Skip to content

docs: research local AI options for M3 Pro 36 GB #323

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
bearmug wants to merge 3 commits into
base: main
Choose a base branch
from
Open

docs: research local AI options for M3 Pro 36 GB #323

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
f2ee93b
docs: research local AI options for M3 Pro
bearmug Jul 14, 2026
2980c88
docs: clarify M3 Max portability
bearmug Jul 14, 2026
3bfd7db
docs: record research validation status
bearmug Jul 14, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 9 additions & 0 deletions docs/benchmark-evidence/local-ai-candidate-inventory-2026-07-14.csv
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
candidate,artifact_or_endpoint,revision,artifact_or_tree_bytes,artifact_or_tree_gb,role,disposition,confidence,evidence_status,retrieved_at,runtime_target,license,source_url,notes
Current Qwen3.6 GGUF Q4_K_M,http://127.0.0.1:8004/v1,5bc3e238d916f48a861bac2f8a1990a0e9b7e98d,22663387424,22.663,control,preserve-and-repair-reproducibility,high,local-observation-plus-canonical-size-match,2026-07-14,llama.cpp-b9740,apache-2.0,https://huggingface.co/unsloth/Qwen3.6-35B-A3B-MTP-GGUF,"Healthy live server; 128K Q8 KV; MTP depth 2; unlinked local file exactly matches canonical Q4_K_M byte size but content hash was not recoverable"
Active Qwable 3.6 35B Q4_K_M,http://127.0.0.1:8005/v1,c6faceabf608fb1e58679ad80d8f4d72d6be9fd0,21166757920,21.167,non-candidate-active-service,deactivate-with-approval-before-benchmarking,high,locally-observed-and-backed-by-pinned-cache,2026-07-14,llama.cpp-b9740,mit,https://huggingface.co/Mia-AiLab/Qwable-3.6-35b,"CPU-only 8K service; vmmap reports 19.7 GB writable memory swapped out; do not delete its cache until ownership and restart policy are understood"
North Mini Code 1.0 4-bit,mlx-community/North-Mini-Code-1.0-4bit,dfbe084dfa26e241345af99ca32848f38fd865f9,18514868006,18.515,coding-worker,benchmark-first,medium,vendor-documented-plus-upstream-runtime-listed,2026-07-14,mlx-vlm-0.6.4,apache-2.0,https://huggingface.co/mlx-community/North-Mini-Code-1.0-4bit,"First new candidate; vendor coding lead not independently reproduced; test MLX-VLM and Pi tools"
North Mini Code 1.0 6-bit,mlx-community/North-Mini-Code-1.0-6bit,8e32b0f6e2e6aa33335564b4518f03563d825d9d,25901561738,25.902,coding-worker,defer,medium,vendor-documented-plus-upstream-runtime-listed,2026-07-14,mlx-vlm-0.6.4,apache-2.0,https://huggingface.co/mlx-community/North-Mini-Code-1.0-6bit,"Test only if 4-bit wins and a measured quality defect justifies the tighter memory and disk fit"
Qwen3.6 35B-A3B OptiQ 4-bit,mlx-community/Qwen3.6-35B-A3B-OptiQ-4bit,cf989459c872374b9622f97f7e3b6493d6a2032d,24693956069,24.694,general-agent,benchmark-mtp-off-on,medium,publisher-documented-unverified-locally,2026-07-14,mlx-optiq-0.3.3,apache-2.0,https://huggingface.co/mlx-community/Qwen3.6-35B-A3B-OptiQ-4bit,"Complete tree includes core quant plus 1.645 GB MTP head and 0.893 GB vision package; 1.4x is a publisher claim"
Gemma 4 26B-A4B uniform 4-bit,mlx-community/gemma-4-26b-a4b-it-4bit,0d77464eeb233a2da68ebf9d7dc4edaac7db956d,15373588575,15.374,reasoning-vision,benchmark-before-larger-quant,medium,vendor-documented-plus-upstream-runtime-listed,2026-07-14,mlx-vlm-0.6.4,apache-2.0,https://huggingface.co/mlx-community/gemma-4-26b-a4b-it-4bit,"Lowest-cost Gemma candidate; use MLX-VLM for vision and verify parsed OpenAI tool calls"
Gemma 4 26B-A4B QAT OptiQ 4-bit,mlx-community/gemma-4-26B-A4B-it-qat-OptiQ-4bit,7f948260d6bea23a9b7a926c39d6437491531e7e,21887491387,21.887,reasoning-vision,defer-until-uniform-result,medium,publisher-documented-unverified-locally,2026-07-14,mlx-optiq-0.3.3,apache-2.0,https://huggingface.co/mlx-community/gemma-4-26B-A4B-it-qat-OptiQ-4bit,"Extra 6.5 GB versus uniform 4-bit needs a measured quality benefit"
DiffusionGemma 26B-A4B 4-bit,mlx-community/diffusiongemma-26B-A4B-it-4bit,252183330817f96e9cba0b20cc400b2947a575cf,16575473079,16.575,experimental,spike-only,medium,vendor-documented-plus-upstream-runtime-listed,2026-07-14,mlx-vlm-0.6.4,apache-2.0,https://huggingface.co/mlx-community/diffusiongemma-26B-A4B-it-4bit,"MLX-VLM 0.6.4 has specialized runtime and parser fix; Google reports lower capability than autoregressive Gemma 4"
Loading
Loading