A focused feed of models, agents, research and open-source releases for people building with AI.
Community quant repack of Qwen-Image-2.1-Turbo hits 2.5M downloads and 4.5k likes on Hugging Face. Eleven formats: GGUF (Q4 to Q8), FP8, INT8 ConvRot, NVFP4 safetensors, and MLX 4/6/8-bit for Apple Silicon. Base model generates in 8 denoising steps at CFG 1, versus 40 steps for Qwen-Image-2.1.
Partially decensored GGUF of Swift 1.5 Qwen3.8-27B built with Heretic , with vision and MTP intact. Scores 97.6% GSM8K, 86.67% MMLU-Pro, 92.98% IFEval prompt-strict on the quantized file. Follows through on roughly 1 in 5 prompts the original Swift 1.5 refused; not fully uncensored.
Topaz Labs released Wonder 4 , a one-click upscale, denoise, and sharpen model. First Topaz model built on a mixture-of-experts architecture that routes enhancements per image region. Combines upscaling, denoising, deblurring, compression artifact removal, and face recovery in a single pass.
Google's Project Suncatcher ran Gemma 4 on a data-center TPU in orbit, returning a haiku to Earth. Prototype satellite named MVP, built with Planet, launched Oct 1 on SpaceX Transporter-18. Carries four Trillium-generation TPUs running on roughly a 1 kW solar power budget.
Community release puts MiniMax H3 on consumer GPUs via pruned GGUF quantizations. Checkpoints range from 8.9 GB (Q3_K_M, ~12 GB VRAM) up to 21.6 GB (Q8_0 reference). Two modes shipped: FL2VA first/last-frame-to-video-audio and Ref2VA omni-reference generation.
HoppouAI released Breeze-TTS-2.cpp , GGUF quants plus a C++ runtime for BreezeBlue's 3B TTS model. Runs on NVIDIA, AMD, Intel GPUs via Vulkan, with CPU fallback and no Python required. Hits roughly 1.2x realtime at Q8_0 on an RTX 3060, 24 kHz bilingual English and Mandarin output.
Tiny Corp cut the exabox price from $10M to $5M, with 2 preorder slots left at $100k deposit. GPU config switched from 720 unreleased RDNA5 cards to 256 shipping AMD MI300X accelerators. Specs: 669 PFLOPS compute, 49 TB HBM3, 1.36 PB/s memory bandwidth, 500 kW at 480V.
ISTA-DASLab released a 1.05 bpw compression of DeepSeek-V4.1-Flash-1017 , a 763B parameter MoE model. 72.1 GB of transformer weights fit on a single GPU, down from 138 GB for IQ2_XXS. Retains 94.8% of SWE-bench Verified and 98.7% of LiveCodeBench v6 against base.
AutoTrust ships a lossless MTP speculative-decoding recipe for Qwen3.8-27B-NVFP4 , taking DGX Spark from 12.5 to 48.2 tok/s. Weights are bit-exact identical to NVIDIA's upstream NVFP4 release; only the drafter changes.
NVIDIA released an NVFP4 quantized DeepSeek-V4.1-Flash-1017 , roughly 400 GiB versus 475 GiB source. Accuracy on GPQA Diamond, AA-LCR, SciCode, IFBench, MMMU-Pro, and Terminal-Bench 2.1 matches source within noise.
Sharp-Spark-X2.5-4B-GGUF is an imatrix-quantized repack of Spark-X2.5-4B targeting 6 GB VRAM GPUs. Q6 ties upstream Q8 on MMLU-Pro while using 43% fewer tokens at the median. Ties Anthropic Haiku 4.5 on an SWE-bench-Live slice at 4B parameters.
Liquid AI released LFM3 , six open-weight 2B-class models sharing one backbone across text, vision, audio, omni, and decision tasks. New tri-hybrid architecture mixes 13 gated convolutions, 5 sliding-window attention, and 6 global attention layers across 24 layers.
EdgeCIM is a CIM-based accelerator targeting the memory-bound decoding phase of decoder-only SLMs up to 4B parameters. On LLaMA3.2-1B it hits 400 tokens/s, 7.3x faster and 49.59x more energy-efficient than Jetson Orin Nano.
Prime Intellect released Prime Mega MoE , adding a backward kernel to DeepSeek's forward-only Mega MoE. Up to 3.11x faster MoE kernels versus a PyTorch baseline, 1.34x faster end-to-end training. Qwen3-30B-A3B steps 48% faster and Qwen3-235B-A22B 55% faster than TorchTitan.
FrostWolfAI released Spark-X2.6-4B-abliterated-GGUF , a 4B uncensored local model in 14 quantization formats. Refusals drop from 100/100 to 2/100 at just 0.0177 KL divergence from the base model. Retains 1M-token context, controllable chain-of-thought reasoning, and native tool calling intact.
A research team proposes using thin-film lithium niobate modulators as analog replacements for Softmax and Sigmoid inside transformer attention.
The new release flips the rewritten NCCL backend on by default, adds one-sided transports, and accelerates NVFP4 decode on Blackwell.
A community port by the handy-computer team converts NVIDIA's multilingual Parakeet-TDT-0.6B-v3 into GGUF files for transcribe.cpp, enabling local CPU and GPU inference.
Qdrant's ONNX port of multilingual-e5-large hits 926K downloads, giving teams a lightweight way to run a top multilingual embedder in production.
QuantStack's GGUF conversion of Wan2.2-I2V-A14B shrinks a 27B Mixture-of-Experts video model into quantizations that fit on consumer GPUs through ComfyUI.
A six-page follow-up uses Hypothesis to stress-test the recent DeepMind softmax fix, finds a new edge case, and ships a tougher implementation.
Qwen's compact vision-language model gets a 4-bit AWQ build, cutting memory to run multimodal reasoning, OCR, and agent tasks on modest GPUs.
A 4-bit Gemma 4 MoE build turns a telecom LLM into a probability-driven decision engine, running at 32k tokens per second on a single AMD MI350P.
Slotstream streams a 125B mixture-of-experts model from SSD on Apple Silicon, letting 16 to 64 GB Macs run a 105 GB model locally.
No stories match these filters.
