Qwen
46 MLX aliases · Alibaba's full open-weights line — Qwen3.5 / 3.6 / Coder / VL / legacy / 2.5 / Qwopus.
The Qwen family on rapid-mlx covers every Alibaba open-weights line we support — from the workhorse Qwen3.5 (4B → 122B-A10B MoE) and the post-3.5 refresh Qwen3.6, through the code-tuned Qwen3 Coder and the vision-language Qwen3-VL, down to the smaller legacy Qwen3 / Qwen2.5 baselines and the community Opus-aligned Qwopus. All variants use the hermes tool-call envelope (Qwen3.6 swaps to qwen3_coder_xml) and the qwen3 reasoning parser, so structured tool calls and <think>-style chain-of-thought stream out of /v1/chat/completions and /v1/responses with no extra config.
- family
- Qwen
- aliases
- 46
- lines
- 7
- install
- rapid-mlx serve <alias>
- OpenAI base URL
- http://localhost:8000/v1
Lines in this family
- Qwen3.6 (11 aliases)
- Qwen3.5 (15 aliases)
- Qwen3 Coder (2 aliases)
- Qwen3-VL (4 aliases)
- Qwopus (3 aliases)
- Qwen3 (legacy) (10 aliases)
- Qwen2.5 (1 alias)
Qwen3.6 · 11 aliases
Post-3.5 refresh. Same hybrid-cache MoE architecture as 3.5-35B but a tighter post-training mix and an XML-style tool-call envelope, parsed with our qwen3_coder_xml parser. We ship 8 aliases including DWQ (data-aware quant) and Unsloth's UD-MLX quants alongside the standard 4/6/8-bit set.
parser: qwen3_coder_xml
| alias | hf repo | tool parser | reasoning | flags |
|---|---|---|---|---|
| qwen3.6-27b-4bit | mlx-community/Qwen3.6-27B-4bit | qwen3_coder_xml | qwen3 | — |
| qwen3.6-27b-6bit | lmstudio-community/Qwen3.6-27B-MLX-6bit | qwen3_coder_xml | qwen3 | — |
| qwen3.6-27b-8bit | unsloth/Qwen3.6-27B-MLX-8bit | qwen3_coder_xml | qwen3 | — |
| qwen3.6-27b-ud | unsloth/Qwen3.6-27B-UD-MLX-4bit | qwen3_coder_xml | qwen3 | — |
| qwen3.6-35b-4bit | mlx-community/Qwen3.6-35B-A3B-4bit | qwen3_coder_xml | qwen3 | hybrid · moe |
| qwen3.6-35b-6bit | mlx-community/Qwen3.6-35B-A3B-6bit | qwen3_coder_xml | qwen3 | hybrid · moe |
| qwen3.6-35b-8bit | mlx-community/Qwen3.6-35B-A3B-8bit | qwen3_coder_xml | qwen3 | hybrid · moe |
| qwen3.6-35b-dwq | mlx-community/Qwen3.6-35B-A3B-4bit-DWQ | qwen3_coder_xml | qwen3 | hybrid · moe |
| qwen3.6-35b-mxfp4 | mlx-community/Qwen3.6-35B-A3B-mxfp4 | qwen3_coder_xml | qwen3 | hybrid · moe |
| qwen3.6-35b-nvfp4 | mlx-community/Qwen3.6-35B-A3B-nvfp4 | qwen3_coder_xml | qwen3 | hybrid · moe |
| qwen3.6-35b-ud | unsloth/Qwen3.6-35B-A3B-UD-MLX-4bit | qwen3_coder_xml | qwen3 | hybrid · moe |
Notes & caveats
- Both 27B and 35B variants ship — pick 27B for a 32 GB Mac, 35B-A3B for 64 GB+.
- DWQ + UD-MLX quants are higher-quality 4-bit than the stock release.
- Hybrid cache on all 35B variants — long context with bounded KV growth.
Qwen3.5 · 15 aliases
The workhorse line on rapid-mlx. 10 MLX-native quants from 4 GB-RAM-friendly 4B up to the 122B-A10B MoE that needs a Mac Studio with 192 GB+ unified memory. Default tool-call parser is hermes; reasoning parser is qwen3.
parser: hermes
| alias | hf repo | tool parser | reasoning | flags |
|---|---|---|---|---|
| qwen3.5-122b-6bit | mlx-community/Qwen3.5-122B-A10B-6bit | hermes | qwen3 | hybrid · moe |
| qwen3.5-122b-8bit | mlx-community/Qwen3.5-122B-A10B-8bit | hermes | qwen3 | hybrid · moe |
| qwen3.5-122b-mxfp4 | nightmedia/Qwen3.5-122B-A10B-Text-mxfp4-mlx | hermes | qwen3 | hybrid · moe |
| qwen3.5-27b-4bit | mlx-community/Qwen3.5-27B-4bit | hermes | qwen3 | — |
| qwen3.5-27b-6bit | mlx-community/Qwen3.5-27B-6bit | hermes | qwen3 | — |
| qwen3.5-27b-8bit | mlx-community/Qwen3.5-27B-8bit | hermes | qwen3 | — |
| qwen3.5-35b-4bit | mlx-community/Qwen3.5-35B-A3B-4bit | hermes | qwen3 | hybrid · moe |
| qwen3.5-35b-6bit | mlx-community/Qwen3.5-35B-A3B-6bit | hermes | qwen3 | hybrid · moe |
| qwen3.5-35b-8bit | mlx-community/Qwen3.5-35B-A3B-8bit | hermes | qwen3 | hybrid · moe |
| qwen3.5-4b-4bit | mlx-community/Qwen3.5-4B-MLX-4bit | hermes | qwen3 | — |
| qwen3.5-4b-6bit | mlx-community/Qwen3.5-4B-6bit | hermes | qwen3 | — |
| qwen3.5-4b-8bit | mlx-community/Qwen3.5-4B-8bit | hermes | qwen3 | — |
| qwen3.5-9b-4bit | mlx-community/Qwen3.5-9B-4bit | hermes | qwen3 | — |
| qwen3.5-9b-6bit | mlx-community/Qwen3.5-9B-6bit | hermes | qwen3 | — |
| qwen3.5-9b-8bit | mlx-community/Qwen3.5-9B-8bit | hermes | qwen3 | — |
Notes & caveats
- Default workhorse — start with
qwen3.5-9b-4biton a 16 GB Mac,qwen3.5-27b-4biton 32 GB+. - The 35B and 122B variants are MoE (A3B / A10B) — heavy weights, light active params.
- Hybrid-cache enabled on the 35B and 122B quants for long-context throughput.
Qwen3 Coder · 2 aliases
Alibaba's code-specialised line — instruction-tuned on code corpora and aligned for IDE-style completion + agentic edit workflows. Both aliases use the hermes tool envelope so Cursor / Claude Code / Aider / Continue all work without per-model config.
parser: hermes
| alias | hf repo | tool parser | reasoning | flags |
|---|---|---|---|---|
| qwen3-coder-30b-4bit | mlx-community/Qwen3-Coder-30B-A3B-Instruct-4bit | hermes | — | moe · spec |
| qwen3-coder-4bit | lmstudio-community/Qwen3-Coder-Next-MLX-4bit | hermes | — | hybrid |
Notes & caveats
- 30B-A3B-Instruct is the recommended pick for general code work on a 32 GB+ Mac.
- The MLX-Next 4-bit quant runs hybrid cache for long-file context.
Qwen3-VL · 4 aliases
Vision-language extension of Qwen3. rapid-mlx serves the OpenAI-compatible /v1/chat/completions vision shape — pass image URLs or base64 in the standard message-content array. The 2B / 4B / 8B variants fit a 16 GB Mac; the 30B-A3B MoE wants 32 GB+.
parser: hermes
| alias | hf repo | tool parser | reasoning | flags |
|---|---|---|---|---|
| qwen3-vl-2b-4bit | mlx-community/Qwen3-VL-2B-Instruct-4bit | hermes | — | spec |
| qwen3-vl-30b-4bit | mlx-community/Qwen3-VL-30B-A3B-Instruct-4bit | hermes | qwen3 | moe · spec |
| qwen3-vl-4b-4bit | mlx-community/Qwen3-VL-4B-Instruct-4bit | hermes | qwen3 | spec |
| qwen3-vl-8b-4bit | mlx-community/Qwen3-VL-8B-Instruct-4bit | hermes | qwen3 | spec |
Notes & caveats
- Use the standard OpenAI
image_urlmessage-content shape — no custom transport. - 30B-A3B variant has the
qwen3reasoning parser wired so you also getreasoning_contenton/v1/responses.
Qwopus · 3 aliases
Community-trained "Qwen meets Opus" line — Qwen3 base weights post-trained on an Opus-style alignment mix, with hybrid cache enabled. Three MLX quants: 9B-4bit, 27B-4bit, 27B-8bit.
parser: hermes
| alias | hf repo | tool parser | reasoning | flags |
|---|---|---|---|---|
| qwopus-27b-4bit | Jackrong/MLX-Qwopus3.5-27B-v3-4bit | hermes | qwen3 | hybrid |
| qwopus-27b-8bit | Jackrong/MLX-Qwopus3.5-27B-v3-8bit | hermes | qwen3 | hybrid |
| qwopus-9b-4bit | Jackrong/MLX-Qwopus3.5-9B-v3-4bit | hermes | qwen3 | hybrid |
Notes & caveats
- Hybrid cache on all variants — long-context cost bounded.
- Reasoning parser is
qwen3, so<think>-style CoT streams cleanly.
Qwen3 (legacy) · 10 aliases
The original Qwen3 line — kept for users who want a small, well-understood baseline. Includes the 2507 instruct and thinking checkpoints, which give you the same family on two different alignment recipes.
parser: hermes
| alias | hf repo | tool parser | reasoning | flags |
|---|---|---|---|---|
| qwen3-0.6b | mlx-community/Qwen3-0.6B-4bit | hermes | qwen3 | spec |
| qwen3-0.6b-4bit | mlx-community/Qwen3-0.6B-4bit | hermes | qwen3 | spec |
| qwen3-0.6b-8bit | mlx-community/Qwen3-0.6B-8bit | hermes | qwen3 | spec |
| qwen3-1.7b | mlx-community/Qwen3-1.7B-4bit | hermes | qwen3 | spec |
| qwen3-1.7b-4bit | mlx-community/Qwen3-1.7B-4bit | hermes | qwen3 | spec |
| qwen3-4b-8bit | mlx-community/Qwen3-4B-8bit | hermes | qwen3 | spec |
| qwen3-4b-instruct-2507-4bit | mlx-community/Qwen3-4B-Instruct-2507-4bit | hermes | — | spec |
| qwen3-4b-thinking-2507-4bit | mlx-community/Qwen3-4B-Thinking-2507-4bit | hermes | qwen3 | spec |
| qwen3-8b-4bit | mlx-community/Qwen3-8B-4bit | hermes | qwen3 | spec |
| qwen3-8b-8bit | mlx-community/Qwen3-8B-8bit | hermes | qwen3 | spec |
Notes & caveats
qwen3-0.6b-4bitis the smallest practical chat-quality model in the registry — useful for laptop-class CI smoke tests.- The
-thinking-2507-variant streamsreasoning_content; the-instruct-2507-variant does not.
Qwen2.5 · 1 alias
One Qwen2.5 alias kept in the registry — the 14B Instruct 4-bit — as a reference baseline for anyone benchmarking the Qwen3 line against its predecessor. Same hermes tool envelope as the rest of the Qwen family.
parser: hermes
| alias | hf repo | tool parser | reasoning | flags |
|---|---|---|---|---|
| qwen2.5-14b-4bit | mlx-community/Qwen2.5-14B-Instruct-4bit | hermes | — | spec |
Notes & caveats
- Use Qwen3.5-9B or Qwen3.6-27B instead unless you have a specific reason to pin Qwen2.5.
Notes & caveats
- Tool-call parser is
hermesacross the family except for Qwen3.6, which usesqwen3_coder_xml. - Reasoning parser is
qwen3across the family —<think>-style CoT streams toreasoning_contenton/v1/responses. - For the larger MoE variants (35B-A3B, 122B-A10B) hybrid cache is on by default — long-context cost stays bounded.