# Local AI Release Board
> Tracks new open-weight model releases (Hugging Face) and local-AI runtime/harness releases (GitHub), with a "runs on" fit check for NVIDIA/AMD/Intel Arc GPUs, DGX Spark, Strix Halo, Macs and CPU-only machines. Window: last 30 days; older items move to the archive. Generated 2026-10-24T23:23:29Z.

- [Board](https://recentmodel.com/): Just landed + main board, filterable (devices, custom memory, GPU count, lowest quant, task, architecture, params, license, source); filters persist in the URL query string
- [Archive](https://recentmodel.com/archive/): searchable archive of models that left the board (never deleted), same filters
- [JSON feed](https://recentmodel.com/api/releases.json): models[], filtered_models[], harnesses[], ollama_newest[], x_posts[], devices[], estimate_note
- [Archive JSON](https://recentmodel.com/api/archive.json): models[] with archive_reason, archived_at, low_signal, last-known stats and fit
## Notes
- Facts come from the HF API/model cards, GitHub releases, ollama.com/library and public X posts. Unknown fields are left blank.
- Main board threshold (non-trusted orgs must clear one): >= 50 HF likes, >= 5000 HF downloads, or >= 100 likes on X posts naming the model. Trusted orgs bypass it. Anything created or first seen in the last 72h skips the threshold and is shown as 'Just landed' (models[].lane=='fresh', vetted=false means 'new, unvetted'), sorted by velocity; unknown uploaders still need >= 2 likes or an X post linking the repo, and tests/checkpoints/merges/abliterated/re-uploads/empty cards/upload spam are dropped. After that window, items below threshold go to the archive marked low_signal (also listed in filtered_models[] for that run).
- Items older than 30 days (release or first-seen) leave the board for the archive.
- Advisory · OpenCode: Security incident Oct 17, 2026: rotate provider keys saved in OpenCode Console (https://ctxsec.com/items/opencode-console-provider-key-breach-2026-10)
- Summaries are assembled from config.json and the model card; quoted claims are vendor-reported, not verified.
- "Runs on": Need = weight size + FP16 KV cache for 8192 tokens (from config.json when available, otherwise a rough 8% of Q8 weights) + 1.0GB overhead. Weight size comes from the real quant files on Hugging Face when they exist (Q4_K_M, Q8_0 and BF16/F16 GGUF from the official org, unsloth, bartowski, ggml-org, lmstudio-community or mradermacher; MLX 4-bit/8-bit/bf16; or the official BF16 safetensors; another community GGUF only as a last resort, marked community), checked against the parameter count. When no matching file exists it is ESTIMATED at ~4.8/8.5/16 bits per weight and labelled 'est'. Q4 may use Unsloth UD-Q4_K_XL or Q4_K_S when no Q4_K_M exists. Usable memory per device is assumed below physical (GPUs ≈ VRAM − 1GB; Mac ≈75% of unified RAM; Strix Halo ≈96GB; CPU = RAM minus OS, and slow). est.fit in the JSON is for one card per device; the page can multiply discrete GPUs by 2 or 4, which assumes the runtime splits the model across cards (tensor/pipeline parallel or llama.cpp layer split). Arc classes only count Q8/FP16, matching what performs well on Arc. MoE models still need all weights resident; active params only affect speed. Fit is still a sizing calculation, not verified by running.

## Optional
- [Full board as markdown](https://recentmodel.com/llms-full.txt)
- [Board markdown](https://recentmodel.com/index.md) (also served for Accept: text/markdown on /)
- [API catalog](https://recentmodel.com/.well-known/api-catalog)
- [Agent skill](https://recentmodel.com/.well-known/agent-skills/recentmodel/SKILL.md)