HF/Daily
THU 23 JUL 2026
pulled 00:09 UTC · 24 Jul

A thin day on the Hub, a loud day on the wire. Nothing frontier-scale landed on Hugging Face, but Upstage, Motif and Nanbeige put fresh Korean and Chinese open weights on the board, HF quietly dropped The Stack v3 — 4.9T tokens of code — and Microsoft open-sourced an MIT-licensed image model. Off-Hub, the White House accused Moonshot of distilling Anthropic's Fable to build Kimi K3 — four days before K3's weights are scheduled to go public.

Daily papers
19
Top entry: DeepSeek-V4 post-training on Ascend SuperPOD, 45 upvotes.
Notable drops · 72h
6
Two Korean MoEs, a 3B Apache LLM, an MIT image model, a security agent, one corpus.
Biggest new weights
250B
upstage/Solar-Open2-250B — MoE, EN/KO/JA, landed 22 Jul.
Stack v3 corpus
4.9T tok
713 languages, 173M repos, ODC-By. Released today.
Channel · Models

What the Hub is actually chasing

Trending score is Hugging Face's own momentum metric — recent likes, downloads and activity — not a quality signal. Read it as attention, not as a benchmark.

Top trending models

trending score · hover a row for detail
0400800
Table view — same data as numbers
#ModelScoreDownloadsLikesCreated

The same 30 models, sorted by what they are

count of top-30 trending
0612

A third of the trending board is the community quant-and-uncensor economy — GGUF, MLX, 1-bit and abliterated repacks of models somebody else trained. That is the Hub's real center of gravity, and it moves faster than the labs do.

Channel · New in 72h

Everything that actually shipped

RepoFromSizeWhat it isLicenseDate
HuggingFaceCode/stack-v3-train Hugging Face 4.9T tok The Stack v3. Direct GitHub crawl (Aug 2025 snapshot) replacing the 2023 Software Heritage base; 713 languages, 173M repos, contents inlined, language-agnostic near-dedup. ODC-By 23 Jul
upstage/Solar-Open2-250B Upstage (KR) 250B MoE Largest new open-weight release of the week. EN/KO/JA, vLLM-ready. 407 likes on 362 downloads — everyone is watching, almost nobody can run it yet. custom 22 Jul
Nanbeige/Nanbeige4.2-3B Nanbeige 4.2B dense Small EN/ZH instruct model with published eval results and a live demo Space. The only genuinely edge-sized new release of the window. Apache-2.0 21 Jul
microsoft/Mage-Flow Microsoft 4.1B Rectified-flow text-to-image and instruction editing in one 4B model, with a diffusers pipeline. Turbo and Edit variants already mirrored by the community. MIT 21 Jul
fdtn-ai/antares-1b Foundation AI 1.8B RL-trained terminal agent for vulnerability detection, built on IBM Granite 4.0 MoE-hybrid. A security co-pilot small enough to run on a Jetson. Apache · gated 21 Jul
Motif-Technologies/Motif-3-Beta Motif (KR) 315B MoE Long-context multilingual MoE, EN/KO, shipped as a preview with no license file yet — treat as look-don't-deploy until they publish terms. none stated 20 Jul
Channel · Papers

Daily papers, 23 July

Nineteen papers listed, and the shape of the list is telling: the top entry is a systems paper about training on non-NVIDIA silicon, and four of the top ten are about video, retrieval or robot policy rather than language modeling.

Ranked by upvotes

top 10 of 19
02550
Channel · The wire

Off-Hub, and bigger than the Hub

Allegation · unproven

White House accuses Moonshot of distilling Fable

OSTP director Michael Kratsios said publicly that Moonshot AI distilled Anthropic's Fable to build Kimi K3, calling it covert industrial distillation aimed at US technology, and separately alleged Moonshot obtained restricted Nvidia GB300s routed through Thailand.

The supporting evidence is circumstantial: Redwood Research found K3 identifies itself as Claude disproportionately often, which web-scraped training data also explains. Distillation is genuinely hard to prove from outputs alone. Treat this as a contested claim, not an established fact.

Dated · 24–27 Jul

DeepSeek V4 and Kimi K3 weights land this week

Two major open-weight releases are scheduled inside the next four days — DeepSeek V4 on the 24th, Kimi K3 on the 27th. K3's weights now arrive into an open provenance dispute, which is an adoption question for anyone with compliance or government exposure, separately from how well the model performs.

OpenAI

Presence, and a $30B data center

OpenAI launched Presence, an enterprise agent deployment platform with governance, permissions and audit across voice and chat — and announced Project Camellia, a 3.2-gigawatt Georgia campus with power delivery phased across 2028–2032. Enterprise agent platforms are now a four-way race with Google, Meta and NVIDIA–ServiceNow, competing on governance rather than model quality.

Google · 21 Jul

Three Gemini models, still no 3.5 Pro

Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber — a security-tuned variant limited to governments and trusted partners. Gemini 3.5 Pro has now missed its target repeatedly and remains unshipped. Google says its largest pretraining run yet is underway for Gemini 4.

Channel · Fleet relevance

What of this touches our stack

None of the numbers on this page were measured on our hardware. Per the model-diary law, anything below is a candidate to bench, not a verdict.

Mage-Flow
Strongest single pickup. MIT-licensed T2I plus instruction editing at 4B — it fits our 96 GB box easily and it is commercially clean, which is exactly the constraint that keeps biting us on CC-BY-NC weights. Worth a ComfyUI lane next to the Ideogram path.
antares-1b
1.8B RL-trained vulnerability-detection terminal agent. Small enough for a 24 GB laptop-GPU node or a Spark to hold permanently. Natural backing model for the cso audit skill, which currently spends frontier tokens on work a specialist could pre-filter.
Nanbeige4.2-3B
Apache-2.0, 4.2B, EN/ZH. The only new release in the window sized for a Jetson-class edge node or an always-resident lane. Worth a same-prompt run against whatever currently holds the small-model slot.
OvisOCR2
Not new today but climbing — 0.8B base, Apache-2.0, tuned for document parsing to markdown with tables and formulas. Directly relevant to the OCR intake alias in the doorbell registry, where the current backing model is much larger than the job needs.
MOSS-Transcribe-Diarize
Already our Scribe backend on our 96 GB box, and still trending at 86 with 111K downloads two months post-release. No action — just confirmation that the single-pass ASR + diarization call was the right one.
Kimi K3
Flag, not an emergency. We run the kimi CLI as a super-team lane. The provenance dispute is about the weights and the vendor, not about our usage — but it is a fact worth knowing before K3 goes anywhere near client-facing work.
Stack v3
4.9T tokens, ODC-By, 15.9TB for the training subset. Only matters if the fleet-model-9B fine-tune line restarts — but it is now the most current code corpus in the open, and the old Stack v2 base is three years stale.
Solar-Open2 / Motif-3
250B and 315B MoE. Both are past what our 96 GB box holds at usable quant, and Motif-3 ships with no license file. Watch for GGUF; do not plan around either.