Qwen

Qwen3-235B-A22B-Instruct-2507

235.1B parameters · mixture-of-experts · Instruct · Qwen3 family

Mixture-of-experts — what it means for memory & speed → Long context — how it uses memory →
apache-2.0 hash verified source matched revision pinned
Good for: Long context

Which version should I download?

Default pick: GGUF · UD-Q2_K_XL — Unsloth dynamic 2-bit (UD-Q2_K_XL) — accessible 235B-A22B MoE giant lead (~89 GB, 2 shards; preservation copy). Apache-2.0.

Set your rig to confirm it fits and to see if a higher-quality quant runs fully on your hardware.

Set your rig

Quantizations

Why these quants? Which quant should I pick? What do these mean?

GGUF · UD-Q2_K_XL

82.67 GB

Set your rig
Est. speed
—
Swarm
1S / 0L webseed OK

Unsloth dynamic 2-bit (UD-Q2_K_XL) — accessible 235B-A22B MoE giant lead (~89 GB, 2 shards; preservation copy). Apache-2.0.

Magnet .torrent

Want a different quant? Request it on the board →

Runs fully on

RTX Pro 6000 Blackwell 96GB · UD-Q2_K_XL Apple M4 Max (128 GB) · UD-Q2_K_XL Apple M3 Ultra (512 GB) · UD-Q2_K_XL Apple M3 Ultra (256 GB) · UD-Q2_K_XL Apple M3 Max (128 GB) · UD-Q2_K_XL

Green = the model's best quant fits fully in GPU/unified memory at 8K context. Tap a card for its full "what runs on it" page.

Context

Advertised 256k · usable ≈ 512k–1M (measured — RULER)

How we know

Evidence grade: measured — a published evaluation we can cite, with task and length stated.

Model card RULER table (DCA, 260 samples/length): 91.0 at 256k, 90.9 at 512k, 84.5 at 1000k. Usable band exceeds the advertised native window; card: extendable to 1,010,000.

Reviewed on Jul 18, 2026.

Capabilities (as declared by the maintainer): tool calling

Run it

Runtime completeness (UD-Q2_K_XL torrent): llama.cpp ✅ Ollama – vision sidecar –

  • – Ollama — runs the GGUF directly; no Modelfile bundled
  • – vision sidecar — not a vision model

Context / KV 8,192 tokens · FP16 set above the quant table

Generic commands (no rig set). GPU-offload values assume the model fits on your GPU — set your rig for values tuned to your hardware.

Start the server 
llama-server -m Qwen3-235B-A22B-Instruct-2507-UD-Q2_K_XL-00001-of-00002.gguf -c 8192 -ngl 999

Use llama-cli in place of llama-server for a one-shot prompt.

Save as Modelfile next to the GGUF (Modelfile) 
FROM ./Qwen3-235B-A22B-Instruct-2507-UD-Q2_K_XL-00001-of-00002.gguf
PARAMETER num_ctx 8192
PARAMETER num_gpu 999
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|endoftext|>"
PARAMETER temperature 0.7
PARAMETER top_p 0.8
PARAMETER top_k 20
# Chat template: Ollama uses the template embedded in the GGUF (no TEMPLATE directive needed).
Create 
ollama create qwen-qwen3-235b-a22b-instruct-2507 -f Modelfile
Run 
ollama run qwen-qwen3-235b-a22b-instruct-2507

GGUF is not a first-class vLLM format.

vLLM GGUF support is experimental and single-file only; prefer safetensors/GPTQ/AWQ for production. If you must, pass the .gguf path to `vllm serve` with --load-format gguf on a recent vLLM.

transformers does not load GGUF weights.

GGUF weights run under llama.cpp, Ollama or LM Studio.

Load Qwen3-235B-A22B-Instruct-2507-UD-Q2_K_XL-00001-of-00002.gguf, set the context length to 8192 tokens. Set GPU offload to Max (all layers).

MLX runs MLX-format weights only (Apple Silicon). This quant is a GGUF build.

GGUF weights run under llama.cpp, Ollama or LM Studio.

Technical details

Chat template

{%- if tools %}
    {{- '<|im_start|>system\n' }}
    {%- if messages[0].role == 'system' %}
        {{- messages[0].content + '\n\n' }}
    {%- endif %}
    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
    {%- for tool in tools %}
        {{- "\n" }}
        {{- tool | tojson }}
    {%- endfor %}
    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
{%- else %}
    {%- if messages[0].role == 'system' %}
        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
    {%- endif %}
{%- endif %}
{%- for message in messages %}
    {%- if message.content is string %}
        {%- set content = message.content %}
    {%- else %}
        {%- set content = '' %}
    {%- endif %}
    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
    {%- elif message.role == "assistant" %}
        {{- '<|im_start|>' + message.role + '\n' + content }}
        {%- if message.tool_calls %}
            {%- for tool_call in message.tool_calls %}
                {%- if (loop.first and content) or (not loop.first) %}
                    {{- '\n' }}
                {%- endif %}
                {%- if tool_call.function %}
                    {%- set tool_call = tool_call.function %}
                {%- endif %}
                {{- '<tool_call>\n{"name": "' }}
                {{- tool_call.name }}
                {{- '", "arguments": ' }}
                {%- if tool_call.arguments is string %}
                    {{- tool_call.arguments }}
                {%- else %}
                    {{- tool_call.arguments | tojson }}
                {%- endif %}
                {{- '}\n</tool_call>' }}
            {%- endfor %}
        {%- endif %}
        {{- '<|im_end|>\n' }}
    {%- elif message.role == "tool" %}
        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
            {{- '<|im_start|>user' }}
        {%- endif %}
        {{- '\n<tool_response>\n' }}
        {{- content }}
        {{- '\n</tool_response>' }}
        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
            {{- '<|im_end|>\n' }}
        {%- endif %}
    {%- endif %}
{%- endfor %}
{%- if add_generation_prompt %}
    {{- '<|im_start|>assistant\n' }}
{%- endif %}

Sampling defaults

top_k
20
top_p
0.8
temperature
0.7

Stop strings

<|im_end|> <|endoftext|>

Evidence & provenance

Source

Provenance
https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507
Revision pin
ac9c66cc9b46af7306746a9250f23d47083d689e
Manifest
Present

License

Name
apache-2.0
Commercial use
yes
Access
Open

How verification works →

File hashes (SHA-256)

  • qwen3-235b-a22b-instruct-2507-ud-q2-k-xl/LICENSE cfc7749b96f63bd31c3c42b5c471bf756814053e847c10f3eb003417bc523d30
  • qwen3-235b-a22b-instruct-2507-ud-q2-k-xl/Qwen3-235B-A22B-Instruct-2507-UD-Q2_K_XL-00001-of-00002.gguf f1c17a179cdfe393d381ece4f6e2c6393342eacb3cd82990f26908edeceb9ba7
  • qwen3-235b-a22b-instruct-2507-ud-q2-k-xl/Qwen3-235B-A22B-Instruct-2507-UD-Q2_K_XL-00002-of-00002.gguf 7c955f3676def5d48bb016fdc670a2ff01c0696f0ac5653059aac335ef5899f6
  • qwen3-235b-a22b-instruct-2507-ud-q2-k-xl/README.md 8cd13370d19c6398c9b142c3cd0903d2222954e9c42b3f1b4b30eadaf3fb6028

Explore further

Appears in

What runs on a 128 GB Mac What runs on a 192 GB Mac

Also consider

Get an email when a better quant fits your rig.

Set your rig once; we'll alert you when a new or better-fitting build lands.

Set your rig

Performance reports

Real-world throughput reported by the community (and scraped sources).

View performance leaderboard
Community

Reviews, sampler presets and community runtime reports. Be the first to contribute.

Reviews (0)

Log in to write a structured review of this model.

No reviews yet.

Sampler presets

Log in to share a sampler preset or vote on presets.

No presets yet.

Community runtime reports

Unofficial, community-submitted "it loads for me" reports. These are not the official verified-working badge — an admin reviews reports before anything is marked officially verified.

Log in to report whether a quant loads in your runtime.

No community reports yet.

selected to compare · pick at least 2

Compare ()