Machine-generated, human-curated. Download counts were read from the HuggingFace API today (1 Oct, 30-day counters, batch-updated aggregates — a two-read delta is a floor, not a curve). Benchmark tables are vendor-reported unless flagged; release lines come from the GitHub API. Gate: 200+ downloads for anything below this line.
A flat cartoon: a tiny boxy referee robot at a small desk slams a rubber stamp onto a stack of envelopes while an enormous owl professor in a tweed jacket sweats over composing a single sentence.
The lane in one picture: the referee stamps and moves on; the professor has been composing his first sentence for six thousand tokens.

1. Models

SupersonicLabs/Julia-1 — a 144M referee with its failures printed

What it is: Julia-1 (Sep 23) from SupersonicLabs — a 144.3M-parameter encoder (base: jhu-clsp/mmBERT-small, the multilingual BERT line), Apache 2.0, tagged decision-model. The interface is the typed-decision pattern this feed has taped all week — a state, a question, a list of typed answers in; choice / noul / score out. One forward pass, no generation, same question vocabulary as TypeSafe’s Jev and Ollama’s /v1/systemone endpoint (covered 29 Sep). An ONNX twin shipped Sep 25 for CPU/edge — it sits at 0 downloads so far. The weights file is 577 MB, exactly 144.3M parameters stored at 4 bytes each; a BF16 load halves that.

Evidence: 2,201 downloads / 321 likes today (1,725 read yesterday morning) — trending on download volume, not likes alone.

The card does something rare: it prints its own loss.Vendor-run (H200, BF16, dated 2026-09-24): 2,000 typed decisions — 73.15% vs a Jev reference’s 72.70%, a statistical shrug at parity. AG News 94% vs 91%. Emotion (DAIR, 6 labels) 86% vs 48%. And Banking77, 72 labels: 64% vs 87% — a −23-point loss, published by the authors, with the mechanism disclosed: 100-example pilots, and Banking77 handled through a ranking top-16 shortlist rather than a native 72-way call. Abstentions count as wrong. The Jev column is labeled on the card as a supplied reference, not a fresh Jev run — say credit where the disclosure itself earns it. MASSIVE scenario classification, 52 locales: 71.50% macro accuracy across 154,648 cases.

Why it matters: the decision lane has been running on decoders-with-heads (CLM freezes a Qwen3-8B), laptop LoRAs (kev-0.5b) and API services (Jev). Julia is the first entrant that is just a small native encoder — phone-and-lap-class inference, a permissive license, and a documented ceiling.

Verdict — Marginal — adopt only if you need a tiny permissive referee at volume. The incumbent for tool-pick/triage is Ollama’s /v1/systemone with nimble or tev1; the incumbent for self-hosted judging is CLM-8B. Julia’s delta is two orders of magnitude less compute and 52-locale breadth; its ceiling is exactly the row it lost — composition over 72 labels. Fine for routing and triage; keep a real model for judgement.

NaiveAI/Naive-N0.5-Flash — a 309B coder with no full attention, sub-path unproven

New model of the window (created Sep 27, MIT): Naive-N0.5-Flash — 309B total / 15.5B active, built for coding and AI R&D on a Xiaomi MiMo-V2.5 base, native 1M context with a 39 sliding-window + 9 DeepSeek-sparse-attention layer layout and no full-attention layers at all. First coverage of a NaiveAI release here; 989 downloads / 122 likes — above the gate, one step off the shelf.

The deployment honesty has to come from me, because the card’s framing is sales-side: the FP8 weights occupy ~315 GB and require FP8-capable NVIDIA GPUs plus trust_remote_code (custom naive_n05_flash architecture). The advertised 2,000 tok/s “Ultrafast” figure comes from NaiveRT, the lab’s own engine (mega-kernel fusion, speculative decode), with no fidelity context against base and no third-party measurement. Local-engine builds exist (GGUF, EXL3, NVFP4) but every one of them is still under 260 downloads. Verdict: watch, don’t set up — the architecture line is genuinely interesting to anyone who cared about the GLM-5.3-Flash and DeepSeek-SWA threads; the running story isn’t.

The lane’s tape, two days in

CLM-v0.1-8B: 1,271 (29 Sep, the day it led this feed) → 2,392 downloads / 581 likes — nearly doubled in two days, 8 quants still up. Jev-Omni → 1,132. Xing4.0-29B-A4B → 47,613 / 1,819. And the applause column, unchanged in shape: laya 4,713 likes / 0 downloads, #2 trending on hearts alone; Ming-Image-0.1-Design 360 / 0. ZDTaichu5.0-9B is the drift to watch: 12,139 downloads but 2,468 likes — since the 29 Sep tape, likes +715 against downloads +401. Bookmarking, not running.

2. Quantizations

GLM-5.3-Flash: a 1.8M-downloads-a-month quant shelf gets a native mainline engine

What it is: GLM-5.3-Flash — the 320B/18B-active hybrid MoE this feed led the Sep 2 roundup with (MIT, 1M-token context, 34 KDA linear-attention layers + 11 DeepSeek-style sparse-attention layers, mHC connections, text + vision). Yesterday morning (30 Sep 06:20Z), llama.cpp merged PR #27773 — native GLM5-Next architecture support: KDA layers reuse the Kimi-K3 implementation, attention is no-position MLA, the mHC helpers were pulled out of the DeepSeek-V4 graph builder into graph_context so both models share them, and a new DSA indexer scores pools of 4 consecutive tokens while always keeping the incomplete tail. Two follow-up fixes landed since (#29745 gives dead indexer slots unique scatter rows; #29712 wires it into the Metal CI matrix).

The quant family, first full count here (created late Aug, 30-day downloads read today): unsloth/GLM-5.3-Flash-GGUF 971,744 / 442 likes — incumbent-shelf volume. orcarouter’s uncensored FP8 200,148, antirez/glm-5.3-flash-gguf 188,765, four NVFP4 builds (RadixArk 74,298, nvidia 57,004, RedHatAI 55,128, LibertAIDAI 42,194),incoai’s DFlash2 drafter build 42,116, the uncensored GGUF riders at 34,034 / 32,361 / 28,772, and pfeifferj’s independent GSQ-RCO reproduction at 24,498 — it nearly doubled from the 12,886 read yesterday. Sum the thirteen and the shelf is pulling ~1.8 million downloads a month for a model the default engine could not load until this merge existed.

Deployment verdict — Real deployment option for the cluster-and-big-Mac class. The incumbent at this weight class is DeepSeek-V4.1-Flash on ds4/DwarfStar (antirez’s pack below). GLM’s delta: 18B active vs 16B is a wash, but KDA hybrid attention + native vision + 1M context + MIT — and a mainline llama.cpp path. Nightly binaries carry it; the merge is a day old, so expect first-run friction and file it upstream rather than assuming rot.

The MOPD fix ships its quant wave — and it’s 10× overnight

What it is: the builds for MiMo-V2.6-Flash-MOPD and Pro-MOPD — the tool-call-repetition fix checkpoints covered 29 Sep (66-key config diff says identical: every serving recipe from 24 Sep carries over, a weights-only swap). The wave exploded between reads: QuantaPlanta’s MXFP4 GGUF was 4,429 downloads yesterday morning — 43,639 today. ggml-org’s official GGUF 1,447 → 2,019. Eleven more builds above the gate: AesSedai 709, MikeRoz’s Pro 4.04bpw EXL3 645, ProCreations NVFP4 637, ddh0 617, drowzeys’ fleet hybrid 449, then 282 / 273 / 268. Bases: Flash-MOPD 3,434, Pro-MOPD 985.

Same window, same architecture, engine side: llama.cpp merged #29650 — “mimo: support dflash (convert + feature extraction)” (30 Sep 14:04). The DFlash drafter that made the 2× DGX Spark recipe fast (structured-call acceptance 6.85/7 vs 0.83/7 for prose, measured) now converts inside mainline instead of a side script.

Verdict — Real deployment option if you were already running Flash-RL builds: swap to the MOPD pack, keep every fit and launch flag. Nobody advertised these quants; twelve of them crossed 200 on their own. That is what a fix that people actually want looks like.

Tape — the two 1M stories and the runners-up

antirez’s DeepSeek-V4.1-Flash GGUF: 1,151,996 (yesterday read) → 1,193,974 / 54 likes. unsloth still has not shipped a V4.1 GGUF — day 23; their newest DeepSeek shelf entry is the Aug-31 Vision-Exp. ukisai Swift-1.5 GSQ-RCO 116,252 → 175,005 (+50% in a day; verdict unchanged — Marginal: vendor speed claim, enterprise licence). Viggle 6-step DMD 190,649 → 205,137 — adoption still voting ahead of the fidelity eval nobody has published. Bonsai-2 3,676,692 (batch caveat), DavidAU NEO-MAX 1,777,756 / 1,308. Qwen-Image-2.1 keeps compounding: Comfy-Org repack 4,699,089 → 5,032,483 in a day, abenzerps uncensored 1,232,685, base checkpoint 70,687.

3. Fine-tunes

Nothing created since 27 Sep clears the 200 gate, and nothing older is new. The fresh repos of the last 48 hours sit at zero downloads — the HF counter batch-lags, so that queue gets judged tomorrow, not pre-guessed. Tape on fine-tunes already judged: Confucius4-R2T2 11,535 / 483, akatz’s H3 Character-Swap 8,078 / 200, OrcaSAQ-2 Cyber GGUF 7,186, Hemmingway-1 8,524 / 793. Verdicts unchanged.

4. Technical term of the week — the reranker, or: why a decision is not an essay

The week’s anchor is a merge line: llama.cpp PR #29627 — “model: support classifier_pooling for ModernBERT rerankers” (merged 29 Sep 20:33Z). Translation, then stakes.

A reranker is a model that takes a (query, candidate) pair and returns one number: how relevant / correct / safe / on-topic is this. It comes from search engineering — retrieve 1,000 candidates cheaply, then re-rank the top handful with a heavier, smarter model that actually reads the query and the candidate together (a “cross-encoder”). A classification head is the same machinery one step further: run the encoder over the input, pool the final hidden states down to a single vector — take the special-token slot, or average, or the last position — then a small matmul turns that vector into a score per label. That pooling choice is classifier_pooling: metadata a model file declares and the loader honors. Before this merge, a GGUF could describe its head all day and the engine would ignore it and hand you tokens. Now the server can serve a score where it used to serve only text.

Why it matters for local AI: decode is the expensive organ. Every generated token is a full forward pass (13B, 27B, whatever), and a “decision” phrased as prose pays that price per word — the 13 Sep reasoning-budget arithmetic still stands: 8k thinking tokens at 40 tok/s is ~200 seconds and a whole invoice for one judgment call. An agent that picks a tool, triages a ticket, or gates a risky command makes that call dozens of times a task. A pooled head does it in one forward pass, no decode loop, a fixed output space (no “sure! here are 7 approaches…” failure mode), and a probability you can calibrate — which is why the 29 Sep Ollama endpoint returns {score: 0.87}, why Pi v0.99 (below) wires classifiers into the loop within a day of the merge, and why Julia-1 fits the decision lane in half a gigabyte.

The trade-offs, honestly: a head can only score the options you give it — no out-of-set reasoning, no “actually the right move is C, which you forgot to list.” Label sets are decided at call time (or trained in, and then you have a specialist). Long or compositional questions hit the cross-encoder ceiling — Julia’s own −23-on-Banking77 is this week’s worked example. Cross-encoders cost per candidate pair: fine at N=16, a real compute line-item at N=100k. And a score is only as good as its calibration — the good actors in this lane publish ECE (expected calibration error), not just accuracy.

What just changed: a reranker-class head is no longer a research artifact or an API vendor’s product. It is a named op in the default local engine, a tool inside the default local harness, and a 144M Apache download. That’s what a term crossing from idea into plumbing looks like.

5. Harness improvements

  • Pi — v0.87 → v0.99.2 in ten days, and the repo moved. The harness is now earendil-works/pi (110,841★). v0.99.0 (29 Sep) is the big one: codemode runs model-written JavaScript in a QuickJS sandbox that calls Pi’s own tools — parallel fan-outs and glue code as one scripted call instead of N round-trips of prose; MCP servers are built-in (stdio + streamable HTTP, OAuth, per-project trust); tool_search stops declaring the whole catalog up front (the fix for MCP’s context-tax); virtual models route each request to a different physical model; and the System One bolt-on: classifier models as first-class — Jev classifiers from codemode scripts, or any llama.cpp model used as a classifier, with a llama.cpp integration doc shipped in the release. That is the classifier_pooling merge consumed as a harness feature within 24 hours. v0.99.1: GPT-6.1 Sol default on the Codex provider. v0.99.2 (30 Sep): default-codemode MCP servers no longer clutter the tool list or block the first prompt — scripts find them via searchTools().
  • llama.cpp — a dense 48 hours. GLM-5.3-Flash native support (#27773) and classifier_pooling (#29627), both above; #29638 — speculative decoding stops accepting draft tokens at end-of-generation (the quiet correctness fix: a drafter could keep a finished answer talking); #29598 faster model loading; #26979 rejects GGUF tensors whose size wraps after padding. A UI layer started landing(five PRs, 30 Sep): HF Hub data layer, model download pipeline, memory-fit estimation — the engine growing the “will it run here?” question into a first-class screen. Also: MiMo DFlash convert (#29650), llama_prefetch_rows (#29599), BF16 unary/GLU/binary ops on CPU+CUDA (#29675, #28937), a HIP fix that explicitly refuses to treat CDNA as a DGX Spark (#29572), and Metal BF16 math for mxfp4 (#29770).
  • mlx-vlm v0.7.4 (28 Sep). Batched num_imagesgeneration for qwen_image, ming_image, flux2 and ideogram4 — N images per prompt in one pass — plus an engine-completeness ledger update: MiMo-V2.6 audio requests no longer fail across server threads. Also the Qwen3.5 batch left-padding fix and packed-width fixes for 3/5/6-bit quantized KV caches.
  • Ollama v0.35.1-rc0 (29 Sep, first RC after v0.35.0’s/v1/systemone). oh-my-pi v18.4.5 → v18.4.8 (30 Sep → today 06:08Z): a segfault-at-startup fix for Apple-silicon Macs on macOS pre-27 (Apple Foundation Models now loads only on macOS 27+), afterToolCall trusted post-tool guidance, forced-tool-call fixes for Claude on Bedrock, Factory Droid OAuth. mlx-serve: no new tag since v26.9.6.
  • Hermes — no tag since v2026.9.24; main is busy with desktop-queued terminal selection and session plumbing. PI-Desktop 6,195★; ds4 22,818★ (docs-quiet since 20 Sep).

The one to try today: the week’s thesis fits in one afternoon. Grab a llama.cpp nightly after the #29627 merge, point it at any classifier-capable GGUF, and time the difference between asking a 27B to say a number and having a pooled head return one. If you run Pi:/mcp and codemode are both in v0.99, and the sandbox writes your tool fan-outs for you. System One used to be a research demo. This week it shipped as factory options.