A public, anonymous Cloudflare R2 mirror of MLX-quantised LLM weights — the same files rapid-mlx pull downloads under the hood. Browse the catalog, copy the install command, and have a model serving on your Mac in minutes.
$ rapid-mlx pull qwen3.5-4b-4bit
Six families rapid-mlx supports as first-class citizens — each R2-mirrored alias here is verified against its shard sum at render time and tagged to the same RAM tier the hardware-tiers installer recommender uses. Hunyuan 3 (HY3) is the one exception: it is Ultra-only (M3 Ultra 256 GB), is not on the R2 mirror, and is pulled straight from HuggingFace by alias (rapid-mlx serve hy3-preview-4bit). Skip to the full catalog below for the long tail.
The 35B-A3B MoE anchors the 48-95 GB tier (3B active, 35B total). Native MTP head baked into every checkpoint — pair with a matching MTP drafter (below) for speculative decode.
qwen3.6-35b-8bit35B-A3B MoE at 8-bit MLX. Anchor pick of the 48-95 GB tier.
$ rapid-mlx pull qwen3.6-35b-8bitqwen3.6-35b-4bit35B-A3B MoE at 4-bit MLX. Fits at 32-48 GB with headroom.
$ rapid-mlx pull qwen3.6-35b-4bitqwen3.6-27b-4bitDense 27B at 4-bit MLX. Hybrid attention layout.
$ rapid-mlx pull qwen3.6-27b-4bitqwen3.5-4b-4bitQwen 3.5 4B 4-bit — the small-tier stand-in until Qwen 3.6 ships a <8B SKU.
$ rapid-mlx pull qwen3.5-4b-4bitGoogle's dense Gemma 4 line. QAT variants ship for sharper low-bit quants; the 12B / 26B / 31B checkpoints land cleanly on Mac Studio.
gemma-4-31b-4bitDense 31B at 4-bit MLX. Strong chat/tools baseline.
$ rapid-mlx pull gemma-4-31b-4bitgemma-4-26b-4bitDense 26B at 4-bit MLX. Trimmed from 31B.
$ rapid-mlx pull gemma-4-26b-4bitgemma-4-12b-8bitDense 12B at 8-bit MLX. Higher fidelity than the 4-bit twin.
$ rapid-mlx pull gemma-4-12b-8bitgemma-4-12b-4bitDense 12B at 4-bit MLX. Fits a 16 GB MacBook.
$ rapid-mlx pull gemma-4-12b-4bitR1-Distill line — chain-of-thought reasoning baked into the base weights via distillation from the DeepSeek R1 teacher.
deepseek-r1-32b-4bitR1-Distill Qwen 32B at 4-bit MLX. Reasoning-first flagship distill.
$ rapid-mlx pull deepseek-r1-32b-4bitdeepseek-r1-8b-4bitR1-Distill Qwen3 8B at 4-bit MLX. Reasoning on a 16 GB Mac.
$ rapid-mlx pull deepseek-r1-8b-4bitOpenAI's harmony-native MoE. Ships MXFP4 weights with 8-bit accumulators; parser pair auto-selects harmony/harmony.
gpt-oss-120b-mxfp4-q8OpenAI 120B MoE at MXFP4-Q8. Frontier reasoning + tool calling on-device.
$ rapid-mlx pull gpt-oss-120b-mxfp4-q8gpt-oss-20b-mxfp4-q8OpenAI 20B MoE at MXFP4-Q8. Anchor pick of the 24-47 GB tier.
$ rapid-mlx pull gpt-oss-20b-mxfp4-q8Tencent's Hunyuan 3 — a 295B-total / 21B-active MoE with a 3.8B MTP head. Ultra-only: peak resident is ~156 GB, so it needs an M3 Ultra with 256 GB unified memory. Will not fit smaller Macs. Weights pull from HuggingFace (not R2-mirrored).
hy3-preview-4bit295B-A21B MoE at 4-bit MLX. Ultra-only — requires an M3 Ultra 256 GB Mac (~156 GB peak). Pulls from HuggingFace, not the R2 mirror.
$ rapid-mlx serve hy3-preview-4bitA ternary-quantised Qwen 3.5-class 27B that packs into 7.9 GB and runs on a 16 GB Mac — the quant is stock MLX 2-bit affine, no custom kernel. Strong on the mainstream: code, math, reasoning, EN/ZH general writing, and agent/framework tool-calling (verified end-to-end with the OpenAI SDK, LangChain, pydantic-ai, smolagents, and Aider). Known limit: strict Chinese classical regulated verse (五言绝句 with fixed rhyme) can trip a repetition loop — a niche edge case.
bonsai-27b-2bit27B ternary (stock MLX 2-bit affine) in 7.9 GB — fits a 16 GB Mac. ~46 tok/s on M3 Ultra. Strong at code/math/reasoning + EN/ZH writing + tool-calling.
$ rapid-mlx pull bonsai-27b-2bitSmall drafter heads that pair with their parent Qwen 3.6 checkpoint. Pass --enable-mtp on rapid-mlx serve to activate speculative decode.
qwen3.6-35b-mtp-4bitSpeculative-decode sidecar for qwen3.6-35b-8bit. Not standalone weights.
$ rapid-mlx pull qwen3.6-35b-mtp-4bitqwen3.6-27b-mtp-4bitSpeculative-decode sidecar for qwen3.6-27b-4bit. Not standalone weights.
$ rapid-mlx pull qwen3.6-27b-mtp-4bitNot sure which model? Pick your Mac RAM tier; we'll surface the alias that gives the best single-user throughput at that footprint. The full catalog is below.
Every alias rapid-mlx ships with, joined against what is currently mirrored on R2. Files are immutable per HuggingFace revision and cached aggressively at the edge.
Every file is a public, CORS-enabled URL served straight from Cloudflare R2. Grab one file, a whole repo, or point any MLX runtime at the result — egress is free, so pull as much as you like. The files are the exact upstream MLX weights, usable with mlx-lm, LM Studio, mlx-vlm, or anything that reads a local model folder.
curl -LO https://models.rapidmlx.com/mlx-community/Qwen3.5-9B-6bit/model.safetensorsREPO=mlx-community/Qwen3.5-9B-6bit
curl -s "https://models.rapidmlx.com/$REPO/" | grep -o '"name":"[^"]*"' | cut -d'"' -f4 | xargs -P4 -I{} curl -fL --create-dirs -o "$REPO/{}" "https://models.rapidmlx.com/$REPO/{}"In a browser, open any /<owner>/<repo>/ URL (the “browse files” link on each card) for a clickable file index.
# mlx-lm
mlx_lm.generate --model ./$REPO --prompt "Hello"
# ...or open the downloaded folder in LM Studio, or load it with mlx-vlm.