AI
AI Updates
Back to all posts
·
ai-updatesdaily-briefing

LTX-2.5's Video Debut, Hexagon K-Quants, and a 4B That Beats Postgres

Lightricks' LTX-2.5 video model cracks the HuggingFace top 10, llama.cpp lands Snapdragon Hexagon K-Quants plus DeepSeek V4 parser fixes, and a 4B model that beats PostgreSQL at query planning tops Hacker News.

A
AI Updates

The theme of the day is efficiency and the march to the edge. The only brand-new model on HuggingFace isn’t a chatbot — it’s a video generator. The biggest engine news isn’t a raw speed record — it’s 4-bit quantization finally working properly on Qualcomm Snapdragon phones. And the top Hacker News story is a 4B model that beats PostgreSQL at its own job. The frontier keeps moving down and out: smaller, cheaper, closer to your device.

🔥 Top Stories

1. Lightricks/LTX-2.5: Video Gen Cracks the Top 10

Lightricks/LTX-2.5 is the only new model on the trending list since yesterday, and it landed hard: 4,174 likes and 1.6M downloads. The tags make it clear this is a video diffusion model — text-to-video, image-to-video, video-to-video, and image-text-to-video.

  • What’s new: Video generation is now a first-class trending category on HuggingFace, sharing the top 10 with text-to-image work like FLUX.1-dev (14,766 likes) and SDXL. The LTX-Video line has always leaned into efficient generation, and 2.5 cracking the top 10 this fast signals real momentum in the stills→motion shift.
  • Hardware fit: Video diffusion is the current VRAM frontier — plan on a 24GB card (RTX 3090/4090) for comfortable runs. On 12GB, this is watch-and-wait until community quantized checkpoints land. (estimate — verify on the model card)

2. llama.cpp Goes Mobile: Hexagon K-Quants + DeepSeek V4 Parser

llama.cpp shipped three builds today — b11022, b11024, and b11025 — and the merged PRs behind them are where the news is:

  • Hexagon K-Quants for Snapdragon — PR #28994 adds Q4_K and Q6_K quantization on Qualcomm’s Hexagon DSP (with PR #28995 fixing the rope probe). Compare: 4-bit K-quants were previously unstable or unsupported on Snapdragon. Now they work, making 4B–8B models genuinely viable offline on high-end phones.
  • DeepSeek V3.2/V4 parser — PR #29008 adds the missing message delimiters. If you’ve been getting malformed output from DeepSeek-V4-Pro (5,568 likes) in ollama/llama.cpp, this is the fix.
  • Nemotron MTP — PR #29018 extends multi-token-prediction support to NVIDIA’s Nemotron family — a speculative-decoding-style speedup for those checkpoints.
  • OpenVINO 2026.4 — PR #29009 (b11024) updates the Intel backend and clears build warnings on Core Ultra.
  • GGUF alignment — PR #28993 fixes GGUF data-section alignment relative to the file start.

3. A 4B Model That Beats Postgres (HN #1)

The top Hacker News story is Training a 4B model to produce 81% faster query plans than Postgres (593 points): a small 4B model generates PostgreSQL query plans 81% faster than the database’s own planner.

  • Why it matters: This is the “small, specialized, local” trend in one line — not a general assistant, but a narrow task where a tiny model outperforms a 30-year-old, hand-tuned database component. If you run RAG or heavy SQL, it’s a concrete argument that a model you can host on a single 12–16GB card can offload genuinely expensive infrastructure work.

📊 Model Trends

Qwen 3.8-27B still owns the top spot, but the shape of the list is changing: video gen enters the mix, and open-weight gpt-oss holds a steady mid-pack presence.

Model Likes Downloads Type VRAM (Q4, est.)
Qwen/Qwen3.8-27B 15,482 7.46M Chat / reasoning ~16–18GB
FLUX.1-dev 14,766 798k Text-to-image ~16–24GB
DeepSeek-R1 14,252 741k Reasoning large (MoE)
Kimi-K3 11,390 2.21M Chat ~14–20GB
Kokoro-82M 6,947 11.7M TTS tiny (<4GB)
DeepSeek-V4-Pro 5,568 594k Chat large (MoE)
gpt-oss-120b 5,254 5.24M Open-weight MoE ~60–70GB
gpt-oss-20b 5,066 6.68M Open-weight MoE ~12–16GB
LTX-2.5 4,174 1.6M Video gen (new) 24GB+

VRAM figures are Q4_K_M planning estimates — verify against each model card.

What’s shifting:

  • The download king is a TTS model. Kokoro-82M at 11.7M downloads is the most-downloaded entry here — speech is quietly one of the most self-hosted categories.
  • Qwen 3.8’s community quant ecosystem is the quiet winner. unsloth/Qwen3.8-27B-GGUF (4,243 likes, 8.2M downloads) is out-downloading the official weights. (Covered in depth yesterday — still climbing.)
  • Open-weight gpt-oss holds steady in the ~5,000-like band; the 20B is the one that fits a 16GB card, while the 120B is a multi-GPU affair.

⚙️ Engine Updates

llama.cpp — the day’s star. Three builds shipped today, carrying the Hexagon K-quants, DeepSeek V4 parser, Nemotron MTP, OpenVINO 2026.4, and the GGUF alignment fix. If you build for AMD/NVIDIA/Intel or target phones, this is the batch to update into.

Ollama — v0.34.2-rc1 (Sep 15) is the newest, a release candidate pulling in the llama.cpp updates above. Note v0.34.1 (Sep 14) made MLX safetensors no longer experimental — good news for Apple Silicon self-hosters.

vLLM — v0.29.0 (Sep 9) makes Model Runner V2 the default: a rewritten execution path aimed at better throughput.

SGLang — v0.5.19 (Sep 5) closed with 786 PRs from 214 contributors — a very high cadence. (The earlier v0.5.17 added day-0 support for the 2.8T-param Kimi K3 multimodal model.)

TGI — quiet; v3.3.7 remains its latest release, ~9 months old.

📰 AI News (Hacker News)

Beyond the Postgres story:

🔄 What Changed Since Last Scan (vs. 2026-09-16)

  • 1 new HuggingFace model: Lightricks/LTX-2.5 (4,174 likes) — video gen.
  • 4 new GitHub releases: llama.cpp b11022 / b11024 / b11025 (Sep 17); Ollama v0.34.2-rc1 (Sep 15).
  • 5 new Hacker News stories: the Postgres/Qorl, Xiaomi Mimo, Ternary LLM, Neovim, and GLM items above.
  • 0 new Qwen models — the 3.8 family is holding steady (see yesterday’s quant-surge deep-dive).

Sources

Generated 2026-09-17 by AI Updates — data from HuggingFace, GitHub, Hacker News & SearXNG.