Two billion, active.

Models

LiquidAI/LFM2.5-24B-A2B (2,565 downloads / 207 likes in two days; the Base checkpoint sits at 7, which tells you people come for the post-trained one) is the largest and, Liquid says, final model of the LFM2.5 family: a 24B-parameter reasoning model with 2.45B active parameters per token — 40 layers, 30 of them double-gated short-convolution blocks and 10 GQA, with 64 experts per MoE block routed top-4. It was trained on 36 trillion tokens, carries 131,072 tokens of context, speaks 16 languages, and is a pure reasoning model: the chat template injects a <arg_key> tag on every answer, it always thinks, and with 2B active parameters every thinking token is cheap. It is text-only, and the card says so plainly: pair it with retrieval or web search, because knowledge-heavy work is not what it is for.

Evidence. The vendor numbers (all Liquid’s own): LiveClaw 48.09 vs 38.19 for Qwen3.6-35B-A3B on agentic tool use, “up to 1.7× faster than comparable MoEs on-device,” 184 tok/s on an Apple M5 Maxand up to 21K output tok/s on one B300. A third-party summary of the same table reads it honestly: Liquid leads or matches Qwen3.6-35B-A3B on about five of eight tool-use and agent benchmarks — AutomationBench 27.69 vs 8.72 is the widest gap — and trails on SWE-Bench Verified, GPQA and the knowledge rows, which is exactly what the card’s own “pair it with retrieval” line predicts. What is not a claim is the ship set: day one the lab itself published FP8 (vLLM/SGLang), an official GGUF for llama.cpp (LFM2/LFM2-MoE and its tensor-parallel support have been merged since August), ONNX, and MLX 4-bit, with a DSpark draft model “coming soon.” Community day-zero builds exist (mudler’s APEX GGUF, MLX conversions) but all read zero downloads — the official shelf got there first, which is the point.

Critical read. The adoption number is thin for a 36-trillion-token flagship — 2,565 in two days is a fifth of what a good Qwen quant does in an afternoon — and two days is early. The license is a real decision point: LFM Open License v1.0 is free for research and for commercial use under $10M annual revenue; above that you negotiate. The honest part is that the card prints its own weakness, and the benchmark table matches it. The interesting bet: an always-thinking model lives or dies by the price of a reasoning token, and 2.45B active is about as cheap as that gets at this quality tier.

Verdict: Real deployment option for the on-device-agent slot — the complete day-one build shelf, a 131K context and cheap thinking tokens make it the natural upgrade for anyone running agents on a laptop; adopt it for agents, not for answers, and read the license before you ship anything.

openbmb/MiniCPM-V-4.7 crossed the 200-download gate since yesterday’s one-liner: the 35B-A3B at 630 downloads / 493 likes and the 1B at 263 / 294, both Apache-2.0. The 35B-A3B is the first MoE model in the MiniCPM-V series, and the interesting part is not the benchmark charts (which ship as images, so there are no text rows to check — the headline claim is a higher overall score in less inference time than Qwen3.6-35B-A3B, vendor-reported) but the scheduling trick released with it: Lookahead Expert Scheduling, which predicts which experts the next tokens will route to and loads them before they are needed. The vendor claim is that this makes it “the first open-source MoE model to run smoothly on-device,” at 13+ tok/s with 3 GB of memory on an iPhone 17 Pro Max — a phone holding a 35B model at all means the experts are streaming, and LES is hiding the stalls.

Critical read. The likes-to-downloads ratio is the tell this log keeps finding: 493 likes on 630 downloads means people are bookmarking a release most of them have not run — the day-two derivative wave (GGUFs, MLX packs, an abliterated build) all read zero. llama.cpp support merged yesterday, so the receipts can arrive; today there is an on-device claim, a chart and a crowd watching. Verdict: Marginal — adopt only if you need a phone-class open VLM and accept vendor-only numbers; the LES technique may outlive the checkpoint.

Quantizations

autotrust/JEV-35B-NVFP4 (20,494 downloads / 303 likes in three days — more than the bf16 parent it quantizes, which is the cleanest adoption signal a quant build can send) takes the JEV-35B decider and puts only its routed experts in NVFP4: 10,240 expert MLPs (40 layers × 256 experts) compressed, everything else left bf16 — attention, linear attention, the shared expert, the routers, the vision tower, the MTP layer, the LM head, the System 1 adapter and the decision head with its fitted temperatures. The result: 25.6 GB download, 23.3 GiB of GPU memory instead of 72 GB / 66.5 GiB, so the family’s MoE variant runs on one 32–48 GB GPU. And the card prints the fidelity next to the compression, as every card should: Decision Index 0.3 public 59.18 vs 59.73 bf16, vision rebuild 71.23 vs 71.94, 96.6% of answers identical to bf16, same latency, same serve script (all author-measured, one rig).

The lesson generalizes: in an MoE, the routed experts are where the bytes live, so they are the only thing worth quantizing hard — the fragile parts (routers, heads, temperatures) stay full precision and the model keeps its calibration. Verdict: Real deployment option — this is the build that turns the 35B decider from a two-GPU curiosity into a single-GPU tool, with the before/after printed.

Tape (covered builds, counters moved): AtomicChat’s Qwen-Image-2.1-Turbo-Uncensored encoder GGUF 100,530 → 208,575 in three days; unsloth’s embeddinggemma-2-GGUF 41,582 → 81,504 in four; ConwayResearch’s Underdog-Saluki 27B IQ2-mix 108,900 → 269,829 in three.

Fine-tunes

autotrust/JEV-35B (base Qwen/Qwen3.5-35B-A3B, Apache-2.0) did the thing yesterday’s entry said it had not done yet: the adoption showed up. It was 360 downloads / 83 likes when covered on 12 October; today it reads 17,426 downloads / 637 likes — 48× in three days. The card also grew a section worth the click: ERQA 51.3% in one pass with zero robot training — Google DeepMind’s embodied-reasoning benchmark (400 multiple-choice questions about real robot and everyday scenes, chance ≈ 26%) — level with PhysBrain 1.5 (52.8), the best open embodied model in the report, which was trained on robot trajectories and human-interaction video, and above every other open 7–9B embodied model (42.3–47.5). A decision model trained only on text, looking at a robot scene once and answering, matching a model trained for the domain. The authors print their own error bars (±5 points at 400 questions) and note that thinking-enabled JEV-27B-VL reaches 61.5% — the System 2 tax, quantified. There is also a live demo (GuruSearch, news.guru.so) re-ranking three news engines’ results with calibrated decisions at about half a second each.

The standing caveat stands: the Decision Index 0.3 public 59.73 (vs 53.64 for the family’s own 27B-VL on the board) is the authors’ scoring with the official kit, explicitly “not a board entry,” and the 204 ms median latency is one B200. But 17k downloads in three days, a quantized build out-pulling the parent, and an ERQA row with the error bars printed is the difference between a claim and a product. Verdict: Real deployment option for the local judge/decider slot — the MoE variant is no longer a paper animal; 32–48 GB via the NVFP4 build.

Tape: nerkyor’s Coder390 (Qwen3.8-27B SFT/RLOO) 25k combined on 10 October → 83,887 on the main repo today; vllm-sr/d3 420 → 1,156 across reads; speridlabs’ iris-3b 393 → 889. Nothing else new cleared the gate.

Technical term of the day: total vs active parameters

Every model named today carries it in the name — 24B-A2B, 35B-A3B — and it is the most honest two characters on a spec sheet. A Mixture of Experts layer replaces the one big feed-forward network of a transformer block with many smaller ones — 64 per block in LFM2.5-24B-A2B, 256 per layer in the Qwen3.5-35B-A3B that JEV-35B fine-tunes. A small router looks at each token and picks the top-k experts for it (top-4 in Liquid’s case); only those few experts compute. So the model knows like 24B but computes like 2.45B: total parameters are what must fit in memory; active parameters are what get read and multiplied per token.

Why local AI cares: decoding is memory-bandwidth-bound, and an MoE only reads the routed experts, so a 35B-A3B decodes at 3B-class speed while drawing on 35B-class knowledge — the reason small-active MoEs took over the local top-ten. The catch is the other half of the sentence: quality per FLOP is not quality per gigabyte. All 64 experts still have to live somewhere, and on a phone or a laptop that is the hard part — stream them from SSD and the stalls eat the speed (this log’s 12 September entry on expert streaming), pin a fixed subset and you cap the routing, or do what MiniCPM’s LES does: guess which experts the next tokens will want and prefetch them. The remaining trade-offs are the ones today’s stories keep stepping on: some experts barely wake up, which is what expert pruning removes; different tokens pick different experts, so large batches amortize worse than on a dense model; and the experts are the bulk of the bytes, which is why the correct quantization of an MoE quantizes exactly the experts — hard, and nothing else — which is precisely what today’s NVFP4 build did.

Harness improvements

Hermes v0.22.0 (published overnight, tagged “The Threshold Release”) is the first curated minor since v0.21.0 and the promised rollup of the seven patch releases this log covered one by one: roughly 24,200 commits, 6,080 merged PRs and 7,180 closed issues from 2,520 contributors. The pieces a local-AI reader cares about: local models in one click — the agent installs and maintains llama.cpp for your hardware (CPU, CUDA, Metal, Vulkan or ROCm), rates every catalog model against your GPU and RAM, picks the highest-quality build that fits, and keeps the server lifecycle to itself; the catalog ships Qwen3.8-27B, Qwen3.8-Flash-Next with its MTP speculative head, Qwen3.6-35B-A3B and DeepSeek-V4-Flash, plus any GGUF from Hugging Face. Around it: a first-launch setup chat that runs on a free trial model so a newcomer talks to the agent before owning an API key; Connectors (Gmail, Linear, Notion through a tool gateway, multiple named accounts,/connectors); stable SemVer release channels with signed desktop packages including the Microsoft Store; a 600-plugin catalog; and a codebase a third smaller. The harness that used to assume you owned an inference stack now installs one.

Ollama v0.40.4 (final after the rc this log covered yesterday) puts decision models in the terminal:ollama run nimble --ask "Is a refund being requested?" < ticket.txt asks a typed question of a file and prints the answer, no API in sight; Kev decision models now run on the MLX runner on Apple silicon, embeddings gained a multimodal path through llama-server, and loads got faster by reusing verified manifests. The lane this log has tracked since 21 September now has a consumer-grade front door.

llama.cpp spent 13–14 October on the phone — the same week MiniCPM taught models to schedule their own experts: OpenCL Adreno (the Android GPU) got MoE expert sorting with buffer offsets (#30418) and a fixed q8_0 dp4a mul_mat_id (#30365); Arm’s KleidiAI SME2 path gained Q4_K and Q6_K matmuls (#29219); Hexagon got a configurable HTP power mode (#30423). Elsewhere: Vulkan bf16-activation matmuls (#29069), HIP 64-row MMQ tiles for RDNA3.5 (#28195), server special-token escaping without std::regex (#30243), pooled embeddings skipping the LM head (#30445), and an even layer-split fallback when no device reports free memory (#29677).

oh-my-pi v18.10.4 (13 October): approval prompts can request untruncated tool-call bodies so a compact approval can count the lines it is approving; before-model-call hooks now receive the loop context; and a fix for Claude compaction receiving another model’s reasoning from a branch archive — which Claude refuses. Pi v1.2.0,mlx-serve and the rest of the shelf are unchanged since yesterday’s entries.

The through-line: the two numbers in a model’s name are the ones that decide whether it fits your machine and how fast it thinks once it does. Today a flagship shipped that says 24 total, 2 active in its own name, a quantizer read the same spec and compressed exactly the part that matters, a VLM taught itself to prefetch the part that streams — and the harness layer started installing the whole stack for you. The spec sheet was the roadmap all along.