Pick your unified memory, compare verified model families, and copy one command. Public weights, no account, built for Apple Silicon.
rapid-mlx pull qwen3.5-4b-4bitThe models people run most, from tiny and fast to frontier-scale. Each speed is the fastest measured run we know of, with the chip that ran it.
Liquid AI's newest. Runs on any Mac.
Ternary weights: tiny and blazing fast.
Small and quick, with tool calling.
The top pick for coding agents and chat.
The speed pick for everyday chat and code.
Google's open MoE, with vision.
NVIDIA's new MoE for reasoning and tools.
Frontier-class reasoning.
Frontier DeepSeek on a single Mac.
Pick your Mac's unified-memory tier and see the two models rapid-mlx recipe surfaces for it — smart first, fast second. Same picks as the CLI recommender. The full catalog stays available below.
Not for coding
rapid-mlx pull lfm2.5-2.6b-4bitBasic chat
rapid-mlx pull lfm2.5-1b-4bit| RAM | Smart pick | Fast pick |
|---|---|---|
| 8 GB | LFM2.5 2.6B 4bit | LFM2.5 1B 4bit |
| 16 GB | Qwen3.5 4B 4bit | LFM2.5 1B 4bit |
| 18 GB | Qwen3.5 9B 4bit | Qwen3.5 4B 4bit |
| 24 GB | Ternary Bonsai 27B 2bit | Qwen3.5 4B 4bit |
| 32 GB | Qwen3.8 27B 4bit | Qwen3.5 4B 4bit |
| 48 GB | Qwen3.8 27B 4bit | Qwen3.6 35B-A3B 4bit |
| 64 GB | Qwen3.8 27B 4bit | Qwen3.6 35B-A3B 4bit |
| 96 GB | Qwen3.8 27B 4bit | Qwen3.6 35B-A3B 4bit |
Picks come straight from the engine recommender (rapid-mlx recipe / hardware tiers): smart = highest capability that fits, fast = highest throughput. Footprint is measured resident memory; context length, KV cache and concurrent requests consume more.
Verified aliases, shard sizes, RAM tiers and copy-ready commands. Hunyuan 3 is the one family fetched from Hugging Face rather than the R2 mirror.
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.
35B-A3B MoE at 8-bit MLX. Anchor pick of the 48-95 GB tier.
rapid-mlx pull qwen3.6-35b-8bit35B-A3B MoE at 4-bit MLX. Fits at 32-48 GB with headroom.
rapid-mlx pull qwen3.6-35b-4bitDense 27B at 4-bit MLX. Hybrid attention layout.
rapid-mlx pull qwen3.6-27b-4bitQwen 3.5 4B 4-bit — the small-tier pick.
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.
Dense 31B at 4-bit MLX. Strong chat/tools baseline.
rapid-mlx pull gemma-4-31b-4bitDense 26B at 4-bit MLX. Trimmed from 31B.
rapid-mlx pull gemma-4-26b-4bitDense 12B at 8-bit MLX. Higher fidelity than the 4-bit twin.
rapid-mlx pull gemma-4-12b-8bitDense 12B at 4-bit MLX. Fits a 16 GB MacBook.
rapid-mlx pull gemma-4-12b-4bitDeepSeek's on-device lineup: the V4 Flash frontier MoE (1M context, Ultra-class) leads, backed by the accessible R1-Distill reasoning models that bake chain-of-thought into smaller weights.
V4 Flash frontier MoE at 2-bit (DQ). Flagship with 1M context. Ultra-class — needs ~96 GB+ unified memory.
rapid-mlx pull deepseek-v4-flash-2bitR1-Distill Qwen 32B at 4-bit MLX. Reasoning-first flagship distill.
rapid-mlx pull deepseek-r1-32b-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.
OpenAI 120B MoE at MXFP4-Q8. Frontier reasoning + tool calling on-device.
rapid-mlx pull gpt-oss-120b-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).
295B-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 verse (five-character jueju quatrains with fixed rhyme) can trip a repetition loop — a niche edge case.
27B ternary (stock MLX 2-bit affine) in 7.9 GB — fits a 16 GB Mac. 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. They are not standalone checkpoints — pull one alongside its parent, then enable speculative decode on the parent.
rapid-mlx serve qwen3.6-35b-4bit and rapid-mlx serve qwen3.6-27b-4bit turn MTP on by default; on the other quants of the parent, pass --speculative-config '{"method":"mtp"}'. rapid-mlx info <alias> shows which. There is no standalone serve command for a drafter on its own.
Speculative-decode sidecar for qwen3.6-35b-8bit. Not standalone weights.
rapid-mlx pull qwen3.6-35b-mtp-4bitSpeculative-decode sidecar for qwen3.6-27b-4bit. Not standalone weights.
rapid-mlx pull qwen3.6-27b-mtp-4bitEvery rapid-mlx alias joined against the live R2 mirror.
loading catalog…Context is read from the config.json of the exact build each alias pulls, so two quantisations of one model can differ. AA index is the Artificial Analysis Intelligence Index for the base model at full precision with reasoning on — a property of the model, not a score for our quantised build. Sort by it to find the most capable model your Mac can actually hold.
Not in the catalog? Point it at a Hugging Face repo instead: rapid-mlx serve <org/repo> checks format, architecture and memory fit before it downloads, and rapid-mlx import quantizes a bf16 checkpoint for your Mac. Bring your own model →
The most-pulled models, each with its own page. Each table is our own reference run: one machine, one engine version, one method. Fastest known run adds the fastest community benchmark when it beats ours, with the chip that ran it.
| Model | Architecture | Decode | Fastest known run | First token · 2K prompt | Peak memory · Metal | Weights |
|---|---|---|---|---|---|---|
| LFM2.5 1.2B | 1.2 B hybrid conv+attention · 4-bit | 314.0 tok/s | this run | 0.8 s | 1.8 GB | 0.66 GB |
| Qwen3.6 35B (A3B MoE) | MoE · 35 B total / 3 B active · 4-bit | 104.5 tok/s | 151 tok/s · M5 Max | 2.4 s | 23.2 GB | 17.5 GB |
| Qwen3.5 4B | 4 B dense · 4-bit | 85.8 tok/s | 227 tok/s · M5 Ultra | 2.9 s | 4.3 GB | 3.06 GB |
| Gemma 4 26B (A4B MoE) | MoE · 26 B total / 4 B active · 4-bit | 78.5 tok/s | 130 tok/s · M5 Max | 3.1 s | 20.9 GB | 13 GB |
| GPT-OSS 20B | MoE · 20 B total · MXFP4-Q8 · reasoning | 77.2 tok/s | 110 tok/s · M5 Max | 2.8 s | 15.7 GB | 10 GB |
| Qwen3.5 9B | 9 B dense · 4-bit | 63.0 tok/s | 114 tok/s · M3 Ultra | 5.3 s | 8.0 GB | 5.98 GB |
Median of 3 runs, temperature 0, 512-token decode; first token is a cold ~2K-token prompt. Raw data and method.
| Model | Architecture | Decode | Fastest known run | First token · 2K prompt | Peak memory · Metal | Weights |
|---|---|---|---|---|---|---|
| Qwen3 0.6B | 0.6 B dense · 4-bit | 384.2 tok/s | this run | 0.5 s | 12.7 GB | 0.35 GB |
| Ternary Bonsai 1.7B | 1.7 B ternary · 2-bit | 236.4 tok/s | 422 tok/s · M5 Ultra | 1.3 s | 13.1 GB | 0.5 GB |
| Llama 3.2 3B | 3 B dense · 4-bit | 121.2 tok/s | this run | 2.2 s | 14.4 GB | 1.5 GB |
| Qwen3 4B Instruct (2507) | 4 B dense · 4-bit | 93.0 tok/s | this run | 2.9 s | 16.2 GB | 2.3 GB |
| Qwen3 4B Thinking (2507) | 4 B dense · 4-bit · reasoning | 93.0 tok/s | this run | 2.9 s | 16.2 GB | 2.3 GB |
| Gemma 3 4B (QAT) | 4 B dense · QAT 4-bit · vision | 91.3 tok/s | this run | 2.7 s | 7.6 GB | 3.03 GB |
| Qwen3 8B | 8 B dense · 4-bit | 54.0 tok/s | this run | 5.3 s | 18.5 GB | 4 GB |
| Gemma 4 12B | 12 B dense · 4-bit · vision | 33.5 tok/s | 51 tok/s · M5 Max | 8.7 s | 13.8 GB | 6 GB |
| Ternary Bonsai 27B | 27 B ternary · 2-bit | 22.4 tok/s | 57 tok/s · M5 Max | 19.4 s | 13.5 GB | 7.9 GB |
| Qwen3.6 27B | 27 B dense · 4-bit | 23.4 tok/s | 25 tok/s · M5 Max | 18.9 s | 22.8 GB | 13.5 GB |
| Qwen3 Coder 30B (A3B MoE) | MoE · 30 B total / 3 B active · 4-bit · code | 91.5 tok/s | this run | 2.6 s | 27.3 GB | 15 GB |
| LFM2.5 2.6B | 2.6 B hybrid conv+attention · 4-bit | — | 93 tok/s · M2 Pro | Not yet measured on 0.15.5 | 1.6 GB | |
Same Mac, harness and method as the table above, on rapid-mlx 0.15.5. Peak memory spans the whole run (up to a ~16K-token prompt and 4 streams) and includes the engine's prefix cache. The model marked not yet measured did not complete this run; its page leads with the fastest community run instead. Raw data, method and run record.
Every file is a public, CORS-enabled Cloudflare R2 URL. Grab one file, a whole repository, or point any MLX runtime at the result — the files are the exact upstream MLX weights, usable with mlx-lm, LM Studio, mlx-vlm, or your own runtime.
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/$REPO/{}"mlx_lm.generate --model ./$REPO --prompt "Hello"Open any /<owner>/<repo>/ URL on the mirror for a clickable file index.