Hugging Face Daily Digest Oct 22, 2026

Recursive Flint is on Hugging Face's default trending list this morning: an Apache-2.0 39.8 billion parameter coding Mixture-of-Experts from RecursiveSI, with official INT4 and NVFP4 llama.cpp files beside it. Tencent added three smaller official GGUFs of Hy-Memory-30B-A3B. IBM Granite posted Granite Speech 5.0, announcing a 2B successor. Daily Papers moved to an October 22 stack led by Cache Engineering, with Looking-Ahead Apache-2.0 weights on the Hub.

On the Hub

Recursive Flint coding MoE

RecursiveSI published Flint on 21 October. It is a decoder-only Mixture-of-Experts coding model, 39.8 billion parameters with about 2.5 billion active, Apache-2.0, and a 262,144 token context.¹

The BF16 tree is 79.591 GB. Transformers modeling code is not in the repo yet. The card serves it with the in-repo SGLang kit on one B300 or H200. Official INT4 GGUF is 25.993 GB and official NVFP4 GGUF is 22.864 GB, both dated 21 October.²³

Unsloth's Dynamic GGUF ladder of Flint opened on 20 October and now sits on the GGUF trending list, with IQ1_M at 11.352 GB.⁴ Recursive's own post claims more than 125 tokens per second on an M4 Pro and more than 150 on a 64 GB RTX Spark at short context through its rec engine. Those figures are Recursive's, not llama.cpp.⁵

Tencent Hy-Memory smaller GGUFs

Tencent's official llama.cpp pack of Hy-Memory-30B-A3B grew this morning from three files to six. The card was last modified on 22 October at 06:50 UTC. The new files are Q3_K_M at 14.45 GB, expert-mixed IQ2_XXS at 9.14 GB, and expert-mixed IQ1_M at 8.03 GB. The original Q8_0, Q6_K, and Q4_K_M files remain. The tree is now 106.554 GB.⁶

Hy-Memory is a memory writer for Mem0 and MemOS. The BF16 parent is still the 15 October card.⁷ Tencent lists LoCoMo Judge at 57.25 percent for both Q3_K_M and the 8.03 GB IQ1_M mix, versus 58.78 percent for Q6_K. IQ1_M and IQ2_XXS keep Q8_0 on the rest of the model. Those scores are Tencent's.

IBM Granite Speech 5.0

IBM Granite posted Granite Speech 5.0 on 21 October. The post announces granite-speech-5.0-2b as the successor to 4.1-2B, plus revisions of the 470m Turbo CTC and 4.1-2B-plus cards.⁸

The 2B card is Apache-2.0, 5.050 GB, and dates to 7 October. IBM claims an Open ASR Leaderboard average English WER of 5.06 percent, second among listed open-weight ASR as of 21 October. Support is claimed in Transformers, vLLM, llama.cpp, and mlx-audio. Those figures are the post's.⁹

An official llama.cpp pack exists on the Hub. That pack is not on the trending lists.¹⁰

Daily Papers flip to Oct 22

Daily Papers moved off yesterday's GUI Agents, PAR, Ropedia W0, and Sage-A1.5 list. Cache Engineering for Agents leads with 65 upvotes and has a project page with no Hub model card.¹¹ WildAWM sits at 57. Its Hub card is a community 35B-A3B from 18 October.¹²

Looking Ahead sits at 55. TencentARC posted Apache-2.0 image-to-video weights, 67.882 GB, created 21 October at 12:08 UTC.¹³ The paper is arXiv 2610.29989.¹⁴