The small one wins.

Models

black-forest-labs/flux-3-image-klein (3,241 downloads / 290 likes in about three days; repos created 13 October, launch announced 15 October) is Black Forest Labs’ new open-weights image model — and the smallest model to top this feed’s image lane since it started. It is a 5.6B-parameter generator that does text-to-image, image editing and multi-reference generation with up to ten reference images, outputs native 4K, edits pixel-precisely, and accepts bounding boxes that either place an element or confine an edit to a region. The release ships a base checkpoint built for fine-tuning and a step-distilled checkpoint that runs 16 steps instead of 50, plus FP8, INT8 and NVFP4 builds put together with NVIDIA. BFL says the family shares training signal across images, video and audio — Klein is the second open-weights FLUX 3 release, after FLUX 3 Action, a 7B robot-control model from 22 September that this log never covered (its base repo sits at 2,424 downloads).

Evidence. The distilled checkpoint cleared the 200-download gate in three days against license-gated repos (you have to click through terms before the files even show), and the ecosystem is forming at the usual speed: BFL’s own ComfyUI repack at 39 downloads / 279 likes, the fine-tuning base at 134/76, a third sibling repo — flux-3-image-klein-pup — at 13/32 that the blog does not explain, and a day-zero community quant wave (GGUFs, dynamic 8/4-bit builds, an OrbitQuant W4A8, an INT4 ConvRot for ComfyUI) all still at zero downloads because it is hours old. Like ratios near 9% read as genuine pull, not bookmarking.

Vendor numbers, once: BFL’s internal Elo evaluations rank Klein first among open models in text-to-image, multi-reference and editing — above Ideogram V4.5 and BFL’s own FLUX.2 [pro] — with +91 Elo in text-to-image and +70 in editing over FLUX.2 [klein] (9B) at 38% fewer parameters. Those are the lab’s own evals, and nobody independent has scored it yet. The speed table is NVIDIA-measured: with the distilled 16-step model, a 1-megapixel image takes about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070 through ComfyUI; the 50-step base model needs 43.8 seconds on a DGX Spark; a GB300 does it in 0.55 seconds. The generator weights are 11.2 GB in BF16 and 5.6 GB in FP8, and the INT8 build runs on RTX cards with as little as 8 GB of VRAM. LoRA training works from day one in Diffusers and Ostris’ AI Toolkit, with NVIDIA NeMo AutoModel recipes; serving recipes exist for vLLM-Omni, SGLang Diffusion and TensorRT. Stock llama.cpp does not run image DiTs — the GGUF wave rides ComfyUI-class tools, as always.

The quietly important part is the license: the FLUX Kommunity License allows free commercial use for individuals and companies under $5M annual revenue, research and non-commercial use for everyone. That is the most generous commercial line an image model in this feed has shipped since Liquid’s $10M LFM license — Qwen-Image-2.1, the incumbent, gates commercial use behind a research license entirely. Verdict: Real deployment option — for anyone on an 8–12 GB card, or anyone who needs to sell what the model makes. If you already run Qwen-Image-2.1 on bigger hardware and never sell an image, nothing here forces a switch; if you do, the 5.6B challenger just became the default to evaluate.

Quantizations

openbmb/MiniCPM-V-4.7-35B-A3B-gguf (100,853 downloads / 139 likes) is the official llama.cpp shelf of MiniCPM-V 4.7’s first-ever MoE, the 35B with 3B active that crossed this feed’s gate on 14 October at 630 downloads. It has since pulled 10.9× the bf16 checkpoint (9,278) — the widest official-shelf-as-distribution ratio yet in a pattern this log has tracked all month: LFM2.5-24B-A2B’s GGUF now leads its bf16 repo 45,676 to 6,286. The shelf is the full ladder (F16, Q8_0, Q6_K, Q5s, Q4_K_M/S, Q4_0/1, IQ4_XS), a mixed 3-bit build that publishes its per-tensor type map alongside the weights, two vision encoders (f16 and a lighter f16-4x), and — added in today’s repo update — a DSpark drafter GGUF for speculative decoding; the drafter’s bf16 repo sits at 1,140 downloads. The 1B sibling went 263 → 5,274 downloads in two days as the Ollama listing compounds. llama.cpp support merged last week (#29416), and OpenBMB’s Lookahead Expert Scheduling — the prefetch trick behind the “13+ tokens/s at 3 GB on a phone” claim — lives in their own llama.cpp-omni fork, not stock. Verdict: Real deployment option — this shelf is how the phone-class open MoE VLM actually reaches Macs and PCs; the on-device smoothness numbers remain vendor-reported, flagged when covered.

Second: nota-ai’s ternary MLX build of nvidia/Nemotron-3.5-Lightning-30B-A3B (337 downloads / 37 likes, created 12 October, openmdw-1.1) — a 1.58-bit ternary squeeze of NVIDIA’s Mamba-2/Transformer hybrid MoE whose BF16 repo has quietly accumulated 676,533 downloads (NVFP4: 907,200) since August. An August model this feed never covered, surfacing now through its quant: the ternary playbook keeps spreading past Bonsai and the parakeets into NVIDIA’s hybrid line. Verdict: Marginal — early, MLX-only, no fidelity table printed.

Tape (covered builds, counters moved): AtomicChat’s Qwen-Image-2.1-Turbo-Uncensored encoder GGUF 258,320 → 296,253; the official Turbo checkpoint 4,702 → 15,315 in five days; Conway’s Underdog-Saluki 311,969 → 344,177; clef-omni 297 → 2,949; autotrust’s GLM5.3-Flash-E192-DGX-Spark 1,394 → 7,230. Klein’s own day-zero quant wave is at zero downloads everywhere — hours old, worth a re-check, not a verdict.

Fine-tunes

Nothing new cleared the gate. Tape: nerkyor’s Coder390 (Qwen3.8-27B SFT/RLOO) 114,351 → 148,660, about +34,000 in a day; Liquid’s d1-3B decision model → 16,731; vllm-sr’s d3 → 2,747; speridlabs’ iris-3b → 1,257.

Technical term of the day: flow matching

An image generator does not draw a picture in one jump; it starts from pure noise and refines it over many steps. Classic diffusion learns this as denoising: at every noise level, predict and remove a little noise, repeat hundreds of times. Flow matching — the method behind FLUX, Stable Diffusion 3’s lineage and Qwen-Image’s DiT — learns something different: a velocity field, a direction-and-speed attached to every point in image space, that carries a sample of noise along a path to a real image. “Flow” because it defines a flow of points through space; “matching” because training matches the model’s field to the paths connecting noise to data.

The efficiency trick is rectified flow: choose the simplest possible paths — straight lines from noise to image — and train the model to reproduce their constant velocity. Straight paths can be walked in a few large steps; curved ones need many small ones. That single choice is why Klein ships two checkpoints and why their tables look the way they do: the base model runs 50 steps, the distilled one runs 16, and on an RTX 4070 that is the difference between a 40-second-class image and a 9-second one. Step count is wall-clock: every image you generate pays steps × compute per step, so the method that converges in tens of steps instead of hundreds is the method that wins on local hardware.

The trade-offs are the usual speed-for-fidelity exchange, made concrete: distilled few-step models can flatten diversity and rare-mode detail, guidance has fewer steps to correct course in, and a card that says “step-distilled, 16 steps” tells you the speed, never the quality — that is what the 50-step base checkpoint is for. One more local-AI note: flow transformers quantize well, because the weights dominate the bytes and the method is forgiving of small numeric errors — which is exactly why a 5.6B flow model fits an 8 GB card at INT8, and why the interesting question on every new image release is not whether it quantizes but what the text encoder and the step count cost you.

Harness improvements

llama.cpp b11642 → b11681 shipped a version bump to 0.7.0 (ggml 0.27.0) and a batch of model support: Nemotron Super 3.5 VL and Nemotron Bolt 74B (#30520 — the same models Ollama’s release headlines, landing in stock llama.cpp the same day), Muse Spark Open with text, vision and audio (#30616), and mtmd_audio_stream with voxtral-realtime streaming audio (#30629). Around it: conversion support for text-only pplx-decider checkpoints (#30575), Mellum reading its MoE gating parameters from the GGUF (#30578) with MELLUM precision handling (#30584), the server saving drafter state in the slot file (#30574) and fixing an off-by-two in the SWA reuse threshold (#30580), SYCL reordered Q8_0 MoE weights (#29337) and wide stores for q4_K/q5_K dequant (#29696), Vulkan host-memory fixes (#30468), a 4 GB cap on CUDA pinned host buffers (#30449), a Hexagon tensor-repacking rework, and a unicode layer that now rejects overlong UTF-8 (#30485).

Ollama v0.40.5 (14 October) went final with a NVIDIA headline: Nemotron 3.5 Super — a 120B-parameter, 12B-active hybrid Mamba-2/Transformer MoE with a C-RADIO vision encoder, tool calling, a thinking toggle and 1M context — plus Nemotron 3.5 Bolt, a 74B/9B distill tuned to fit a 24 GB GPU with a drafter included for speculative decoding. Decision models moved to the cloud lane (clef:cloud, clef-flash:cloud — run without local pulls), Responses conversations with custom tool calls stopped erroring, and cloud-failure messages finally say what failed. The adoption tension is worth printing: the official HuggingFace repos of Super and Bolt are all sub-gate (Super-VL BF16 142 downloads / 51 likes, NVFP4 81/93, Bolt-24GB 58/87, Bolt GGUF 119/65) — Ollama distribution without HuggingFace adoption yet, the likes-beating-downloads pattern this log keeps flagging.

Pi v1.3.0 (15 October) lets you choose where /share uploads go (sessionShareDestination: auto, GitHub gist, or radius); changing the thinking level mid-session on GPT-6 now keeps the prompt cache; Kimi Code over Chat Completions no longer invalidates its cache when tools are added mid-conversation; switching models drops the previous model’s thinking instead of feeding it to the new one as text; and resumed sessions keep their active tools, with +name/-name deltas applying on top. oh-my-pi v18.11.1 (15 October) added Agent.prioritizeSteering — fresh steering messages jump ahead of stale ones in the delivery queue — fixed a Cursor tool call re-sent by the server persisting two results under one id, and added Tokenizer.countLatestTurn so Anthropic replays stay byte-for-byte.

Hermes v0.22.2 (15 October) opens with unusual honesty: “v0.22.1 was built from the wrong commit. It shipped unreleased gateway work, and the Microsoft Store app could stick on ‘Connecting.’ 0.22.2 is the release 0.22.1 was meant to be.” The fixes that were meant to ship: the gateway starts after a reboot again(a freshly rebooted Linux can hand an early PID to an unrelated process, which the gateway misread as itself already running — it now records its boot and checks the host lock), and long local-model turns survive on laptops(prompt-reading progress always streams, stall timers skip the time the machine was asleep, a recovered local server gets another try instead of tripping the breaker for every later message, and compression no longer times out while a slow server reads a long conversation). Known issue, stated plainly: sessions that ran under a 0.22.1 gateway cannot be deleted in 0.22.2. mlx-serve v26.10.2 went stable on 12 October (storyboards into films up to two minutes, YuE2 songs with editable scores, Stable Audio 3, native Linux) — the stable tag postdates the pre-release this log covered on 11 October, with nothing newer since.

The through-line: the small model keeps winning the week. A 5.6B image generator takes on 12B-class rivals with a license that actually lets people sell the output; a VLM’s official GGUF shelf does ten times the distributing of its own weights; a ternary build reaches NVIDIA’s hybrid MoE line. Underneath, the harness layer spent the week on the boring failure modes — gateways that survive reboots, laptops that survive sleep, steering that arrives in order — which is what the boring failure modes are for.