Local Deployment of AI Models on the R9700 Suite
This article covers local deployment of AI models and a measured R9700-suite comparison between Swift 1.5 and the stock Qwen3.8-27B. The point is straightforward: on one machine, one server build, and one test suite on the stated date, Swift 1.5 finished the suite sooner and scored higher on MMLU-Pro and coding, while decode speed was essentially equal. That is a useful practical signal for local deployments, not a universal ranking.
Key Takeaway
In our R9700 suite, Swift 1.5 finished sooner than the stock Qwen3.8-27B Q4 run and scored better on MMLDU-Pro and coding, while GSM8K was equal. Decode speed was nearly the same: 60 versus 62 tokens per second. The most defensible conclusion is that this Swift 1.5 quant-and-drafter setup completed our suite faster; it is not a controlled proof that the fine-tune alone caused the difference, because the two GGUF recipes differ.
About Swift 1.5
UkisAI's Swift 1.5 card identifies Swift 1.5 as a reasoning-efficient derivative of Qwen3.8-27B. It is a 27B dense model, with the same broad Qwen hybrid language-and-vision architecture and a 262,144-token context configuration. The underlying Qwen weights are Apache 2.0; UkisAI's adapted weights and contributions use Swift Open License v1.0. That separate license permits free personal, research, educational, evaluation, and commercial use for organizations with up to US$1,000,000 gross annual revenue, including affiliates; above that threshold, commercial use requires enterprise terms. UkisAI says Swift 1.5 builds on Swift 1.0's post-training, scaling reinforcement learning and online preference distillation with a focus on long-horizon, agentic, and coding tasks. The card describes a training approach that penalizes tokens associated with pathological overthinking, then regains capability through RL and OPD. Its own evaluation reports 58.5% fewer median thinking tokens and a 0.35% higher aggregate score than the base on its chosen tasks, while individual task results vary. These are vendor claims under the card's own evaluation setup, not our R9700 test results.
Swift's card recommends temperature 1.0, top-p 0.95, top-k 20, min-p 0, presence penalty 0, and repetition penalty 1. It also says the model includes Qwen's MTP head and documents speculative decoding support.
R9700 Comparison Results
The suite used 400 MMLU-Pro questions, 50 GSM8K questions, and six coding tasks. We decoded at temperature 0, capped each response at 2,048 tokens, and ran four parallel slots. Counts in parentheses identify answers that reached the cap. Swift used a Q4_K_M GGUF with the stock Qwen MTP drafter at depth 2; the Qwen comparison used Unsloth UD-Q4_K_XLL and the same draft depth. They are close in nominal bit range but not the same quantization recipe.
Setup. AMD Radeon AI PRO R9700 (32 GB GDDR6); desktop AM5 platform, DDR5-6000, Windows 11; llama.cpp llama-server, Vulkan backend, build b11836; flash attention on; 131K context; ubatch 2048; four parallel slots sharing one unified KV cache. Tested October 2026.
| Model and run | MMLU-Pro | GSM8K | Coding | Decode | Suite wall time |
|---|---|---|---|---|---|
| Swift 1.5 Q4_K_M + stock MTP d2 | 89.5% (13 at cap) | 94% (0 at cap) | 6/6 | 60 tok/s | 97 min |
| Stock Qwen3.8-27B UD-Q4_K_XL + MTP d2 | 86.75% (31 at cap) | 94% (1 at cap) | 5/6 | 62 tok/s | 149 min |
Swift improved the observed MMLU-Pro score by 2.75 points and completed one more coding task; GSM8K was the same. The Qwen row recorded one more GSM8K answer at the cap. Decode rate favored the stock Qwen by 2 tok/s, a small difference in these measurements, while the complete suite took 52 fewer minutes on Swift. A suite wall time combines all prompts and tasks; it is not the same as per-token decode speed. This is also a comparison between two model packages, not a perfectly isolated fine-tune experiment. Swift used Q4_K_M; stock Qwen used UD-Q4_K_XL. Quantization can affect both runtime and task scores. We did not measure Swift on MLX, and we did not test the two models with one identical quantization method in this suite. The vendor's "fewer thinking tokens" claim and our shorter suite runtime point in the same practical direction, but only our run is direct evidence on this R9700 and suite.
Recommended Run Settings
For reproducing these measurements, use Swift Q4_K_M with MTP depth 2 and the standardized suite settings above. The card's generation recommendation is temperature 1.0, top-p 0.95, top-k 20, min-p 0, presence penalty 0, and repetition penalty 1. For an apples-to-apples local comparison, keep context, prompt template, output cap, parallelism, and quantization recipe fixed across Swift and stock Qwen, then record the exact model revision and runtime. Our results do not establish how much of the wall-time gap comes from the fine-tune versus the quant recipe. Swift includes the base model's MTP head, but our measured comparison uses a separate stock Qwen drafter at depth 2. In a separate short decode sweep, Swift accepted 82% of proposed tokens at depth 2. Our sweeps found depth 2 faster than deeper draft settings for the tested 27B models; use it as the starting point on this machine rather than assuming a higher draft depth will be faster. Always recheck output quality on the tasks that matter to you.
Verdict
Swift 1.5 is a promising option for a local coding and general reasoning workflow where finishing a batch sooner matters. On the tested package, it achieved higher MMLU-Pro and coding results than stock Qwen Q4 and cut suite wall time, despite essentially equal decode speed. The quant recipes differ, so the next useful controlled comparison would hold quantization constant. For now, the result supports choosing the tested Swift package as a practical option, not a universal model ranking.
Related reading: "MiniCPM-V 4.7 on the R9700: a fast multimodal MoE, with a DSpark caveat"; "Sage-A1.5 on the R9700: why thinking mode changed the score"; "Qwen3.8-27B on the R9700: MTP depth and quantization trade-offs"; "Qwen3.8-27B GSQ-RCO IQ3_S: 11.3 GiB and a useful quality floor"; and "Flint 40B-A2.5B on the R9700: fast decode, longer work".
Disclaimer. These are measurements from one machine, one server build, and one test suite on the stated date. Results vary with hardware, software version, settings, and load. They are not a general ranking or a recommendation.
