Skip to content

Models

The app's built-in catalog. Every entry is a public Hugging Face repo — no token — including the loopl MLX models; only the loopl GGUF builds (gated: yes) are private and need your own token (Settings › Hugging Face). Any other public MLX or GGUF repo loads the same way. Usable on 8 GB is the context that fits beside the weights on an iPhone 16 Pro; it is the app's default.

The loopl models

Qwen3.5 0.8B, 2B and 4B post-tuned on cagataydev/loopl-train (private): the agent loop and how this runtime behaves, tool calls in both dialects loopl speaks, recovering from a tool error instead of repeating it, who they are. The vision tower is the base's, kept — the MLX exports read photos; the GGUF exports are text-only. They are public, in one place: the loopl collection on Hugging Face — the three MLX 4-bit exports (what the app downloads) and the three bf16 checkpoints (the base for continual post-tuning with Models › Train your own). Each card has the scores against the base, the recipe and how to run it in the app, with mlx_vlm or from Swift. The GGUF builds, LoRA adapter and the training dataset stay private.

MLX 4-bit · vision · public bf16 · post-tuning base · public GGUF Q4_K_M · text · private
loopl 0.8B loopl-0.8b-4bit loopl-0.8b loopl-0.8b-GGUF (token)
loopl 2B loopl-2b-4bit loopl-2b loopl-2b-GGUF (token)
loopl 4B loopl-4b-4bit loopl-4b loopl-4b-GGUF (token)

In the app: Models › loopl models → download (no account). On a Mac: pip install mlx-vlm then python -m mlx_vlm.generate --model cagataydev/loopl-2b-4bit --image photo.jpg --prompt "What is this?".

The loopl-vl models

The same training data on Qwen3-VL Instruct bases (2B and 4B): these read pictures and video natively — the app sends sampled frames (see Video input) and transcribes the audio track on the device, because Qwen3-VL has no audio input. Tool calls use the Qwen3-VL (hermes JSON) dialect; no thinking blocks. No GGUF — llama.cpp cannot see. They sit next to the loopl models in the collection, 4-bit + bf16, public.

MLX 4-bit · vision + video · public bf16 · post-tuning base · public
loopl-vl 2B loopl-vl-2b-4bit loopl-vl-2b
loopl-vl 4B loopl-vl-4b-4bit loopl-vl-4b

In the app: Models › loopl models → the loopl-vl rows (the + button then offers Photo or Video). On a Mac: python -m mlx_vlm.generate --model cagataydev/loopl-vl-2b-4bit --video clip.mp4 --fps 2 --video-max-frames 8 --prompt "What happens in this video?".

The catalog

Model Runtime Params 4-bit size ctx max usable on 8 GB min RAM gated tool calls
Qwen3.5 0.8B MLX 0.8B 652 MB 262,144 16,384 2 GB no ✓
loopl 0.8B MLX 0.8B 652 MB 262,144 16,384 2 GB no ✓
Qwen3-VL 2B MLX 2.0B 1.80 GB 262,144 12,288 3 GB no ✓
Qwen3-VL 2B Thinking MLX 2.0B 1.80 GB 262,144 12,288 3 GB no ✓
Qwen3.5 2B MLX 2.0B 1.75 GB 262,144 12,288 3 GB no ✓
loopl 2B MLX 2.0B 1.75 GB 262,144 12,288 3 GB no ✓
loopl-vl 2B MLX 2.0B 1.80 GB 262,144 12,288 3 GB no ✓
DeepSeek-R1-Distill-Qwen-1.5B (4-bit) MLX 1.8B 1000 MB 131,072 8,192 6 GB no —
Llama-3.2-1B Instruct (4-bit) MLX 1.2B 695 MB 131,072 8,192 6 GB no ✓
Llama-3.2-3B Instruct (4-bit) MLX 3.2B 1.81 GB 131,072 8,192 8 GB no ✓
Phi-4-mini-instruct (4-bit) MLX 3.8B 2.16 GB 131,072 8,192 8 GB no ✓
Qwen3-0.6B (4-bit) MLX 0.6B 335 MB 40,960 8,192 6 GB no ✓
Qwen3-1.7B (4-bit) MLX 1.7B 968 MB 40,960 8,192 6 GB no ✓
Qwen3-4B (4-bit) MLX 4.0B 2.26 GB 40,960 8,192 8 GB no ✓
Qwen3-VL 4B MLX 4.0B 3.11 GB 262,144 8,192 5 GB no ✓
Qwen3-VL 4B Thinking MLX 4.0B 3.11 GB 262,144 8,192 5 GB no ✓
SmolLM3-3B (4-bit) MLX 3.1B 1.73 GB 65,536 8,192 8 GB no ✓
gemma-3-1b IT (4-bit) MLX 1.0B 733 MB 32,768 8,192 6 GB no —
gemma-3-1b IT-qat (4-bit) MLX 1.0B 733 MB 32,768 8,192 6 GB no —
loopl 4B MLX 4.0B 3.06 GB 262,144 8,192 6 GB no ✓
loopl-vl 4B MLX 4.0B 3.11 GB 262,144 8,192 6 GB no ✓
gemma-3-4b IT-qat (4-bit) MLX 4.3B 3.00 GB 131,072 6,838 8 GB no —
gemma-3-4b IT (4-bit) MLX 4.3B 3.40 GB 131,072 5,383 12 GB no —
Gemma 3 1B IT · Q4_K_M (llama.cpp) GGUF 1.0B 806 MB 32,768 8,192 6 GB no —
Llama 3.2 3B Instruct · Q4_K_M (llama.cpp) GGUF 3.2B 2.02 GB 131,072 8,192 8 GB no ✓
Qwen3 0.6B · Q4_K_M (llama.cpp) GGUF 0.6B 397 MB 40,960 8,192 3 GB no ✓
Qwen3 1.7B · Q4_K_M (llama.cpp) GGUF 1.7B 1.11 GB 40,960 8,192 6 GB no ✓
Qwen3 4B · Q4_K_M (llama.cpp) GGUF 4.0B 2.50 GB 40,960 8,192 8 GB no ✓
SmolLM2 1.7B Instruct · Q4_K_M (llama.cpp) GGUF 1.7B 1.06 GB 8,192 8,192 6 GB no —
loopl 0.8B · Q4_K_M (llama.cpp) GGUF 0.8B 529 MB 262,144 8,192 3 GB yes ✓
loopl 2B · Q4_K_M (llama.cpp) GGUF 2.0B 1.27 GB 262,144 8,192 4 GB yes ✓
loopl 4B · Q4_K_M (llama.cpp) GGUF 4.0B 2.71 GB 262,144 4,096 6 GB yes ✓
Apple Intelligence (on-device) appleFM 3.0B — 4,096 4,096 — no —

Notes per model

  • Qwen3-0.6B (4-bit) — 40960 = 32k native; thinking mode on by default (/no_think or enable_thinking=false) loopl passes enable_thinking=false for the agent loop. KV 112 KiB/token → memory-bound ctx ≈39,799 on 8 GB (estimate).
  • Apple Intelligence (on-device) — iOS 26+, Apple Intelligence enabled, iPhone 15 Pro or later; zero download; offline; SystemLanguageModel.default.availability
  • Qwen3-1.7B (4-bit) — 40960 = 32k native; thinking mode on by default (/no_think or enable_thinking=false) loopl passes enable_thinking=false for the agent loop. KV 112 KiB/token → memory-bound ctx ≈34,283 on 8 GB (estimate).
  • Qwen3-4B (4-bit) — 40960 = 32k native; thinking mode on by default (/no_think or enable_thinking=false) loopl passes enable_thinking=false for the agent loop. KV 144 KiB/token → memory-bound ctx ≈17,883 on 8 GB (estimate).
  • Qwen3.5 0.8B — Sees pictures. Smallest Qwen3.5 (image-text-to-text), 4-bit MLX. Proven on Mac: image → description + tool call.
  • Qwen3.5 2B — Sees pictures. Qwen3.5 2B, 4-bit MLX — the sweet spot for photos + tools on an 8 GB phone.
  • Qwen3-VL 2B — Sees pictures. Qwen3-VL 2B Instruct, 4-bit MLX.
  • Qwen3-VL 4B — Sees pictures. Qwen3-VL 4B Instruct, 4-bit MLX — best vision quality that fits an 8 GB phone.
  • Qwen3-VL 2B Thinking — Sees pictures and thinks before answering. Qwen3-VL 2B Thinking, 4-bit MLX. The template always opens a block (no enable_thinking switch) — loopl shows the reasoning, then the answer. Slower than Instruct: expect a few hundred reasoning tokens per turn.
  • Qwen3-VL 4B Thinking — Sees pictures and thinks before answering. Qwen3-VL 4B Thinking, 4-bit MLX — best vision reasoning that fits an 8 GB phone. The template always opens a block (no enable_thinking switch) — loopl shows the reasoning, then the answer. Slower than Instruct.
  • Llama-3.2-1B Instruct (4-bit) — KV 32 KiB/token → memory-bound ctx ≈128,317 on 8 GB (estimate).
  • Llama-3.2-3B Instruct (4-bit) — KV 112 KiB/token → memory-bound ctx ≈26,964 on 8 GB (estimate).
  • gemma-3-1b IT (4-bit) — hybrid local/global attention → real KV below the fp16 formula; template has no tools branch → JSON-in-prompt fallback No tool role in the template — loopl folds tool results into user turns. KV 26 KiB/token → memory-bound ctx ≈32,768 on 8 GB (estimate).
  • gemma-3-1b IT-qat (4-bit) — hybrid local/global attention → real KV below the fp16 formula; template has no tools branch → JSON-in-prompt fallback No tool role in the template — loopl folds tool results into user turns. KV 26 KiB/token → memory-bound ctx ≈32,768 on 8 GB (estimate).
  • gemma-3-4b IT (4-bit) — hybrid local/global attention → real KV below the fp16 formula; template has no tools branch → JSON-in-prompt fallback No tool role in the template — loopl folds tool results into user turns. KV 136 KiB/token → memory-bound ctx ≈10,766 on 8 GB (estimate).
  • gemma-3-4b IT-qat (4-bit) — hybrid local/global attention → real KV below the fp16 formula; template has no tools branch → JSON-in-prompt fallback No tool role in the template — loopl folds tool results into user turns. KV 136 KiB/token → memory-bound ctx ≈13,676 on 8 GB (estimate).
  • Phi-4-mini-instruct (4-bit) — KV 128 KiB/token → memory-bound ctx ≈20,919 on 8 GB (estimate).
  • SmolLM3-3B (4-bit) — KV 72 KiB/token → memory-bound ctx ≈42,995 on 8 GB (estimate).
  • DeepSeek-R1-Distill-Qwen-1.5B (4-bit) — reasoning model, output, no tool template KV 28 KiB/token → memory-bound ctx ≈131,072 on 8 GB (estimate).
  • Qwen3 0.6B · Q4_K_M (llama.cpp) — Single-file GGUF; built-in chat template with tools.
  • Qwen3 1.7B · Q4_K_M (llama.cpp) — Single-file GGUF; built-in chat template with tools.
  • Qwen3 4B · Q4_K_M (llama.cpp) — Single-file GGUF.
  • Llama 3.2 3B Instruct · Q4_K_M (llama.cpp) — Public mirror (no licence gate).
  • Gemma 3 1B IT · Q4_K_M (llama.cpp) — No tool template — manual JSON prompting.
  • SmolLM2 1.7B Instruct · Q4_K_M (llama.cpp) — Small and fast; manual JSON prompting.
  • loopl 0.8B — loopl v1.2 — Qwen3.5-0.8B post-tuned by Cagatay Cali on cagataydev/loopl-train (agent loop, tools, robots, identity + who Cagatay Cali is and his projects; trained on loopl-train v2.2.1 rev 47d27b44; right tool 174/250 (v1.1 113), installs a missing tool 25/48 (v1.1 1/48)). Vision kept (mlx_vlm export: vision_tower + preprocessor_config, same layout as the mlx-community base). Public — no token. bf16 base for Train your own: cagataydev/loopl-0.8b.
  • loopl 0.8B · Q4_K_M (llama.cpp) — loopl v1.2 GGUF (text-only, llama.cpp b11401 --no-mtp; trained on loopl-train v2.2.1 rev 47d27b44; right tool 174/250 (v1.1 113), installs a missing tool 25/48 (v1.1 1/48) — identity/bio family, file cards, repair trajectories). Private: needs the user's HF token.
  • loopl 2B — loopl v1.2 — Qwen3.5-2B post-tuned by Cagatay Cali on cagataydev/loopl-train (agent loop, tools, robots, identity + who Cagatay Cali is and his projects; trained on loopl-train v2.2.1 rev 47d27b44; right tool 163/250 (v1.1 135), installs a missing tool 19/48 (v1.1 3/48)). Vision kept (mlx_vlm export: vision_tower + preprocessor_config, same layout as the mlx-community base). Public — no token. bf16 base for Train your own: cagataydev/loopl-2b.
  • loopl 2B · Q4_K_M (llama.cpp) — loopl v1.2 GGUF (text-only, llama.cpp b11401 --no-mtp; trained on loopl-train v2.2.1 rev 47d27b44; right tool 163/250 (v1.1 135), installs a missing tool 19/48 (v1.1 3/48) — identity/bio family, file cards, repair trajectories). Private: needs the user's HF token.
  • loopl-vl 2B — loopl-vl v1 — Qwen3-VL-2B-Instruct post-tuned by Cagatay Cali on the same loopl-train data as loopl v1.1 (rev db6fa452: agent loop, tools, identity + who Cagatay Cali is). Sees pictures AND video natively (frames in; the app transcribes the audio track). Vision tower kept (mlx_vlm export). No thinking blocks. Public — no token. bf16 base for Train your own: cagataydev/loopl-vl-2b.
  • loopl 4B — loopl v1.2 — Qwen3.5-4B LoRA r64 post-tuned by Cagatay Cali on cagataydev/loopl-train. Vision kept (mlx_vlm export: vision_tower + preprocessor_config, same layout as the mlx-community base). Public — no token. bf16 base for Train your own: cagataydev/loopl-4b.
  • loopl 4B · Q4_K_M (llama.cpp) — loopl v1.2 GGUF (text-only, llama.cpp b11401 --no-mtp; trained on loopl-train v2.2.1 rev 47d27b44; right tool 198/250 (v1.1 166), installs a missing tool 27/48 (v1.1 5/48) — identity/bio family, file cards, repair trajectories). Private: needs the user's HF token.
  • loopl-vl 4B — loopl-vl v1 — Qwen3-VL-4B-Instruct post-tuned by Cagatay Cali on the same loopl-train data as loopl v1.1 (rev db6fa452: agent loop, tools, identity + who Cagatay Cali is). Sees pictures AND video natively (frames in; the app transcribes the audio track). Vision tower kept (mlx_vlm export). No thinking blocks. Public — no token. bf16 base for Train your own: cagataydev/loopl-vl-4b.

Picking one

  • Default: qwen3-0.6b-4bit — the best balance of tool calls, speed and context on an 8 GB phone.
  • Up to 4B parameters at 4-bit fit comfortably; 7–8B load but leave little room for context.
  • Tool calls are reliable on Qwen3, Llama 3.2 and Apple's model; Gemma and SmolLM manage one tool, not many.

Generated by docs/gen_models.py from App/Resources/models.json (catalog 2026-10-07) on 2026-10-27. Re-run the script after editing the catalog — do not edit this page by hand.