The 8 GB RX 9060 XT has half the advertised memory of its 16 GB sibling. Its candidate Qwen3.5-9B Q4_K_M conversion uses a shorter catalog context to reduce the initial memory demand. That is a planning choice, not a demonstrated maximum context or proof of usable speed.
Check the memory reported free when inference starts, particularly if the card also drives a display. The model file, KV state and compute buffers all need space. If a task needs better answer quality or a longer conversation, compare the 16 GB variant’s distinct Q8 recipe only after testing both exact artifacts. This 8 GB page has no published physical result.
Qwen 3.5 models crashing trying to look at images on LMStudio Rocm
Method (as reported): Informal author report: image inference works with Vulkan and fails with ROCm; no measured benchmark.
Reported GPU setup RX 9060 XT 8GB
GPU count not reported
Operating system Linux (Linux Mint 22.3 Cinnamon 64-bit)
Engine LM Studio
Backend Vulkan versus ROCm
Model Qwen3.5-9B
Actual input input length not reported
Reported context Image/vision use reported; text-only behavior and occupied context unreported.
Topology not reported — One 8GB GPU model is named; total installed GPU count is not stated.
CPU AMD Ryzen 7 5700G
Memory configuration 32 GB DDR4 reported
Not reported: engine build, model file checksum, configured context, topology, system memory, LM Studio, ROCm, Vulkan and AMD driver versions, exact GGUF/quantization and memory placement, total installed GPU count, configured and occupied context, text-only behavior, concurrent clients and throughput
This is the RX 9060 XT 8GB, not the separately reported 16GB variant.
The author describes Linux LM Studio vision behavior, not FastLLM's current Windows text-only serving path.
One user report does not establish universal ROCm failure, full 8GB fit, or a performance recommendation.
DeepSeek-R1-Distill-Qwen
Bitworks
Bitworks: no recipe for this card yet.
Community
Windows · LM Studio Vulkan · by やかもち on Blog
Reported on Blog by やかもち — posted 2025-07-08 Original report
Historical RX 9060 XT 8GB Windows/Vulkan local inference review
Method (as reported): Author requested about 6000 output characters in an English gaming-PC prompt; prose reports 8B generation and 14B shared-memory slowdown.
Reported GPU setup RX 9060 XT 8GB
GPU count not reported
Operating system Windows (Windows 11 Pro 24H2, build 26100.3775)
Engine LM Studio
Backend Vulkan
Driver Radeon 25.6.2
Model DeepSeek-R1-Distill-Qwen GGUF
Actual input input length not reported
Reported context Older 8B/14B GGUFs; requested about 6000 output characters, not measured tokens.
Topology not reported — The ASRock 8GB card is identified; total installed GPU count and serving-device identity are unreported.
Offload other (Author describes 14B shared-memory slowdown; exact CPU/GPU placement is unknown.)
CPU Ryzen 7 9800X3D
Memory configuration 2 x 16GB DDR5-5600
Memory speed 5600 MT/s
Board ASUS TUF GAMING X670E-PLUS
Host notes ASRock Challenger 8GB OC was supplied by ASRock Japan; author says no comparison cards or software were specified.
Not reported: model quantisation, model file checksum, configured context, simultaneous clients, topology, system memory, GGUF artifact versions, quantizations and engine build, Context, prompt/output token counts, repetitions and numeric throughput, Concurrent serving, output correctness and physical VRAM placement
No numeric tokens-per-second result is retained from this review; requested characters are not output tokens.
The 14B slowdown is qualitative and must not be promoted to a rate or a measured VRAM threshold.
The reviewed models precede Qwen3.5/3.8 and do not establish FastLLM fit or performance.
RX 9060 XT 8GB is distinct from RX 9060 XT 16GB and RX 9060 non-XT.
Method (as reported): Author reports two coding-answer rates under LM Studio/Vulkan with maximum GPU offload selected; exact timing formula and repetitions unreported.
Reported GPU setup RX 9060 XT 8GB
GPU count not reported
Operating system Windows (Windows 11 24H2)
Engine LM Studio
Backend Vulkan
Driver AMD Adrenalin 25.10.09.01 DCH
Actual input input length not reported
Reported context Two different older models; configured and occupied context and request token counts are unreported.
Topology not reported — The XFX 8GB card is identified; total installed GPU count and serving-device identity are unreported.
Offload other (Author describes CPU/RAM offload for the 14B run; 8B placement is unproven.)
CPU Ryzen 7 9800X3D
Memory configuration 2 x 16GB DDR5-6400 C32
Memory speed 6400 MT/s
Board MSI MPG X870E CARBON WIFI
Host notes XFX Black Wolf 270 8GB; 2 x 16GB DDR5-6400 C32. Author says the 8B model occupied a little over half the 8GB VRAM.
49.73 tok/s — other (scope not stated) Historical DeepSeek-R1-0528-Qwen3-8B_Q4_K_M coding answer; author-reported rate, timing denominator unknown.
14.23 tok/s — other (scope not stated) Historical DeepSeek-R1-Qwen-14B-Q4_K_M coding answer with CPU/RAM offload; timing denominator unknown.
Not reported: model quantisation, model file checksum, configured context, simultaneous clients, topology, Engine build and exact GGUF artifact digests, Timing denominator, request counts, context, repetitions and variance, Concurrency, model correctness and physical serving-device/VRAM placement, Independent 8B VRAM residency and CPU/RAM offload amount for 14B
These are two different older DeepSeek/Qwen models; neither measures current Qwen3.5/3.8 or the proposed FastLLM 8GB recipe.
The author's roughly half-of-8GB VRAM occupancy description and maximum GPU-offload setting do not independently prove full model or operation residency.
The 14B result includes author-described CPU/RAM offload and is not an 8GB full-residency result.
The exact 8GB card must not be conflated with RX 9060 XT 16GB or RX 9060 non-XT.
Not yet covered by any approved report: No primary Qwen3.5-9B inference measurement has been verified for the RX 9060 XT 8GB. The reported 16GB variant result does not establish performance on this card.
Discussion
Share setups, questions and your own results. Posts are reviewed before they appear. Replies from Bitworks staff are marked.
Discussion
Share setups, questions and your own results. Posts are reviewed before they appear. Replies from Bitworks staff are marked.