Ai2 posted OlmoEarth v1.3 on Hugging Face: 128-dimensional satellite embeddings at 10 meters, a 785 MB Hub checkpoint from 4 September, and a browser that maps the whole planet. The llama.cpp org added official GGUFs of Recursive Flint, with Q4_K_M at 23.388 GB. Microsoft and Hugging Face published ThinkingBox-Training, a GRPO recipe that post-trains Qwen3.8-27B from database-state rewards. Daily Papers moved off yesterday's Cache Engineering list to a board led by Deep Search, with Tencent's FAPL paper posted on 22 October.
On the Hub
Ai2 OlmoEarth v1.3
Ai2 published OlmoEarth v1.3 on 22 October. The model reads a year of Sentinel-1 radar plus Sentinel-2 and Landsat imagery and emits one 128-dimensional vector per 10-meter pixel. The first 64 dimensions also work on their own. The Hub card itself dates to 4 September; the news is the blog, whole-planet embeddings, and a public explorer.12
The checkpoint is 785 MB of weights under the OlmoEarth Artifact License, not Apache-2.0. Use, modify, and distribute are allowed, with restrictions on military, defense, and extractive work covering oil, gas, mining, and deforestation, plus a citation of Ai2. Ai2 claims the 128-d embeddings lead on 20 of 27 tasks against AlphaEarth Foundations and Tessera v2, and that the 64-d prefix still leads 18 of 27. Those figures, and a final pretrain run claimed at $2,720 on rented H100s, are Ai2's.
Precomputed annual embeddings for the planet live on Source Cooperative, not as extra Hub weight files. The explorer needs no login. Training code is on GitHub.345
ggml-org Flint GGUF
ggml-org posted Flint-GGUF on 22 October. The pack is Apache-2.0, converted from RecursiveSI/flint, and ships three files: BF16 at 79.626 GB, Q8_0 at 42.350 GB, and Q4_K_M at 23.388 GB. Serve it with llama serve -hf ggml-org/Flint-GGUF.6
It needs a llama.cpp build with Flint support; a stock build will not load it. In Q4_K_M the small non-expert tensors (embeddings, attention, and the shared expert) stay Q8_0. The source is text-only: no mmproj, MTP, or DFlash sidecars.7
The parent RecursiveSI/flint card is still the 21 October tree. ggml's Q4_K_M sits next to Recursive's official Q4_1 at 25.993 GB and Unsloth's UD-Q4_K_XL at 25.055 GB.89
ThinkingBox-Training on Qwen3.8-27B
Microsoft and Hugging Face posted ThinkingBox-Training on 22 October. The recipe runs GRPO on Qwen3.8-27B using executable database-state rewards, with training code on GitHub. There is no Hub checkpoint for the fine-tune.1011
They report that after 50 GRPO steps, pass@1 rose from 51.70 percent to 60.78 percent, and Observed 20/20 rose from 38 to 107 of 507 tasks. pass@20 fell slightly, from 453 to 446. Those numbers are Microsoft and Hugging Face's table, not an independent run. The checkpoint was evaluated on the sealed 507-task benchmark it did not train on.12
The ThinkingBox-Bench dataset itself dates to 24 August. The earlier post explained why an agent claiming the work is done is not the same as the database agreeing. Today's post is the training loop that follows from that.13
Daily Papers flip to Oct 23
Daily Papers moved off yesterday's Cache Engineering, WildAWM, and Looking-Ahead list. Tencent's FAPL paper (Making MLLM Latents Matter) landed on 22 October at 08:43 UTC with 104 upvotes, a ShotVerse project page, and no Hub weights.1415
Deep Search in Real-World Web Environments leads the board at 112. The arXiv stamp is 19 October. GitHub OpenDS has no Hub model card.16
Deep Fusion Memory sits at 32, arXiv 18 October, with no GitHub and no Hub card. IBM LASAGNA, Policy-Mask, Physical RSI, and LaLa-World fill the rest of the board. They are papers, not checkpoints.1718
