LM Studio Qwen3.5 Failed to Load: Fix 'unknown model architecture: qwen35'

What does 'unknown model architecture: qwen35' actually mean?

When LM Studio opens a GGUF it first reads metadata, including general.architecture. Dense Qwen3.5 GGUFs may store qwen35 there. If the selected llama.cpp engine has no loader registered for that key, loading stops before the model can be initialized normally.

The popup title 'Failed to load model' is generic; the diagnostic line is the architecture error underneath it. 'Unknown' does not automatically mean the GGUF is corrupt. It means this engine does not know how to instantiate that model architecture.

The same symptom can appear on Windows or Linux and with different GPU vendors because it often occurs before meaningful GPU allocation begins.

Failed to load model
error loading model: error loading model architecture: unknown model architecture: 'qwen35'

What are qwen35 and qwen35moe?

The architecture key inside GGUF does not need to match the marketing name character-for-character. Dense Qwen3.5 files may use qwen35, while Mixture-of-Experts variants can use qwen35moe. The runtime uses this key to choose the correct loader and compute path.

A Qwen3.5-27B file may therefore expose qwen35 while a Qwen3.5-35B-A3B MoE file may expose qwen35moe. In both cases an 'unknown architecture' failure points first to runtime compatibility, not model size.

Current llama.cpp source contains Qwen3.5 architecture constants, so an older packaged runtime can fail even while current upstream source already knows the model.

Why this is not primarily a RAM or VRAM error

Memory failures happen after the runtime recognizes the model and starts allocating weights, KV cache or compute buffers. They usually contain messages such as out of memory, failed to allocate or CUDA allocation errors. qwen35 unknown architecture stops earlier in the loader path.

A machine with plenty of RAM and VRAM can still fail if the runtime is too old. After updating, you may then encounter a genuine OOM; that is a second-stage problem and a sign the loader progressed further.

LM Studio app version and llama.cpp runtime are separate layers

The LM Studio desktop application manages the UI, model catalog, chat and server features. GGUF inference is performed by a llama.cpp Runtime Extension Pack. Updating only one of these layers may leave the other outdated.

Official LM Studio documentation says Ctrl + Shift + R opens runtime management on Windows/Linux, and the CLI exposes lms runtime commands. As of August 15, 2026, LM Studio 0.4.21 is the stable release; its changelog references llama.cpp 2.28.1+ engine settings.

If your app only lists 1.x engines, update the desktop application first, refresh runtimes, then verify the selected GGUF engine.

Step-by-step Windows fix before re-downloading the model

Deleting a 10-30+ GB model should not be the first response to an architecture error. Re-downloading the same GGUF and loading it through the same old runtime usually reproduces the same failure.

Apply changes one at a time and reload the model after each major step. If the error text changes, diagnose the new message instead of assuming the original problem remains.

Why v1.104.2 can say 'Latest Version Installed' and still fail

'Latest' can mean the newest package exposed by your current application/channel rather than the newest llama.cpp engine available in the modern runtime line. An older app can therefore present v1.104.2 as latest inside its own catalog.

The real screenshot in this guide shows Vulkan, CUDA, CPU and CUDA 12 all at v1.104.2. Switching between those four old builds does not add qwen35 support; update LM Studio, refresh the catalog and select a current 2.x engine.

After installing a new engine, also verify that the GGUF dropdown points to it. Installed and selected are not the same state.

CUDA 12, CUDA, Vulkan or CPU: which runtime should you use?

For NVIDIA GPUs, a current CUDA-based llama.cpp engine is normally the first choice. CUDA 12 is appropriate when your driver and card support that runtime. The key requirement for this error is a recent engine with Qwen3.5 architecture support.

Vulkan is useful across AMD/Intel and for broad compatibility. CPU runtime is also valuable as a diagnostic path: if a small Qwen3.5 loads on CPU but a GPU backend fails differently, you have narrowed the problem to the backend/driver/runtime combination.

Choose stability first, performance second.

Is 'unknown model architecture: qwen35moe' the same family of error?

Yes. qwen35moe is associated with Qwen3.5 MoE variants such as 35B-A3B. LM Studio's official bug tracker contains reports where the MoE architecture key was not recognized.

The sparse expert design changes performance and memory behavior later, but an unknown architecture failure happens before those optimizations matter. Update and select a compatible runtime first.

Failed to load model
error loading model: error loading model architecture: unknown model architecture: 'qwen35moe'

Updated everything but the error remains: second-level checks

Verify the active engine, not just the list of installed engines. Multiple runtime versions can coexist. Unload the model, restart LM Studio and reload it under the new engine.

Then test a small Qwen3.5 from a trusted catalog. If a 2B/4B Qwen3.5 loads while one third-party GGUF fails, investigate that file/quantization build. If even the small current model reports qwen35 unknown architecture, the wrong engine is still being used.

Use server logs to confirm the architecture value and selected runtime version.

Do GGUF quantizations such as Q4_K_M, Q5, Q6 or Q8 affect this error?

Quantization changes file size, memory demand and quality. It does not add an architecture implementation to an old runtime. If qwen35 is unknown, switching from Q8 to Q4 does not fix the loader's missing model type.

After architecture support works, quantization becomes relevant to RAM/VRAM fit. Remember that model weights are only part of memory use; KV cache, context, multimodal projectors and runtime buffers add overhead.

Use a smaller Qwen3.5 as a compatibility test before deciding whether your large model needs a lighter quantization.

Architecture fixed, now Out of Memory: context length matters

The official Qwen3.5-27B model card states a default context length of 262,144 tokens and recommends reducing the context window if you encounter OOM. Long context can substantially increase KV-cache memory even when the model weights themselves fit.

For local testing, start with a practical context such as 8K or 16K, verify stability, then raise it only when your workload needs more. GPU offload, quantization and context must be balanced against real VRAM/RAM limits.

An OOM after the qwen35 error disappears is a different problem and often means the runtime now recognizes the model.

Model loads, but image input fails: mmproj and vision troubleshooting

This is a different stage. Qwen3.5 can provide multimodal capabilities, but GGUF deployments may require a compatible projector/mmproj file. llama.cpp issue reports show the separate 'image input is not supported' / mmproj error family.

If text chat works, architecture loading succeeded. Check whether the model package includes a projector and whether LM Studio associated it correctly. Use the model card instructions for that exact GGUF build.

Text-only usage may remain available even when the vision projector is missing.

LM Studio 'Failed to load model': read the technical line, not just the title

'Failed to load model' is an umbrella message. The next line distinguishes runtime compatibility, memory, file integrity, GPU initialization and multimodal-projector failures.

Search the exact technical line and inspect Developer Logs or lms log stream --source server if the popup is too short.

Verify runtimes and logs from the command line

The official LM Studio CLI can list and update runtimes without opening the UI. Use lms runtime ls, lms runtime update and lms runtime select to verify what is installed and active.

Use lms log stream --source server for loading logs, lms ls for downloaded models and lms ps for models currently in memory. These outputs are useful when reporting a reproducible bug.

Redact private paths or credentials before posting logs publicly.

lms runtime ls
lms runtime update
lms runtime select

lms ls
lms ps
lms log stream --source server

Which Qwen3.5 size should you choose?

LM Studio lists Qwen3.5 variants including 2B, 4B, 9B, 27B and 35B-A3B. Smaller dense models are convenient for compatibility checks, while larger models demand more memory and may offer stronger quality depending on task and quantization.

After a runtime update, load a small Qwen3.5 first. If it works, then move to 27B or 35B-A3B and solve any capacity errors separately.

For coding, multilingual chat, tool use and document workloads, model size is only one variable; context, quantization and prompt formatting also matter.

Performance tuning after the model finally loads

Benchmark with the same model, quantization, context and prompt. Increase GPU offload only while leaving enough VRAM headroom for KV cache and runtime buffers.

Avoid unnecessarily huge context windows. Record app version, runtime version, GPU, model file and quantization whenever you compare tokens/s because Qwen3.5 performance can change across llama.cpp engine releases.

Measure time-to-first-token and prompt processing as well as decode tokens/s.

When should you actually re-download the GGUF?

A pure qwen35/qwen35moe architecture message is usually not enough reason to re-download. The loader has already read metadata from the file.

Re-download when file size is wrong, the transfer was interrupted, you see unexpected EOF, missing tensor/header errors or a checksum mismatch. Multi-part GGUFs also require every shard to be present.

If the same file opens in another current llama.cpp-based application, that is useful evidence that the file is readable.

Clean reinstall as a last resort

Before a full reinstall, note your model directory and keep the model files. Removing an old runtime and installing a current one is less disruptive than deleting the whole application state.

If you must reinstall, protect chat history, settings and model paths as needed. After reinstalling, test a small Qwen3.5 first and then the large model.

For a bug report, include LM Studio version, OS, GPU, runtime name/version, exact GGUF filename and server logs.