GGUF / MLX conversions of this model exist (they do for most popular models). Sizes are computed from bits per weight.
- • vision: basic, inferred from missing signals (low confidence).
- • Capability ratings are estimated from model size and name — this model hasn't been assessed by hand.
- • Commercial use unknown: Check the model card.
Simple-Attention-Sparsification — 8-bit
Borderline for your workload
Speed
25–35 tok/s
Comfortable
Memory left
27 GB
of 48 GB
First word in
≈25 s
Comfortable
All the details
- Your setup
- MacBook Pro M4 Pro 48 GB
- Workload
- Repository coding with Aider
- Runtime → tool
- LM Studio → Aider
- Memory used
- 9.6 GB
- Context
- 32K · up to 32K practical
- Tool support
- Official
- Use-case fit
- Poorly suited
Why?
- Model fits with approximately 27 GB of memory headroom.
- LM Studio supports hardware acceleration on this machine (Metal (Apple GPU)).
- Aider can connect to LM Studio (officially supported).
- Generation (25–35 tok/s) should be appropriate for interactive coding.
- Ingesting a prompt of 14K tokens takes ≈25 s; prompt processing keeps up with interactive coding.
- 32K window; generation drops to 19–25 tok/s when it is full. KV cache is large (≈4.5 GB).
- Weak coding capability for interactive coding.
Why “Borderline” rather than “Acceptable”?
- Workload suitability: Weak coding capability for interactive coding.
- Context: 32K window; generation drops to 19–25 tok/s when it is full. KV cache is large (≈4.5 GB).
How sure are we?
Medium confidenceEstimated from bandwidth and compute, calibrated against 2 verified benchmarks of other models on this same chip (measured with a different engine, so calibration is partial).
- Calibrated with verified benchmarks of other models on this chip.
- Some specifications come from lower-confidence sources.
Medium confidence