🛠 llama.cpp v0.7.0 Moves llama-server's Default Port From 8080 to 9931

Front-ends, scripts and agent harnesses hard-coded to localhost:8080 stop connecting after the upgrade: pass --port 8080 or repoint them. Docker images keep 8080 inside. The Oct 15 release also tags the MoE expert cache we covered as merged, plus LiliCorr speculative decoding.

🔗 https://github.com/ggml-org/llama.cpp/releases/tag/v0.7.0

#LLM #localModel #AItools

🔬 Microsoft's bitnet.cpp Has Been Giving Wrong Output on CPU for Months, an Independent Check Finds

Against an HF reference, BitNet b1.58 2B4T's perplexity runs ~6× too high on x86 (SiLU where the model uses ReLU², issue #588, open since July) and ARM prints garbage. No fix is merged. ik_llama.cpp is correct on both; PrismML's Bonsai 1.7B was correct wherever it loaded.

🔗 https://github.com/matheus-c-feitosa/lowbit-atlas

#LLM #localModel

🛠 NodeWatch: 98 of ComfyUI's 300 Most-Downloaded Node Packs Have Gone a Year Without a Commit

A free tracker re-checks the top 300 Comfy Registry packs nightly. The 98 idle ones total ~22.7M downloads; ComfyUI_essentials, rank 7, has been maintenance-only since April 2025. It suggests active forks, and warns that "stale" can also mean finished and stable.

🔗 https://phluffhead26.github.io/nodewatch/

#comfyUI #AItools

🤗 FastH3 Omni: MiniMax H3 Reference-to-Video in 8 Steps, Open Weights

Hao AI Lab's 8-step H3 distill turns a prompt plus up to 12 image, clip or voice references into video with synced sound; ComfyUI 0.42.2 ships a template. The 5.6–11.9× speedup is an 8×GB200 figure. The H3 licence bars use in the EU, UK, US and South Korea, outputs included.

🔗 https://huggingface.co/collections/FastVideo/fasth3-omni

#videoGen #comfyUI

🤗 FIBO 2.0: Bria's 8B Image Generator and Editor, Non-Commercial

An 8B DiT with a Qwen3-VL 4B encoder, plus a 4-step edit-only turbo. It leads open weights on Artificial Analysis's edit board at 1087 Elo, within Qwen-Image 2.1's margin (1083). ComfyUI support is merged to master, not yet in a tagged release. CC BY-NC 4.0.

🔗 https://huggingface.co/briaai/Fibo-2.0-dev

#imageGen #comfyUI

🤗 Kyutai's Pocket TTS Small: CPU Voice Cloning at 18.2M, 6.9M and 1.9M Parameters

Three English models in Pocket TTS v4.0.0. Only the 18.2M matches the ~100M original, Kyutai says (WER 0.85% vs 0.92%); the experimental 1.9M nearly doubles it (1.72%). Weights CC BY 4.0, code MIT. Cloning needs an auto-approved gate; the ungated build has preset voices only.

🔗 https://kyutai.org/blog/2026-10-20-pocket-tts-small/

#AItools #localModel #openSource

🧠 KRAFTON's Raon-1 mini preview: a From-Scratch 30B-A4B MoE Under Apache 2.0

A new hybrid-attention MoE (28.8B total, 4.3B active) for Korean chat and agents. It beats Qwen3.6-35B-A3B on IFBench (84.2 vs 66.6) and trails it on MMLU-Pro (76.5 vs 85.9). The Q4_K_M GGUF is 18.3 GiB, but stock llama.cpp can't load it yet: build KRAFTON's own branch.

🔗 https://huggingface.co/KRAFTON/Raon-1-mini-preview

#LLM #localModel #openSource

🧠 Qwen3.8-Flash-Next at 1.66 bpw Fits One RTX 5090, No CPU Offload

turboderp's new EXL3 quant packs Qwen's 125B-A6B MoE into mostly 1.5-bit experts: 28.2 GiB on a headless 5090, with the 18.5 GiB n-gram table streamed from disk. Context tops out at 64K with a K5V4 cache; quality is shown only in charts. Qwen Community License; ExLlamaV3 only.

🔗 https://huggingface.co/turboderp/Qwen3.8-Flash-Next-exl3/tree/1.66bpw_h4_ng3

#LLM #localModel

🧠 Flint: Recursive's From-Scratch 40B-A2.5B Coding MoE, Apache-2.0, With Official GGUFs

Recursive's first open model activates ~2.5B of 39.8B params per token and is built for coding agents. Its NVFP4 GGUF is 22.9 GB and llama.cpp runs it from build b11803, but Recursive recommends 48 GB and tested no 32GB card. Self-reported: 74.2% SWE-bench Verified.

🔗 https://huggingface.co/RecursiveSI/flint

#LLM #localModel #openSource

🤗 IBM Granite Speech 5.0 2B: Apache-2.0 Transcription and Translation in Six Languages

IBM's 2.3B speech model transcribes English, French, German, Spanish, Portuguese and Japanese, and translates to and from English. It averages 5.06% WER on the Open ASR Leaderboard's English sets, behind Qwen3-ASR-1.7B's 4.94%. IBM's Q8_0 GGUF plus audio projector is 3.1 GB.

🔗 https://huggingface.co/ibm-granite/granite-speech-5.0-2b

#localModel #openSource #AItools

🛠 LTX Desktop 1.3.1 Adds an Assistant That Plans and Queues Local Video Jobs

Lightricks' Apache-2.0 app now turns plain requests into queued jobs, via the Gemini API with your key or a local Qwen 3.5 9B. It adds 1440p (20 s) and 4K (10 s) output and says a 20 s 1080p clip now fits 32 GB cards. LTX-2.5 weights need a paid licence above $10M revenue.

🔗 https://github.com/Lightricks/LTX-Desktop/releases/tag/v1.3.1

#videoGen #AItools

🛠 llama.cpp's Web UI Can Now Browse and Download GGUFs From Hugging Face

Since build b11714 a Discover Models page searches the Hub's GGUFs, shows a minimum-memory estimate per quant, pauses and resumes downloads, and can fetch a repo's matching DFlash or MTP draft head. It is off by default and needs router mode: switch it on in Settings.

🔗 https://github.com/ggml-org/llama.cpp/pull/29584

#LLM #AItools

🛠 chaiNNer v0.26.0: TensorRT Upscaling, and the Last Release on Its Python Backend

The GPL-3.0 node-based image-processing app's first release in a year adds TensorRT upscaling and engine conversion with BF16, an optional install for RTX 20xx / GTX 16xx or newer on driver 580+. Its maintainer calls it the last Python-backend release; a Rust rewrite is coming.

🔗 https://github.com/chaiNNer-org/chaiNNer/releases/tag/v0.26.0

#AItools #openSource

🔬 ORL: a Diffusion LLM Decoding 2,788 Tokens/s for One Request on an RTX 5090

Together AI-led researchers post-train diffusion LMs on their own denoising trajectories so each pass commits more tokens. In their dlmrt engine, LLaDA2.1-mini + ORL in NVFP4 (9.88 GiB) hit 2,788 tok/s on one HumanEval request. Code MIT/Apache-2.0, weights Apache-2.0.

🔗 https://github.com/Introspective-Diffusion/ORL

#LLM #localModel #openSource

🧠 MemT-1.0-Nano Gets a Consumer Path: Chat Memory Stored in the Model, Not the Context

MemTensor's Apache-2.0 Qwen3.6-35B-A3B variant files past turns into a 40 MiB latent memory on /commit, then answers without re-reading the chat. A community exl3 build is 15.67 GiB at 3 bpw and needs the MIT exl3-qlora fork; its demo ran 6 bpw on two 3090s.

🔗 https://huggingface.co/UnstableLlama/MemT-1.0-Nano-exl3

#LLM #localModel #openSource

🛠 MOSS-Image Learns to Edit: an Edit LoRA, With Native Support Merged Into ComfyUI

chinoll's ~284 MiB LoRA changes colours, outfits, lighting, seasons and style from text, on MOSS-Image Base or SFT. kijai's PR #17296 adds the reference input to ComfyUI master, not yet in v0.42.5; Chenkin-x's node pack runs it now with up to four references.

🔗 https://huggingface.co/chinoll/Moss-Image-Edit

#imageGen #comfyUI

🧠 Spark-X2.6-4B: SparkLLM's On-Device 4B Retrained for Agents, With Three Thinking Levels

Same architecture as Spark-X2.5, new post-training for coding, tool use and long context, plus none/low/high thinking. Apache-2.0; the Q4_K_M GGUF is 3.30 GiB and runs on llama.cpp b10828+. Its benchmark claims, incl. matching ~9B models, appear only as charts.

🔗 https://huggingface.co/XHToken/Spark-X2.6-4B

#LLM #localModel #openSource

🛠 Coupled Initialization: One ComfyUI Node for More Varied Batches at the Same Cost

Swap it in for RandomNoise and a batch starts from coupled noise; at K=2 the pair is antithetic (ε, −ε), nearly doubling the starting distance between images (2.01 → 3.96). MIT, no dependencies, from the authors' ICLR 2026 work; workflows cover SD1.5, SDXL and FLUX.1-schnell.

🔗 https://github.com/jjia131/ComfyUI-CoupledInitialization

#comfyUI #stableDiffusion #flux #openSource

🛠 llama.cpp b11821 Puts Intel Arc's XMX Engines to Work on Speculative Decoding

Merged PR #29864 runs 2–80-token batches on XMX matrix units in SYCL. On a 32GB Arc Pro B70 with Qwen3.8-27B and MTP drafting, Q4_K_S went 38.7 → 47.6 tok/s and Q2_K 30.2 → 60.0. Plain decode barely moves, and only the B70 was tested.

🔗 https://github.com/ggml-org/llama.cpp/pull/29864

#LLM #localModel

📡 MiniMax Posts M3.1-Flash's Weights: 428B Total, Far Too Big for One GPU

Four days after the API launch, the agentic/coding model's weights are up: ~428B total, 23B active, 1M context, 235 GiB on disk. The MiniMax Community License allows commercial use only with a "Built with" credit and a notice, and written approval above $20M yearly revenue.

🔗 https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash

#LLM #AItools