LM Studio is a desktop llama.cpp wrapper with a local OpenAI-compatible
server. Ubuntu 24.04+ on bare-metal llama.cpp is still the measured AMDai
path —
/setup,
/gpus/r9700,
/benchmarks.
This page is the first-app path: Windows installer or Linux AppImage,
Vulkan runtime first, house
Qwen3.6-35B-A3B-UD-Q5_K_XL.gguf
at 262K. Everything here is
Cited, not measured on the local desktop.
We do not invent LM Studio versions, tok/s, or a bundled-ROCm GPU list.
llama.cpp first.
AMD’s playbook names both runtimes. Official LM Studio docs do not
publish a per-SKU Radeon table. RDNA 2, RDNA 3, RDNA 4, and the R9700
follow this site’s Vulkan-first rule until the picker lists this gfx for
ROCm.
Try this
Vulkan llama.cpp runtime — the first pick on Radeon
AMD’s LM Studio playbook says the app offers Vulkan and AMD ROCm backends (called runtimes) for AMD users. Vulkan is the honest first runtime on this site — the same rule as Vulkan vs HIP. LM Studio’s system-requirements page does not name RDNA 2, RDNA 3, RDNA 4, or the R9700. Open Runtime Selections with Ctrl+Shift+R (AMD playbook) or run lms runtime ls / get / select (LM Studio CLI). Download the Vulkan llama.cpp engine. Do not take a CUDA runtime on a Radeon. Watch a GPU meter: if it stays idle, you are on CPU. Not measured on the local desktop.
RDNA 2, RDNA 3, RDNA 4, and the R9700 (gfx1201) use Vulkan first on AMDai. That is this site’s llama.cpp rule, not an LM Studio SKU table — official LM Studio docs do not publish a per-card Radeon list.
Do not start here
Bundled ROCm llama.cpp runtime — second, and only if the picker lists this gfx
The same AMD playbook documents an AMD ROCm runtime next to Vulkan. gfx completeness of that bundled engine is what the picker shows — not a measured local desktop card, and not a whitelist we invent. We do not claim the bundled ROCm blob lists gfx1201 / R9700, gfx110x (7900 / 7800), or gfx1030 (official RDNA 2 row). If the ROCm runtime is missing, the card is absent, or the GPU meter stays idle, stay on Vulkan or move to plain llama.cpp on /setup. Do not invent an HSA_OVERRIDE on gfx1201. gfx1031 (RX 6700 XT) HSA_OVERRIDE is unofficial community note, not a recommendation — prefer Vulkan. Not measured on the local desktop.
RDNA 2 / 3 / 4 and R9700: only load the ROCm runtime if LM Studio’s picker lists this card’s gfx. Official docs do not publish that list here. Missing or idle → Vulkan, or /setup.
Zero to first token
1 · Cited, not measured on the local desktop
Windows install
Download the installer from lmstudio.ai/download and install. LM Studio’s system-requirements page: Windows x64 and ARM (Snapdragon X Elite); x64 needs AVX2; at least 16 GB of RAM recommended; at least 4 GB of dedicated VRAM recommended. AMD’s playbook says launch once so the lms CLI initializes. Adrenalin is AMD’s Windows driver path on that playbook — we do not invent a version pin. Official 35B cards on this site are Ubuntu llama.cpp, not Windows LM Studio. Not measured on the local desktop.
LM Studio documents Linux as an AppImage (x64 and ARM64). Required: Ubuntu 20.04 or newer. Official note: Ubuntu versions newer than 22 are not well tested. Ubuntu 24.04.5 is the measured AMDai llama.cpp OS — that is not an LM Studio test claim. AMD’s playbook: download the AppImage from lmstudio.ai/download?os=linux, then sudo apt install libfuse2, chmod +x LM-Studio-*.AppImage, ./LM-Studio-*.AppImage. Launch once so lms initializes. Add the login to render and video (AMD playbook). The playbook also mentions a .deb; this page follows the AppImage they spell out. Fedora and Arch can run the same AppImage as a cited try — not verified on our rig. Not measured on the local desktop.
Pick the Vulkan llama.cpp runtime — ROCm only if this gfx is listed
Ctrl+Shift+R opens Runtime Selections (AMD playbook). CLI: lms runtime ls, lms runtime get, lms runtime select. Take the Vulkan llama.cpp engine first on RDNA 2, RDNA 3, RDNA 4, and the R9700. The bundled ROCm engine is second, and only if the picker lists this card’s gfx — we do not invent that list. CUDA is the wrong family on a Radeon. Not measured on the local desktop.
House file: Qwen3.6-35B-A3B Unsloth UD-Q5_K_XL on one R9700 at 262144 context. Discover can search that name, or sideload the GGUF from NVMe. Start at Q6, Q5, or Q4. A cited Q3 file (UD-Q3_K_XL, Hugging Face listing 16.8 GB) is a labeled estimate — no 35B overnight card. Q1/Q2 stay off. Tok/s on /benchmarks and /results are Ubuntu llama-bench on llama.cpp, not numbers measured inside LM Studio. AMD’s playbook example models are not the house pick.
AMD playbook GUI: Model Loader (Ctrl+L), show advanced settings, GPU Offload maximum, Flash Attention On, Remember settings. Their playbook context example is 4096 for their sample models — that is not the house 262K window. LM Studio CLI: lms load <model_key> --gpu max --context-length 262144. --gpu accepts off, max, or a 0–1 fraction; if omitted, LM Studio picks. --context-length sets the token window. The calculator still prints Minimum system RAM needed (the load at the context you picked) and RAM needed at full context (262144 for the house 35B). Overnight UD-Q5 one-GPU Ubuntu llama.cpp peaks are 37.32 GB at 4K and 42.9 GB at 256K — measured on llama.cpp on the local desktop, not inside LM Studio. When a file spills to system RAM, round Minimum system RAM needed to a real kit (192 or 256 GB) with a breakdown and keep it a labeled estimate until measured. Not measured on the local desktop.
Local OpenAI-compatible server — then Hermes or another harness
Developer tab (Ctrl+2) → Server Settings → Status on (AMD playbook). Documented bind is typically http://127.0.0.1:1234. LM Studio OpenAI-compat docs: point the client base URL at http://localhost:1234/v1. Endpoints include GET /v1/models and POST /v1/chat/completions plus /v1/responses. Confirm with curl -s http://127.0.0.1:1234/v1/models. Keep the bind on 127.0.0.1 unless you mean to share the LAN (playbook: Serve on Local Network). Hermes and other harnesses can use that OpenAI-shaped /v1 the same way they use llama-server — house Hermes on this site talks to :8080 / Strata :8083, not a measured LM Studio attach. Not measured on the local desktop.
Leave LM Studio when you need this site’s compile flags, llama-bench cards, or a Stack Lab recipe that matches local desktop. The measured daily driver is llama.cpp Vulkan you built — llama-server on 127.0.0.1:8080, not the bundled engine on :1234. Read First-run setup and Backends. Docker / Podman is the isolated official-image path. We do not invent an LM Studio tok/s or a speed delta versus llama.cpp. Not measured on the local desktop.
GPU meter idle: wrong runtime (CPU or CUDA on a Radeon) or no offload — set GPU Offload max and confirm Vulkan. ROCm runtime missing this gfx: stay Vulkan; do not invent an override on gfx1201. gfx1031 HSA_OVERRIDE is unofficial — prefer Vulkan. AppImage on Ubuntu 24.04: LM Studio says versions newer than 22 are not well tested; libfuse2 is the AMD playbook package. Permission on /dev/dri or /dev/kfd: render and video groups. Start from /troubleshoot and paste a log in the error matcher. Cited rows live on /rocm-problems. Stack Lab remains the Ubuntu companion — packages are not published yet and there is no LM Studio package. Not measured on the local desktop.
Q6, Q5, and Q4 first. Q3 is a cited option labeled estimate. Speeds below
are
measured on llama.cpp directly, not inside LM Studio.
Q1/Q2 stay off.
35B GGUF quants for LM Studio on Radeon, with llama.cpp measurements labeled as not from LM Studio
Quant
Role
llama.cpp on the local desktop (not LM Studio)
Fit note
Q6_K
Higher-quality 35B sibling
Measured on llama.cpp Vulkan on the local desktop, not inside LM Studio: Q6_K 1-GPU 4K decode 103.23 tok/s (23:55 card). 32K and up is partly on DDR5 — that row says so.
Fit VRAM and both RAM lines (Minimum system RAM needed, RAM needed at full context) before you download. Not an LM Studio tok/s.
UD-Q5_K_XL
House default on one R9700 at 262K
Measured on llama.cpp Vulkan on the local desktop, not inside LM Studio: 102.39 tok/s decode at 4K, 90.52 at 32K, 66.43 at 128K, 44.4 at 256K. Peak VRAM 25.97 GB at 4K and 29.34 GB at 256K; peak system RAM 37.32 GB and 42.9 GB (Linux GB). Publisher listing 26.6 GB.
Daily-driver file Qwen3.6-35B-A3B-UD-Q5_K_XL.gguf. Set context to 262144, GPU offload max. Those tok/s are llama.cpp llama-bench cards, not LM Studio.
Q4_K_M
Smaller measured 35B sibling
Measured on llama.cpp Vulkan on the local desktop, not inside LM Studio: Q4_K_M 32K decode 117.69 tok/s; prefill 1902.1 tok/s. One R9700, f16 KV.
Still check Minimum system RAM needed and RAM needed at full context. Not an LM Studio tok/s.
UD-Q3_K_XL
Estimate
Cited Q3 option — labeled estimate
Hugging Face listing for Qwen3.6-35B-A3B-UD-Q3_K_XL.gguf (16.8 GB). No 35B overnight tok/s on this site. We do not invent an LM Studio Q3 speed.
Labeled estimate unless a run card exists — none does for 35B Q3 on the local desktop. Q1/Q2 stay off.
Cited CLI shape
After the first launch, lms ls prints the model key.
Offload and context flags are from the official lms load
page. Confirm the server with a loopback curl. Not measured on the local desktop.
# Cited shape — not measured on the local desktop.
# Replace <model_key> with the id from lms ls.
lms load <model_key> --gpu max --context-length 262144
curl -s http://127.0.0.1:1234/v1/models
RAM lines (llama.cpp-measured)
Use the calculator before you download a second quant. The two labels stay
Minimum system RAM needed
and
RAM needed at full context.
The house 35B peaks below are Ubuntu llama.cpp, not LM Studio.
If the app is not installed, packages are not published yet. There is no public download URL. Read the
Stack Lab install notes
.
If Stack Lab is already running on this machine,
open the local app
.