The interesting question on a GPU with 16 GB of VRAM is not simply which GGUF fits. It is whether a model that is good at writing code is also good at choosing tools—and whether either result survives a realistic context length.
These are two distinct local experiments on a Windows machine with an AMD Radeon RX 9070 16 GB, Ryzen 5 7600, and approximately 32 GB of system RAM. They use different models, prompts, suites and reporting stages, so their scores should not be combined into one leaderboard.
1. A Qwen3-Coder 30B-A3B quality and context test
The first verified local study used Qwen3-Coder-30B-A3B Q3_K_L, llama.cpp b11146 with the Windows HIP backend / ROCm 10.0.0, one concurrent slot, temperature 0, and flash attention. A frozen 19-task local code suite was run with context 4096 and output budget 1024.
| Metric | Observed result |
|---|---|
| Local coding suite | 17/19 passed |
| Failed tests | Email whitespace normalization; HTTP header merge |
| Generated tokens over suite | 1,529 |
| Total runner wall clock | 18.25 seconds |
| Inference wall-clock sum | 16.02 seconds |
| Peak measured server dedicated VRAM | ~14,260 MiB |
The 19-task suite was purpose-built locally. It is not a public LiveCodeBench or SWE-bench score, and its small sample does not show reliable generalization. The result was nevertheless useful for checking that code was generated, parsed and executed against an independent hidden test suite without retries.
The separate context-length tests on the same model used prompt length approximately context minus 256 tokens, generated 128 tokens, flash attention and q4_0 K/V cache. All four runs completed:
| Context allocation | Decode throughput | Prefill throughput | TTFT |
|---|---|---|---|
| 8,192 | 83.295 tok/s | 1,801.897 tok/s | 4.436 s |
| 16,384 | 70.692 tok/s | 2,028.795 tok/s | 7.980 s |
| 24,576 | 59.439 tok/s | 1,935.418 tok/s | 12.613 s |
| 32,768 | 50.078 tok/s | 1,790.343 tok/s | 18.232 s |
These are measured llama.cpp decode/prefill rates for the stated synthetic context workload, not a speed guarantee for a tool-using agent. In particular, 50 tok/s at 32K means the 128-token decode completed at that measured rate; it does not mean the whole request started instantly.
2. The separate Qwen3.8-27B Q3 versus Q4 pilot
On the same RX 9070 host, a later study compared an incumbent Qwen3.8-27B Q3_K_M with a UD-IQ4_XS challenger. The matched exploratory pilot used a 16K context and 2,048-token output allowance and the same cases for both arms.
| Frozen local public-subset pilot | Q3_K_M | UD-IQ4_XS |
|---|---|---|
| LiveCodeBench | 5/12 | 2/12 |
| BigCodeBench-Hard | 5/11 | 4/11 |
| Berkeley Function Calling (BFCL) | 5/12 | 9/12 |
One BigCodeBench case was excluded symmetrically; neither model receives a zero for that excluded item. A prior narrow 15-case hard gate that stopped the challenger after three passes in seven attempts is a different test, not its full capability score.
This is precisely why there is no single winner yet:
- Q3 led on this small coding pilot, particularly LiveCodeBench.
- IQ4 led on this small function-calling pilot.
- Most case pairs tied, the samples are small, and the large same-case main panel had not reached a verified final result in the latest preserved source snapshot.
The pilot used locally frozen public subsets and pinned scoring rules. Its results are not an official leaderboard submission and should not be compared directly with leaderboard percentages using different revisions, generation budgets, or scoring rules.
3. What these experiments do—and do not—say
The 30B-A3B HIP context study and the 27B Q3/Q4 study are intentionally separated. The former answers whether a selected model can perform a defined local coding suite and sustain a long context; the latter probes a coding-versus-tool-use quantization tradeoff.
A bigger numerical score on one task family need not imply a better general assistant. A single tool call can fail because of templates, argument formatting, runtime handling, or underlying planning. We record complete versus attempted tasks, exact runtime, quantization, context, sample count, and independent evaluator together.
The RX 9070 system also has much more memory than an 8 GB MacBook Air. These results cannot be read as evidence that the same models run without swapping on the Mac.
Platform boundary: I also maintain RTX 5070 work on Linux. Those Linux GPU measurements are deliberately not presented as Windows data here. This page covers the Windows RX 9070 measurements only; additional Windows-only hosts will be added when there are verified results.
Reproducibility notes
The 30B-A3B study retained a SHA-256-pinned model and test suite, runtime identifier and separate per-task outcomes. The Q3/Q4 work retained matched case IDs, quantizations, and an operator-reported pilot pending broader validation. Raw local traces, paths, credentials and any restricted fixtures are excluded from the public article. Aggregate comparisons are sourced from the owner’s preserved technical records, not independently rerun for publication.
Related public benchmark projects: LiveCodeBench, BigCodeBench, and BFCL. These are methodology references, not endorsements of the local subset scores.
Updated 4 November 2026. Public generalization evaluation remains in progress; results and interpretations will be versioned instead of silently replacing prior measurements.