Model × hardware
Can Mac Studio (2025) M3 Ultra (32-core CPU / 80-core GPU) 256GB run deepseek-coder-6.7b-instruct?
deepseek-coder-6.7b-instruct at Q4_K_M fits the estimated Mac Studio (2025) M3 Ultra (32-core CPU / 80-core GPU) 256GB memory budget at 4,096 tokens: 6.80 GiB required versus 192.00 GiB usable. 80.2–160.5 estimated tokens/s.
Compare the recorded formats.
| Quant | Weight files | Estimated KV | Estimated total | Fit | Estimated generation |
|---|---|---|---|---|---|
| Q4_K_M | 3.80 GiB | 2.00 GiB | 6.80 GiB | fits | 80.2–160.5 estimated tokens/s |
| Q5_K_M | 4.46 GiB | 2.00 GiB | 7.46 GiB | fits | 68.5–136.9 estimated tokens/s |
| Q8_0 | 6.67 GiB | 2.00 GiB | 9.67 GiB | fits | 45.7–91.5 estimated tokens/s |
What changes with a longer context?
| Context tokens | Estimated KV | Estimated total | Fit |
|---|---|---|---|
| 2,048 | 1.00 GiB | 5.80 GiB | fits |
| 4,096 | 2.00 GiB | 6.80 GiB | fits |
| 8,192 | 4.00 GiB | 8.80 GiB | fits |
| 16,384 | 8.00 GiB | 12.80 GiB | fits |
The formula for this estimate.
Weights GiB = exact recorded file bytes / 2^30. KV GiB = 2 × layers × KV heads × head dimension × context × KV bytes / 2^30. Required GiB = weights + KV + runtime overhead. Usable GiB = device GiB × usable fraction. Estimated decode tokens/s = bandwidth GB/s × 10^9 / (active parameters B × 10^9 × effective bytes per weight) × efficiency. Effective bytes per weight = file bytes / (total parameters B × 10^9). Selected efficiency = 0.6; sensitivity uses 0.4–0.8.
Assumptions and limitations
- Memory values labeled GB in the data contract are GiB (2^30 bytes); bandwidth uses decimal GB/s (10^9 bytes/s).
- The memory budget reserves 25.0% of device memory. Runtime overhead is an assumed 1 GiB.
- KV cache assumes 2 bytes per element, independently of weight quantization. Runtime cache layout and allocation can add memory.
- Memory fit does not establish runtime, driver, operating-system, architecture, or quantization support.
- All speed values are estimated single-stream decode proxies, not benchmarks. Prefill, batching, latency, and concurrent sessions are not predicted.
- Estimated speed sensitivity is not a confidence interval; actual speed can fall outside the range because of compute, cache traffic, kernels, thermal limits, and other workloads.
- Effective bytes per weight is the exact file-size / total-parameter proxy. Metadata, tensors with different precision, padding, and duplicated tensors make it different from nominal quantization bits.
Source records
Official model metadata · Exact weight files · Maker hardware specifications
Model verified 2026-10-22; hardware verified 2026-10-22. This is an estimated memory and bandwidth calculation, not a hands-on benchmark.