Machine-generated, human-curated. Download counts were read from the HuggingFace API today (30-day counters, batch-updated aggregates). 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 split scene: a cheetah barista slams a coffee cup onto the counter in a blur while a turtle in a tweed professor jacket sweats over a towering stack of papers, quill poised.
System One vs System Two, café edition: one serves you in a reflex, the other is still composing its order.

1. Models

CLM-v0.1-8B — the decision model that skips generation entirely

What it is: Contrastive-LM/CLM-v0.1-8B (Sep 21) — a “System One model” from the Contrastive-LM crew (authors incl. Christopher Ré and Azalia Mirhoseini). Architecture: a frozen Qwen3-8B encoder plus two small projection heads — a state head and an action head — trained with a bidirectional InfoNCE loss. Pre-trained on ~60M Nemotron Q&A pairs, mid-trained on ~30M synthetic hard negatives, post-trained on ~1M agentic trajectories. Apache 2.0, both heads and base. It does not generate: you give it a state and typed questions (choice / noul / score — a probability, never an explanation) or a list of candidates to rank.

Evidence: 1,271 downloads / 499 likes in 8 days, trending #7 overall, 8 community quantizations, 3 fine-tunes, 3 Spaces already. Small numbers — but trending position, not download count, is how a two-day-old research repo shows up.

The claim to judge: zero-shot parity with Jev — TypeSafe’s decision model — on computer-use, gaming and tool-calling at up to 9× lower latency; with Action Caching (~1k candidate pool, reused action embeddings) they claim 13× faster than Jev. Fine-tuned as a verifier it reports SOTA on DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%) at 4–6× the speed — and the card is honest that those numbers come from fine-tuned heads, not this checkpoint, and all are the team’s own runs.

Why it matters today specifically: Ollama’s new endpoint (Harness, below) exposes exactly these three question types and names TypeSafe’s Jev API as the spec. CLM is the first Apache-licensed challenger aimed straight at that baseline. Running it is two processes: vllm serve Qwen3-8B --runner pooling for embeddings plus a tiny clm-serve for the heads. The encoder is locked to Qwen3-8B embeddings — a feature for tooling, a cage for architecture.

Verdict — Real deployment option, for the judge lane. The incumbent for “score N candidates” is asking a chat model to emit a digit — a System Two task paid at System Two rates. If you run best-of-N selection, ticket triage, or an agent that must pick a tool 200 times a day, the 9× latency delta is the job’s whole cost. CLM-35B is promised for early October; the 8B is the proof of the loop.

nvidia/Nemotron-3-Diarization — plugging the “who spoke when” hole

What it is: a 99.2M-parameter open-weight speaker diarization model (Sep 23), Sortformer-line: resolves speaker permutation by arrival order, keeps an Arrival-Order Speaker Cache + FIFO queue for streaming. One checkpoint covers 80 ms input buffers (0.32 s recommended) through 30.4 s offline, 10 ms frame resolution, up to 8 speakers, unlimited duration via chunking. Ships with its own lightweight C++ runtime — nemo-speech diarize meeting.wav — and a streaming-ASR demo Space. Explicitly cleared for commercial use.

Evidence: 26,428 downloads / 476 likes in 6 days, 22 community quantizations, 8 fine-tunes, 5 Spaces. That is a deployment shape, not a launch spike.

The incumbent it faces: pyannote’s community-1 (5.4M downloads, but an April-2025 lineage and a heavier Python stack) or WhisperX’s pipeline bolt-on. Last week this feed covered the transcription half of the local audio stack (Audio8’s rolling-KV streaming, Confucius4’s stable-prefix commits, parakeet-redux at 178 MB); diarization was the named missing piece. NVIDIA released a 99M model with a native runtime into that hole.

Verdict — Real deployment option. Small enough to be free, streaming-ready, commercially usable, and the 8-speaker ceiling is the honest limit.

XingChen-AGI/TeleOCR — the document parser that reads bent paper

What it is: a ~1.2B VLM for document parsing, rebranded NaviDC-OCR (Sep 10) from China Telecom’s XingChen line — the Xing4.0 family this feed has watched since 19 Sep. Its thesis is one pipeline for both digital PDFs and camera photos of bent, warped pages — parsing distorted documents directly, no dewarping model upstream of it. Apache 2.0, weights out since 17 Aug, a llama.cpp-compatible GGUF conversion from day two.

Evidence: 27,904 downloads / 826 likes and trending #6 today — an August model that is accelerating four months in, pushed by the EMNLP 2026 Dr. DocBench Challenge. There the third-party eval (eval.ai, not the lab’s own table) puts it at 67.96 overall vs MinerU 2.5 Pro’s 62.26 and PaddleOCR-VL 1.6’s 55.11. Its own OmniDocBench v1.6 table claims #1 at 96.87 among 1B-class parsers — vendor table, noted once.

Verdict — Real deployment option for the “scan a stack of receipts with your phone” lane. 1.2B runs anywhere a 3B OCR runs; the warped-page robustness is the differentiator, and an independent competition board backing it is rarer than a benchmark table.

akhilaaa3/Jev-Omni — the community piles into the lane

577 downloads / 301 likes: a multimodal decision classifier (text, image, audio, video) fine-tuned 30k questions on Gemma 4 12B IT, answering with probabilities only. Its DecisionBench and JevBench numbers are its author’s own datasets and runs, the cross-lab comparison rows are “officially reported by their developers” — a thin, honest, self-scored repo. Verdict — Marginal — a signal that the lane is getting street traffic, not yet a workhorse. One-liner: AlexWortega/openjev sits at 621 likes / 0 downloads — the pre-batch counter plus pure bookmarking, this week’s most crowded category has its usual applause-without-users tail.

2. Quantizations

ISTA-DASLab’s GSQ-RCO Coder — half the experts deleted, 98.7% of the code kept

What it is: Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF (Sep 26) from IST-DASLab — the lab whose GSQ+RCO quantization this feed covered from its first day (4 Sep) and watched cross 1M in three weeks. New move: expert pruning, not just quantization. Flash-Next is a 512-experts-per-layer MoE; the Coder release removes half the routed experts outright — keeping those supporting code, agentic tool use, vision and spatial reasoning — and stores the survivors at the usual 3.5 bpw. Result: 176.9B parameters in a 58.4 GB package of which only 29.6 GB must stay resident (the n-gram shard is a disk lookup), an effective 1.89 bits per parameter of the original transformer (amortising the deleted experts — the card spells out that no single weight is stored at 1.89 bits).

The honest bill, printed by the card itself: at xhigh effort, LiveCodeBench v6 retains 98.7% of the BF16 base (86.28 vs 87.43); SWE-bench Verified retains 91.3% (75.60 vs 82.80). Their reading: single-problem codegen survives capacity cuts; sustained multi-step repo work does not. Capabilities outside the retained set degrade — “an accepted cost of the method.” This is the release’s own framing, which is why it gets a headline instead of a caveat stack.

Evidence: 5,522 downloads / 112 likes in 3 days, and the sibling unpruned Flash-Next GSQ-RCO repo growing 53k → 503k in eight.

Verdict — Real deployment option for the 32 GB-VRAM coder. Incumbent for that class: a quantized dense 27B, or the unpruned Flash-Next GSQ-RCO build needing 48 GB + SSD. This is the first Flash-Next-class MoE that fits a single 32 GB card resident. Adopt it for code and tools; keep a generalist for everything else.

antirez’s DeepSeek-V4.1-Flash GGUF crosses one million

The community Q2/Q4 GGUF of DeepSeek-V4.1-Flash — 1,086 downloads on its first day, 135,900 on day 3 — crossed 1,099,587 downloads (30-day) / 52 likes today, up from 990,509 on 27 Sep, +109k in two days. That is 1M in a month, roughly, at 22× the age-adjusted pace of the biggest lab-shipped GGUFs this feed has taped, for an engine author’s personal quant pack that only runs on DwarfStar. Likes-as-signal stays broken for this repo the way it has been all month — 52 likes against 1.1M downloads means people quietly running it, not bookmarking it. Verdict — Real deployment option, unchanged, now with the milestone to prove it.

Watchlist trigger fires: nvidia’s official V4.1-Flash NVFP4

On 18 Sep this feed flagged nvidia/DeepSeek-V4.1-Flash-NVFP4 at 132 downloads — sub-gate, “re-check.” It re-checks at 20,993 downloads / 93 likes. Official FP4-line weights for the Blackwell class, for the flagship community engine that antirez proved people want. Verdict — Real deployment option (SM100+); the re-quant wave around it (r0b0tlab-style calibrations from the MiMo week) follows demand.

ukisai Swift 1.5 GSQ-RCO — the method as a product shelf

Swift-1.5-Qwen3.8-27B-GSQ-RCO-GGUF (Sep 24, 58,021 dl / 140 likes in 5 days): a commercial house applying ISTA’s GSQ-RCO per-tensor allocations on top of its own Swift 1.5 fine-tune of Qwen3.8-27B — four IQ2_XS→IQ3_S tiers (8.42–11.77 GB) plus MTP variants, full KLD tables per tier, SHA256 manifest. Swift 1.5 claims 58.5% fewer thinking tokens at +0.35% score for a claimed 9.18× end-to-end speed-up (vendor’s own eval, once). The Swift line is itself climbing fast — the stock GGUF ran 9,709 → 144,372 since first logged. One flag the card raises itself: an “enterprise licensing” tier — read the license before shipping on it. Verdict — Marginal — adopt only if the low-thinking-token behaviour is what you want; the quant craft itself is the same method, applied competently.

One-liners with numbers: orcarouter/OrcaSAQ-2-27B — proprietary “sensitivity-aware” 3-bit of Qwen3.8-27B, 54→12.3 GB, +0.02% PPL / 93.2% top-1 agreement (vendor, own eval path), 1,663 dl / 199 likes, vLLM-only, 262K ctx — Marginal, a product pitch with good instrumentation; its Cyber-GGUF rider did 4,599 in 3 days. And the quant shelf this feed keeps re-reading: unsloth still has no DeepSeek-V4.1-Flash GGUF — day 21 (newest shelf entry 9 Sep).

3. Fine-tunes

MiniMax H3 Character-Swap LoRA — the video editors’ casting call

akatz-ai/MiniMax-H3-Character-Swap-LoRA (Sep 25, 3,436 dl / 138 likes, 7 Spaces using it): a rank-16, 1,000-update adapter for H3’s Ref2VA base — give it <Video 1> and a character sheet as <Picture 1>, name the target person in the prompt, and it swaps that character while keeping camera, background, lighting and everyone else. Trained on a single 32 GB Blackwell, published with its own open dataset — and an unusually candid card: 4–5 s continuous shots work better than 14 s, motion timing and hard cuts “remain unreliable,” no duration bound established, no trigger word, no Turbo needed.

This is the third serious H3 adapter family logged (Acc/turbo LoRAs 2 Sep, fused INT8 6 Sep, ComfyUI sharpeners 13 Sep) and the first that edits a source video rather than generating a new one. Verdict — Marginal — adopt only if you specifically want identity-preserving replacement inside existing clips; the card’s own reliability notes are the spec.

Viggle’s 6-step turbo LoRA — the adoption spike keeps climbing

The one this feed flagged 24 Sep at 12,425 dl with “no fidelity eval published” now sits at 175,907 dl / 394 likes — 14× in 5 days on a 6-step DMD distillation LoRA for Qwen-Image-2.1. Nobody has published the before/after the repo doesn’t ship. Adoption says the speed is wanted; the quality question is still delegated to luck. Verdict — Marginal, unchanged, but 176k people voted before the receipts arrived.

Sub-headline fine-tune tape: GLiNER2.5-Decide (the one-forward-pass label router from 27 Sep) 14,772 → 24,250; ZDTaichu5.0-9B 5,078 → 11,738 dl with likes rocketing 272 → 1,753 (the fastest likes-per-download drift this month after laya — watch the ratio, not just the row); Hemmingway-1 7,478 / 765; Confucius4-R2T2 1,864 → 9,336 / 448 — the streaming-ASR lane is filling up.

4. Technical term of the day — expert pruning

The idea: a Mixture-of-Experts layer is many tiny feed-forward blocks plus a router that wakes a few per token. Qwen3.8-Flash-Next runs 10 of 512 experts per layer, 48 layers deep — which means most experts go unused by most tokens, and some go unused by nearly all tokens. Expert pruning deletes those blocks outright. Quantization compresses what the model computes; pruning deletes what the model barely computes.

Why it works at all: a router trained through a soft selection is, in effect, a voting record. Score expert usage across a calibration mix — a few hundred million tokens of the domains you care about — and the bottom fraction turns out to be load-bearing for almost nothing. Removing them saves parameters and memory-bandwidth-per-token, because a smaller candidate pool also means less routing. Delete a top-1 expert and you break a skill; delete a never-selected one and you ship a smaller model that still clears its own benchmark suite.

Why it matters for local AI: MoE is how the big models got cheap to run — but every expert still has to sit on tap (in RAM, on the SSD stream, in the MLX pack). Pruning attacks the only MoE cost that routing doesn’t: the weight you carry to never wake. ISTA’s Coder release is the demo — halving the expert pool is most of how a 176.9B MoE’s 29.6 GB resident set got there before quantization finishes the job. The two methods compound multiplicatively, and the effective-bits number (1.89) is the honest way to report the pair.

The trade-off, in one sentence: you don’t pay for pruning on the benchmark, you pay for it on the task nobody calibrated for — which is exactly why the best releases (this one) print the retained-percentage table and say out loud that sustained agentic work loses 8.7 points while single-shot codegen loses 1.2. A quant of the wrong capability set is a smaller model of the wrong model.

5. Harness improvements

  • Ollama v0.35.0 (28 Sep) — decision models, first-class. New /v1/systemone endpoint “based on TypeSafe’s Jev API”: send a state plus typed questions (choice → choice + per-option probabilities, noul → probability a condition holds, score → rating across ordered criteria), get one-token answers with confidence — usage reports output_tokens: 1 and means it. First library models: nimble (Bespoke Labs — the 9B token-scorer this feed covered 21 Sep, now productized) and tev1 (Together AI). The same week a year ago this idea was a paper trick with a private API; it is now a URL anyone can hit.
  • llama.cpp b11243 — Intel GPUs get a 3.1× day. CI moved to oneAPI 2026.1 after oneDNN vanished from Deep Learning Essentials; re-measured on an Arc B570 with the same code: prompt processing 1,331 vs 434 t/s, generation 50.1 vs 45.3–48.0 (measured notes in the PR — the rare harness win nobody tweeted). Also in the b11194–b11246 window: Muse-Glimmer now honours response_format json_schema under --jinja (#29615), SYCL build docs redone, mtmd build fixes.
  • mlx v0.32.3 (29 Sep). Model-load concurrency now adapts to the machine instead of a fixed cap, the mx.clear_streams() GIL deadlock is gone, and a gather_qmm (quantized MoE) row overflow above 32K contexts is fixed — a quiet one for long-context MoE serving on Apple Silicon.
  • NeMo-Speech.cpp v0.1.0. NVIDIA’s lightweight C++ runtime for NeMo speech models, and the delivery vehicle for the diarization release: nemo-speech diarize meeting.wav, transcribe --diarize --json for word-level speaker tags. The local ASR stack is acquiring llama.cpp-shaped ergonomics.
  • PI-Desktop v0.15.8–v0.15.10 (25–28 Sep). The Pi cockpit keeps shipping fast — now at 6,068★, up from 5,060 on 22 Sep. Hermes has been quiet since v2026.9.24 (24 Sep), Ollama shipped two tags this weekend, mlx-serve no new tag since v26.9.5-final.

Board — the weekly tape

All 30-day counters, HF API, 29 Sep. V4.1-Flash base 668,537 / 3,873 likes (was 640,577); antirez GGUF 1,099,587 — 1M club; Bonsai-2 GGUF 3,457,124 (batch caveat stands); nvidia V4.1 NVFP4 20,993 (trigger fired). MiniCPM5-2B 814,813 (+88k in two days); Xing4.0-29B-A4B 45,834 / 1,802. Qwen-Image-2.1 compounding: Comfy-Org repack 4,351,753, abenzerps uncensored 1,062,921, unsloth stock 220,231, pottokao’s ablated text encoder 158,806 — base checkpoint 58,693, the 60× distribution-channel gap intact. DavidAU NEO-MAX-GGUF 1,683,597; the other uncensored 27B shelf member, HauhauCS Aggressive MTP, tapes in at 2,057,914 — first mention here, worth the look at the… well, it’s a 2M-download abliterated Qwen. MiMo: Pro-RL 76,518 (growth cooling), but bartowski’s 9B distill GGUF 57,826 → 124,344 in two days — the runnable one is the one that moves. Audio8-ASR-Infinite 19,963 / 1,452. And the applause column, for the record: laya 4,367 likes / 0 downloads.

The one to try today: the ISTA Coder quant if you have a 32 GB card and write code with it; otherwise the 5-minute experiment with a genuinely new vibe — ollama pull nimble, then ask it a choice question at /v1/systemone. A model that answers with one token and a probability is a different animal. It’s System One, and it just got an endpoint.