WHAT’S NEW IN LOCAL AI

The log

Published when something is actually worth your read. Silence is a normal outcome, not a gap.

18 OCT 2026 / MODELS + IMAGE

Writes the words.

OpenMOSS opened MOSS-Image — a 5B diffusion transformer built from scratch for the thing open image models are worst at: legible text and exact layout, driven by a JSON scene graph, under Apache-2.0. Its Comfy repack is at 90,456 downloads in two days and an 8-step DMD build at 12,410. Tencent’s Hy-Memory-30B-A3B is a memory writer that replaces the general LLM inside Mem0 and MemOS, with a GGUF at 2,917 on day one. DeepSeek’s V4.1-Flash-1017 is a day old at 1,107 likes and zero downloads, and classifier-free guidance — why every good image costs two forwards a step — explained.

17 OCT 2026 / MODELS + QUANTIZATION

Forty-nine pages a second.

Datalab opened Indra OCR — a document reader whose main recognizer has one million parameters and whose whole model shelf is 18 MB of ONNX, claiming up to 49 pages a second on an ordinary CPU with a better headline score than the full PaddleOCR pipeline. BFL’s own ComfyUI repack of FLUX 3 Klein out-downloaded the model it repacks three-to-one overnight, NVIDIA’s Spark-validated Flash-Next FP4 build stays container-locked, Ollama added ternary Q2_0 support, and the importance matrix — why a 4-bit GGUF still reads like the original — explained.

16 OCT 2026 / MODELS + IMAGE

The small one wins.

Black Forest Labs opened the weights on FLUX 3 Image Klein — 5.6 billion parameters its internal evals rank above every open image model and two closed ones, with an INT8 build that runs on 8 GB of VRAM and a license that lets a small business sell the output. MiniCPM-V 4.7’s official GGUF shelf crossed 100k downloads (10.9× the bf16 checkpoint) and picked up a DSpark drafter, NVIDIA’s 676k-download Mamba hybrid got a ternary 1.58-bit MLX build, Hermes shipped the release it meant to, and flow matching — why 16 steps is enough — explained.

15 OCT 2026 / MODELS + AUDIO

Open mic.

Fish Audio opened the weights on Drama 3 — a 35B MoE TTS you direct like an actor: sixteen voices per script, bracketed stage directions, and word-level edits that re-render only the words you named. A community GGUF out-downloaded the checkpoint in two days, the C++ audio engine layer keeps widening, Ollama taught its MLX backend to run on NVIDIA cards, llama.cpp merged NVIDIA’s LiliCorr speculation, and residual vector quantization — the trick that turns sound into tokens — explained.

14 OCT 2026 / MODELS + HARNESS

Two billion, active.

Liquid AI’s LFM2.5-24B-A2B puts it in the name: 24B parameters of which 2.45B wake up per token, shipped day one with an official GGUF, MLX and ONNX builds and a license that is free under $10M revenue. MiniCPM-V 4.7’s first MoE crossed the gate with a trick that prefetches its own experts on a phone, the JEV-35B decider went 48× in three days and its NVFP4 build out-downloads the bf16 parent with the fidelity table printed, and Hermes v0.22.0 makes local models a one-click install. Plus total vs active parameters, explained.

13 OCT 2026 / MODELS + MEASURED

Everyone ships a decider.

Five decision models from five unconnected directions in eight days — a 27B published by the vLLM router project that answered all 140,178 public requests and printed the score, a 4B that took first place on the public board at 29 ms a question, a 31B whose readout lives in a Python package rather than the weights, a frozen backbone with a 57 MB physics engine, and a 43 MB ONNX file that turns speech straight into tool calls. Plus the conformal set, a quant shelf whose downloads beat the model it came from eight to one, and llama.cpp and Ollama merging decision support in the same weekend.

12 OCT 2026 / QUANTIZATION + MEASURED

Now it fails loudly.

The ternary 27B that loaded silently and produced gibberish in stock llama.cpp now runs correctly — or refuses to load at all, which is the part that matters, and the pull request prints the perplexity that proves both halves (1,258,506.97 before, 10.2343 after, KLD zero). Also: two new open document parsers, a decision model that reads audio and video, the Qwen-Image-Turbo quant shelf, a two-Spark pruned GLM for 16k downloads, and the decision lane’s biggest lab yet.

11 OCT 2026 / QUANTIZATION + MEASURED

How much survives the squeeze.

Three weekend releases all squeeze a model and print what falls out. Qwen’s 8-step image checkpoint crossed the gate in two days; a 12B rewrite model ships 2-bit files it actually trained instead of rounded (70% → 87% top-1 against bf16); and an abliterated image-text encoder with 100,530 downloads comes with the authors’ own measurement showing it changes nothing about the pictures. Plus llama.cpp’s multimodal drafter work, Ollama walking back its background upgrade, and mlx-serve landing on Linux.

10 OCT 2026 / MODELS + MEASURED

Zero tokens, one honest number.

Liquid AI open-weights two decision models that answer a typed question in one forward pass and generate nothing at all — 8 ms a question on a 4090, 30 ms on an M5 Pro — which only works if 0.62 means 62%. Also: the 7.89 GB Qwen3.8-27B that beats its 54 GB parent at tool calling, Z.ai’s expert-tiered Maya quants, JetBrains’ 12B coding MoE, a 34 MB Turkish TTS, and the single-5090 engine with 400k downloads per artifact this log never named.

7 OCT 2026 / MUSIC + MEASURED

The music stack nobody covered.

MiniMax’s 11B song generator has sat on HuggingFace since August at 7k downloads; the C++/GGML desktop app built around it has pulled its GGUFs 101k times and trains LoRAs on your own tracks. Also measured here: a 16.9 MB ASR that turns speech into tool calls in 123 ms, Perplexity’s Decision Index lead, and the prompt expander that became its own model family.

6 OCT 2026 / MODELS + ENGINES

The open judge grew eyes.

The decision lane’s open side: a family distilled from the closed TypeSafe Jev API — 2.16M downloads in a fortnight — as a 451 MB bolt-on that judges text and images in one forward pass, with a third-party GGUF whose fidelity table is printed. Robot arms and computer use included; the failure rows too.

5 OCT 2026 / MODELS + MEASURED

Five levels beat three.

Three days after the ternary parakeet, Fermion’s 164 MB five-value rebuild of the same NVIDIA model beats it on every axis. Measured here: an hour of audio in 18.5 seconds (196×) on the M3 Ultra — plus the silent sentence drops and hotword misses the card doesn’t print.

4 OCT 2026 / ENGINES + MODELS

The wall grew a door.

llama.cpp merged the systemone API and Clef in 48 hours, Ollama serves it with images, a 4-bit MLX pack prints its own fidelity check — and the most-downloaded local Clef is a decoy with no decision head in it. Plus Aleph-Alpha’s first open weights.

2 OCT 2026 / MODELS + HARNESS

System One, 27B.

Cloudflare joins the decision-model lane with a 27B prefill-only scorer — Apache-2.0, vision included, speaking the Jev/Ollama systemone API natively. Its 9B sibling wins half the benchmark rows, and no engine anywhere can run either one yet. Plus: a Mac engine ships a +66% headline that doesn’t divide against its own example.

1 OCT 2026 / MODELS + QUANT + TECH TERM + HARNESS

System One, factory-fitted.

A reranker op lands in llama.cpp, classifiers become a first-class Pi tool, a 144M encoder ships its own failure table, and GLM-5.3-Flash’s 1.8M-downloads-a-month quant shelf finally gets a native mainline path. The MiMo repetition-fix quants went 10× overnight.

29 SEP 2026 / MODELS + HARNESS

A $90,000 fix for a bug RL caused.

Xiaomi published why MiMo-V2.6’s agents loop: RL amplified tool-call flooding because the penalty threshold sat above it — $2.31M to retrain, $90K for the distillation fix. The drop-in checkpoints are up, and one 24 Sep trap is still on disk.

29 SEP 2026 / MODELS + QUANT + TECH TERM + HARNESS

Don’t answer. Choose.

Ollama gives decision models their own endpoint, a Stanford-line crew ships CLM-8B to win them, ISTA deletes half a model’s experts and it still codes, and antirez’s DeepSeek GGUF crosses one million a month.

27 SEP 2026 / MODELS + TECH TERM + HARNESS

The stenographer that never stops.

A streaming ASR model with constant memory, a ternary ASR encoder that fits in 178 MB, Apple’s compressed-document VLM, rolling KV caches explained, and Ollama flipping Apple Silicon to MLX by default.

24 SEP 2026 / FLEETS + MODELS

The MiMo recipes arrived.

Two days after the gate: measured serving recipes for every Spark fleet size, a community NVFP4 re-quant that beat the official checkpoint at the same measured fidelity — and the three traps that explain the model’s mixed reputation with tools.

24 SEP 2026 / TECH TERM + QUANT

How models lose the filter.

Abliteration — the mechanism behind the week’s fastest-moving quant builds. Plus unsloth’s Qwen-Image-2.1 GGUFs, the MiMo build wave crossing every watchlist trigger, and llama.cpp v0.5.0.

23 SEP 2026 / MODELS

MiMo crosses the gate.

The promised re-check: all three MiMo-V2.6 repos cleared the gate with the day-one build shelf in tow, a third-party index put the flagship on top — and grammar-constrained decoding, explained.

22 SEP 2026 / MODELS + HARNESS

Xiaomi published the middle of the recipe

MiMo V2.6 ships the weights, the report and the whole RL machine — and the most interesting file is a 9B labeled “a starting point for open agentic RL.” Plus harness week: Step Code and Unreal Agent quantify the scaffolding.

Looking for the old daily posts? The previous pipeline’s 17 posts live in the Historic Slop archive, kept for the record and the permalinks.