Forty-nine pages a second.

Models

datalab-to/indra-ocr (442 downloads / 140 likes in a week; repo created 10 October, GitHub at 489 stars and version v0.3.1 already out) is the new open OCR engine from Datalab — the lab behind the Marker and Surya document stack most local PDF pipelines already sit on. The shape is the story: the main recognizer is a 1M-parameter CTC model, the whole shelf is about 18 MB of ONNX Runtime models with no PyTorch, and it reads PDFs and images into markdown, plain text, or JSON with a bounding box for every line and labelled layout blocks (headings, lists, tables, captions, footnotes) in reading order. It is text-layer-aware: on digital PDFs it reads the file’s own text in any script and OCRs only what is missing. Install is pip install indra-ocr; a local web viewer ships alongside, and the files never leave the machine. Code is Apache-2.0, weights OpenRAIL-M.

Evidence. The gate is cleared modestly but honestly — 442 downloads against 140 likes in seven days, plus 489 GitHub stars in four — and the author is proven: this is the same org whose tooling already processes a large share of the open document pipeline. It is a tiny-model release in the vein this feed keeps finding real: needle3 (121M), parakeet-redux (178 MB), Phonon-2 (164 MB).

Vendor numbers, once: the accuracy table comes from Datalab’s own OmniParseBench (16,288 pass/fail tests on 2,937 pages — open-sourced, same pages and scorer for every engine, but still the vendor’s bench). Headline 62.1 vs PaddleOCR’s full PP-StructureV3 pipeline at 58.4 and Tesseract at 19.3 on text; with the optional --math mode it reports 66.8, ahead of PP-StructureV3 on text, tables and layout. Speed: a median ~228 ms per page on 4 vCPUs, and ~44.9 pages/second on 32 cores in their throughput run — the “49 pages/s” headline is the same table’s best case. PP-StructureV3 needed a median 40,958 ms per page on the same machine class. The card’s limitations section is unusually honest and worth reading before you adopt: no handwriting (~8% on its own tests), degraded scans (~20%), complex merged-cell tables, and — stated outright — the orientation model turns about half of upright Arabic and Hebrew OCR pages sideways. Scripts outside its OCR list (Latin ~50 languages, Cyrillic, Greek, Devanagari, Bengali, Chinese) work only through a digital PDF’s text layer.

Verdict: Real deployment option for the CPU batch lane. The realistic incumbents are Tesseract and PaddleOCR, and the delta is a class apart: every recent parser this feed covered — LightOnOCR-3 (4B), Youtu-Parsing-Omni (5B), TeleOCR (1.2B) — is a GPU model; Indra is 18 MB of ONNX that saturates a CPU and posts a better headline score than the full Paddle pipeline (vendor-run, flagged once). It is not your handwriting or warped-scan model — the card says so itself, and points you at the paid Chandra tier for that, which is exactly the sort of upsell honesty you want printed in the limitations.

Sub-gate line: Tencent’s Hy-Memory-30B-A3B (created 15 October) sits at 0 downloads / 131 likes — the applause-without-users pattern again. Skip until it moves.

Quantizations

The day’s biggest number belongs to a repack. black-forest-labs/flux-3-image-klein-comfy — BFL’s own ComfyUI-format shelf of yesterday’s FLUX 3 Klein release — went from 39 downloads to 73,767 in one day, at 491 likes. That is 3.2× the generator repo itself (22,790) and roughly 1,900× its yesterday counter. The pattern is the month’s most reliable one: Comfy-Org did this for Qwen-Image-2.1 (3.6M+), OpenBMB’s official GGUF shelf does it for MiniCPM-V 4.7 (10.9×), and now the image-model lab ships its own Comfy-native files on day one because that is where the users are. The unexplained -pup sibling moved 13 → 426 downloads — still unexplained, still liked. Verdict: this is how Klein actually ships — yesterday’s “Real deployment option” stands, license permitting, and if you use ComfyUI the -comfy repo is the download you want.

Second: nvidia/Qwen3.8-Flash-Next-FP4-Local(771 downloads / 156 likes, created 8 October, first coverage) — NVIDIA’s FP4 build of Qwen’s 125B-class hybrid-attention MoE, quantized with Model Optimizer using quantization-aware distillation, ~96 GiB on disk, and validated on exactly the hardware local-AI readers own: a single RTX PRO 6000 Blackwell (850K tokens of FP8 KV with the n-gram embedding table offloaded) and a single DGX Spark(1M tokens of FP8 KV with PLE offload). The catch is printed in the card’s first paragraph: the NVFP4 scale factors use a new lossless compression format whose upstream vLLM support has not landed — you must run NVIDIA’s own containers. Verdict: Marginal — adopt only if you run Blackwell or a Spark and accept the container lock. The open path to Flash-Next stays ISTA-DASLab’s GSQ-RCO GGUF shelf (4.19M downloads), which runs in stock engines.

Tape (covered builds, counters moved): MiniCPM-V-4.7’s official GGUF shelf 100,853 → 125,004(+24k in a day); LFM2.5-24B-A2B-GGUF 45,676 → 60,927; AtomicChat’s Turbo-Uncensored encoder 296,253 → 327,786; Conway’s Underdog-Saluki 344,177 → 370,955.

Fine-tunes

Nothing new cleared the gate. Tape: nerkyor’s Coder390 (Qwen3.8-27B SFT/RLOO) 148,660 → 182,130, about +33,000 a day, with its GGUF-NInfer sibling at 143,580 — the family is past 325k this month; vllm-sr’s d3 → 3,565; speridlabs’ iris-3b → 1,320.

Technical term of the day: the importance matrix

Quantization replaces a model’s weights with coarser values, and the naive way — round everything equally — spends its error budget where it hurts. The importance matrix (imatrix) is the fix the GGUF world settled on: before quantizing, run a corpus of representative text through the model and record how strongly each weight’s output channel actually activates. Those statistics become per-column weights in the quantizer’s error function, so the 4-bit grid is fitted where the activations are large, and relaxed where a tensor barely wakes up. The Q4_K_M file you run by default is an imatrix artifact; at 3–4 bits the difference between a quantized model that reads like itself and one that produces confident nonsense is routinely the quality of that matrix.

Three trade-offs matter if you quantize yourself or pick between shelves. First, the matrix is only as good as its corpus: an imatrix collected on code produces a quant that is kindest to code and can quietly shortchange prose, so shelf-builders publish what they calibrated on — ask. Second, the cost lands on the quantizer, not you: bartowski, mradermacher and unsloth run their corpora through every model they package, which is a real reason the community shelf exists alongside the official one (antirez’s DeepSeek-V4.1 GGUFs, this feed’s most downloaded build, ship a shared 8,192-token imatrix). Third, it fixes error, not structure: the bartowski Clef decoy that scored 20,514 downloads scored them with the decision head missing, and Bonsai’s silent-gibberish packing was missing a runtime rotation — no matrix saves a wrong file. It is the cheap post-training version of a spectrum that ends with quantization-aware training: NVIDIA’s FP4-Local build above pays the compute to distill the quantization into the weights themselves, which is what the “QAD” in its card means. Whatever the method, the receipt is the same one this feed measured on 12 October: perplexity, KL divergence and same-top-p against the bf16 parent, printed next to the file.

Harness improvements

llama.cpp b11682 → b11706 fixed a quiet test blind spot worth naming: the Metal fusion baseline was only ever a lookup table for rows the run produced, so qwen4exp and Mellum had silently lost their fused kernels while CI stayed green — b11684 makes the test fail when a baseline row stops matching. Model support grew: Mellum2.1 gained MTP speculative decoding(#30612, the JetBrains coder now drafts its own tokens), IBM’s Granite Switch 4.1-8B landed as a decision model that can also chat (#30670/#30676 — the decision lane keeps absorbing architectures), and OpenCL got an MXFP4 MoE kernel for Adreno X1 Snapdragon silicon (#30691). Plumbing: -hf downloads finally honor HTTP(S)_PROXY (#21752 — corporate networks stopped failing silently), a SYCL prompt-processing regression after the Q4_0/Q8_0 weight reorder was fixed (#30374), and Hexagon vectorized softplus with HVX.

Ollama v0.40.6 (16 October) quietly closed a loop from this feed’s longest-running saga: its GGUF layer gained q2_0 tensor and file types (#18967) — the mainline-compatible ternary packing that stock engines refused during the Bonsai 2 affair, now first-class in Ollama’s converter and server. Clef arrived in Ollama’s cloud lane, and Windows runners stopped being orphaned.

oh-my-pi v18.11.2 / v18.11.3 (16 October) fixed a bug every structured-output user should care about: JSON-schema dereferencing was rewriting literal values inside const, enum and default fields, silently corrupting schemas on the way to the model. It also fixed Anthropic requests hanging forever on Bun 1.3 when the upstream accepted the request but never sent headers — they now abort and retry — and shipped a breaking auth-gateway API change in v18.11.2. Pi stays at v1.3.0 with one main-line change worth a line: NVIDIA devices now default to Nemotron 3.5 Super VL. Hermes is quiet on v0.22.2, mlx-serve on v26.10.2 stable.

The through-line: the small stuff ships. An 18 MB OCR engine clears the gate on CPU throughput alone; the file that distributes a model now routinely beats the model; a Spark-validated FP4 waits on upstream vLLM; and the humble calibration matrix keeps deciding whether your 4-bit files read like the originals. Nothing new cleared the model gate today — and the day still had this much in it.