Search: “quantization”

Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses

Quesma benchmarks Qwen3.8 27B quantizations: 4-bit Q4_K_M matches BF16 on key benchmarks while 1-bit collapses to random chance.

Quesma spent roughly $3,000 on Modal GPUs testing Unsloth GGUF quantizations of Qwen3.8 27B across GPQA Diamond, IFBench, and Terminal-Bench 2.1. The 17 GB Q4_K_M quantization matched the 55 GB BF16 model on Terminal-Bench 2.1 and showed little degradation down to 4-bit, while the 2-bit UD-Q2_K_XL dropped noticeably. At 1-bit, scores on GPQA Diamond fell to random-guess levels, with longer reasoning making results worse, and reasoning effort settings significantly affected outcomes.

Hacker News · AI·Sep 8, 14:49 UTC · 7d agoAI researchHN 56↑ · 30 comments1

Attention Quantization for Tabular Foundation Models

FP8 quantization of attention queries, keys, and values speeds tabular foundation model inference up to 1.7x with no accuracy loss.

The paper develops an FP8 quantization strategy targeting attention calculations (queries, keys, values) in tabular foundation models, arguing attention matters more than weight or KV cache quantization given their differing size and serving patterns versus LLMs. Aligning quantization error between test rows and training rows proves crucial, since misalignment causes drastic accuracy drops. A Triton kernel using explicit FP8 matrix multiplication achieves up to 1.7x speedup over regular 16-bit kernels, with no relevant accuracy loss on TabPFN-v3 and TabICLv2 across TabArena and BeyondArena benchmarks.

arXiv cs.AI / cs.LG / cs.CL·Sep 11, 16:31 UTC · 4d agoAI research1

SQS: Bayesian DNN Compression through Sparse Quantized Sub-distributions

SQS unifies weight pruning and low-bit quantization via Bayesian variational learning, compressing Llama3.2 and Qwen2.5 at higher rates with comparable accuracy.

SQS introduces a unified Bayesian variational framework performing simultaneous pruning and low-bit quantization, using a spike-and-slab prior for sparsity and Gaussian Mixture Models to model quantized weights. The authors derive an efficient approximation for the intractable objective and provide a consistency result for the variational approach. Experiments on ResNet, BERT-base, Llama3.2, and Qwen2.5 show higher compression rates than prior baselines with comparable performance drops.

Hugging Face daily papers·Sep 6, 20:00 UTC · 9d agoAI research

Quantization-Aware Healing: a compressed, 4-bit model that outperforms its full-precision original

Multiverse Computing details quantization-aware healing, producing a 4-bit compressed model that reportedly outperforms its full-precision original.

A Hugging Face blog post by Multiverse Computing's CAI team introduces quantization-aware healing for compressed models. The post claims the resulting 4-bit model outperforms the original full-precision model. No additional details or benchmarks were available in the provided text.

Hugging Face Blog·Aug 25, 11:39 UTC · 22d agoAI research

Google's TurboQuant cuts AI memory use without losing accuracy

Help Net Security·Mar 25, 00:00 UTC · Mar 25, 2026AI research

Forgetting Only What Matters: Layer-Selective Unlearning toward Robust LLMs

Researchers introduce FOM-UL, a layer-selective machine unlearning framework that improves forgetting-utility trade-offs and resists knowledge recovery after quantization.

FOM-UL selects transformer layers for unlearning using a forget-to-retain significance score, concentrating parameter updates on layers highly influential for the forget set while leaving most of the model unchanged. It reduces residual memorization versus GA, NPO, KLD, SURE, ReLearn, and LUNAR-based baselines on TOFU, KnowUnDo, and MUSE-style evaluations while preserving retain-set utility. Under 8-bit and 4-bit post-training quantization and adversarial prompts, it maintains stronger suppression of forgotten content, addressing brittleness of diffuse unlearning updates.

arXiv cs.AI / cs.LG / cs.CL·Sep 9, 16:50 UTC · 6d agoAI research1

LACE: Layer-Wise Compression for Dynamic Frame Rate Codecs

LACE introduces layer-wise compression for dynamic frame rate audio codecs, cutting sequence lengths and speeding TTS inference while preserving quality.

LACE (Layer-Adaptive Codec Encoding) applies an independent compression step at each quantization layer of a neural audio codec, enabling layer-specific segmentation boundaries instead of shared ones. Union alignment and boundary anchor mechanisms keep durations consistent for downstream text-to-speech. On LibriTTS, LACE achieves a better rate-quality tradeoff than prior dynamic frame rate codecs and improves TTS inference efficiency at competitive synthesis quality. Code is released in the ESPnet3 codec recipe.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 17:46 UTC · 18h agoAI research1

It's Not RoPE that Creates Sinks: The Role of Self-Concentration and Value-Non-Mixing in Attention

Study shows attention sinks and massive activations stem from causal-mask self-concentration and value-non-mixing rather than RoPE, informing quantization work.

The paper analyzes why attention sinks and massive activations emerge at initial sequence positions regardless of which token occupies them. Experiments attribute both phenomena to self-concentration of attention induced by the causal mask and the subsequent value-non-mixing in attention outputs. The findings provide empirical evidence on LLM internal dynamics and may inform low-bit quantization strategies, which massive activations currently complicate.

arXiv cs.AI / cs.LG / cs.CL·Sep 8, 17:32 UTC · 7d agoAI research

Lightweight Vision Transformer Compression for On-Device Plant Disease Detection in Resource-Constrained Agricultural Field Conditions

A unified ViT compression pipeline (H-BAC pruning, quantization, distillation) cuts plant-disease models 54.5x to 6.01 MB while keeping 95.13% accuracy.

Researchers combined Hessian-Balanced Adaptive Block Pruning (H-BAC), guided by second-order sensitivity estimation, with quantization and attention-based knowledge distillation to compress Vision Transformers for on-device chilli plant disease detection in India. On a 3-class cross-village, cross-device out-of-distribution dataset, the integrated pipeline reduced model size from 327.42 MB to 6.01 MB (54.5x) at 95.13 +/- 2.32% accuracy, matching the 95.13% FP32 baseline. Ablations also show a directly trained 6.01 MB INT8 student reaches 94.87% accuracy, indicating where pruning and distillation add limited value.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 16:40 UTC · 11d agoAI research

Kraken: LLM-based Speech-to-Speech Translation via Low-bitrate VQ and Dual-path Source Conditioning

Kraken speech-to-speech translation model builds on Qwen3-8B with low-bitrate vector quantization and source-conditioned vocoding.

Kraken augments a pre-trained Qwen3-8B LLM with speech feature inputs and low-bitrate single-layer vector-quantized tokens trained to reconstruct SSL features. A separate token-to-waveform decoder, Autowave-X, is conditioned on source speech to improve non-linguistic transfer. Training used 150k hours of multilingual and multitask speech data. The model reportedly beats SeamlessM4T-Large v2 and Qwen2.5-Omni in translation quality and speaker/prosody transfer.

arXiv cs.AI / cs.LG / cs.CL·Sep 11, 16:41 UTC · 4d agoAI research1

Privacy-Aligned Personalized Federated Learning with Compact Adaptation and Variable-Length Gaussian Communication

Paper releases a private client context once and confines adaptation to coefficients, matching full-model differential privacy with 2.67x less uplink on CIFAR-10.

The paper addresses the dimensionality misalignment between record-level differential privacy and low-dimensional client variation in personalized federated learning by releasing a private client context once and restricting repeated adaptation to a fixed coefficient space. A variable-length quantized Gaussian mechanism lets quantization error itself serve as the required privacy perturbation. On MNIST and CIFAR-10, the design matches or outperforms full-model private adaptation across privacy budgets and client heterogeneity while cutting protected uplink 2.67x at epsilon=16 on CIFAR-10.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 17:48 UTC · 1d agoAI research1

Retrospectively Reverse-Engineering Apple's Neural Engine

A developer reverse-engineers Apple's M1 Neural Engine architecture, mapping compute cores, MAC datapaths, and schedulers to explain the NPU's decline as transformers displaced CNN workloads.

A developer who previously maintained a reverse-engineered Linux driver for Apple's Neural Engine (ANE) published a retrospective deep dive mapping the M1 ANE's full internal architecture: compute, datapath, scheduler, memory, and execution model. The M1 ANE has 16 compute cores with 128 FP16 (or 256 INT8) MAC lanes each, totaling 2048 parallel MAC lanes, using 32-bit Q16.16 fixed-point accumulation with FP16 readout and an accumulator that saturates at 2^15. The author argues the ANE's dataflow was architected around the predictable reuse patterns of 2017-era CNN workloads (dating to the A11 Bionic), which autoregressive transformer decode broke, limiting its usefulness for general ML. With Apple's M5 folding ANE cores into GPU cores to tout LLM performance, the post frames this as the beginning of the end for the standalone NPU.

Hacker News · AI·Sep 12, 07:54 UTC · 4d agoAI researchHN 54↑ · 7 comments2

StepAudio 3 Gen Technical Report

StepAudio 3 Gen unifies TTS, voice design, music, and sound effects via discrete autoregressive modeling over RVQ tokens.

StepAudio 3 Gen is a general-purpose audio generation model covering zero-shot TTS, voice design, vocal generation, sound effects, music, vibe speech, and mixed audio in one framework. It uses discrete autoregressive modeling over residual vector quantization (RVQ) tokens rather than the diffusion Transformer paradigm, with a StepAudio Tokenizer representing audio at 12.5 Hz in a shared 16x2048 residual code space. Key design principles include interference-aware progressive pretraining, an RVQ Adaptor for multi-codebook acoustic representations, and shared discrete autoregressive modeling. The model reports state-of-the-art performance on TTS and voice design while retaining strong generation across speech, vocals, sound effects, and music.

Hugging Face daily papers·Sep 10, 20:00 UTC · 5d agoAI research

JustFit: 200K-Token LLM Serving on a 24 GiB Laptop with Just-in-Time State Management

JustFit MLX runtime serves 200K-token contexts for Qwen3.8-27B on a 24 GiB MacBook via just-in-time state management.

JustFit is an MLX-based inference runtime combining KVExec for compressed KV execution, PhaseSwap for component residency, and StateTrans for state-preserving serving transitions, independent of weight quantization. On a 24 GiB M4 Pro MacBook running Qwen3.8-27B MXFP4, it completed 196,608 input and 16,384 output tokens, raising single-request context from the mlx-vlm baseline's 30,720 positions to 212,992 (6.93x). Performance tests show 19.11 tokens/s on a 32K-input probe with a 16,374 MiB median peak footprint, and the runtime answered 29 of 30 AIME 2026 problems correctly.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 17:15 UTC · 18h agoAI research

Deep Learning-Based Detection of Electrical Faults and Power Quality Disturbances in Aerospace Power Systems

Compact ResNet detects electrical faults in 400 Hz aircraft power systems with 95.87 percent accuracy after deployment on Xilinx Zynq MPSoC.

The study targets multiclass fault and power quality disturbance detection in 400 Hz aerospace networks using a high-fidelity simulation model inspired by the Boeing 787 electrical architecture, covering 21 normal, disturbance, switching, open-circuit, and short-circuit conditions. Two 73,500-sample datasets are built from 1D waveforms and STFT time-frequency representations, augmented with domain randomization and class-specific GANs, and the time-series dataset is released via IEEE DataPort. A compact ResNet with 175,685 parameters achieved 96.94 percent software test accuracy, and 95.87 percent after 8-bit quantization on a Xilinx Zynq UltraScale Plus ZCU102 with 6.90 ms mean accelerator latency per record.

arXiv cs.AI / cs.LG / cs.CL·Sep 9, 17:19 UTC · 6d agoAI research

NCP-ArchPreview Technical Report: Moving towards Latent Space Language Models through Next Concept Prediction

An 8.9B-parameter latent-space language model using next-concept prediction matches OLMo-3-7B pretraining loss with only 51.3% of the training tokens.

NCP-ArchPreview augments next-token prediction with Next Concept Prediction over a product-quantized concept vocabulary built from hidden states, trained jointly end-to-end. The 8.9B model was trained on 5.73T tokens from the Dolma-3 dataset, the largest latent-space language model demonstration to date. It consumes 51.3% of the tokens to reach OLMo-3-7B's final pretraining loss and outperforms it by 2.45 points on the downstream macro-average, including a 5.99-point GSM8K gain. The learned latent space also enables lightweight domain adaptation via a 17M-parameter VQ module and improves speculative drafting accepted length by 4.17%.

Hugging Face daily papers·Sep 8, 20:00 UTC · 7d agoAI research1

Marigold V2: Revisiting Diffusion Transformers for Monocular Depth Estimation

Marigold V2 adapts diffusion transformers for monocular depth estimation, improving AbsRel 16-26% over the previous best on KITTI and ETH3D.

Huawei's Bayer lab revisits the Marigold approach to repurpose image generation and editing models built on the diffusion transformer (DiT) architecture into monocular depth estimators. The recipes target single-step inference from pretrained multi-step flow-matching models, with remedies including alignment to ground-truth semantic features and a two-stage fine-tuning protocol using a Sinkhorn-based loss. The resulting model produces crisper depth maps that generalize out-of-distribution and also achieves state-of-the-art results on surface normals estimation and intrinsic image decomposition.

Hugging Face daily papers·Sep 7, 20:00 UTC · 8d agoAI research

ReMoMask-2: Latent Retrieval-Augmented Masked Motion Generation

ReMoMask-2 rebuilds retrieval in the generator's latent space for text-to-motion generation, achieving lowest FID on KIT-ML and SnapMoGen.

ReMoMask-2 is a retrieval-augmented text-to-motion framework that constructs its retrieval database directly in the generator's pre-quantization latent space and aligns text queries through a distilled lightweight projector, eliminating the representation gap. The framework combines Hierarchical Bidirectional Momentum contrastive learning, Semantic Spatial-Temporal Attention, and Topology Structured Masking to handle hierarchical motion structure. The retriever achieves state-of-the-art accuracy, and ReMoMask-2 attains the lowest FID on KIT-ML and SnapMoGen, with a single mask-transformer stage outperforming the previous two-stage pipeline while delivering the fastest inference.

Hugging Face daily papers·Sep 7, 20:00 UTC · 8d agoAI research

VoT: Vision-of-Thought for Unified Multimodal Representation Alignment

Researchers propose Vision-of-Thought (VoT), a discrete visual-planning token layer between VLMs and diffusion transformers improving text-to-image semantic alignment.

VoT introduces a discrete visual-thinking layer between vision-language models and diffusion transformers, letting the VLM act as a multimodal planner that emits tokens describing objects and layouts before pixel generation. A specialized VoT tokenizer is trained with VLM alignment, feature reconstruction, and vector-quantization losses. Experiments show improved semantic alignment and a structured, interpretable interface for controllable generation.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 17:53 UTC · 8d agoAI research

H Company Releases NeoMME: A Family of 260M and 800M Single-Tower Multimodal Encoders That Drop the Vision Tower and Causal Decoder

H Company released NeoMME, 260M/800M single-tower multimodal encoders matching 3.75B ColQwen2.5 on ViDoRe v3 while being 14.4x smaller, under Apache 2.0.

H Company released NeoMME, a family of 262,937,906- and 793,715,032-parameter bidirectional encoders that process text and raw 32x32 image patches in a single tower, pretrained via masked diffusion and released under Apache 2.0 with day-zero Hugging Face Transformers support. NeoMME-Retriever-260M reaches 0.523 nDCG@10 on ViDoRe v3, matching 3.75B-parameter ColQwen2.5 while being 14.4x smaller; the 800M model scores 0.556. Hierarchical token pooling with int8 and binary quantization shrinks late-interaction indexes from roughly 1.5 MB to 6 kB per page while retaining 95.19% of nDCG@10; text-only BEIR retrieval remains a weak spot.

MarkTechPost·Sep 6, 21:06 UTC · 9d agoAI research

ShallowStream: Index Shallow then Answer Deep for Streaming Video Understanding

ShallowStream builds streaming-video retrieval indexes from shallow MLLM layers, cutting per-frame prefill latency by up to 52.1x.

ShallowStream is a framework for streaming video understanding with multimodal LLMs that uses the model's shallow layers to simultaneously encode frames and maintain an always-on lightweight retrieval index via shallow-layer KV caches, avoiding full-depth prefill for every incoming frame. At query time, shallow-layer attention scores plus a diversity-aware selection strategy retrieve relevant context frames. It reports performance on par with the strongest existing streaming methods while reducing per-frame prefill latency by up to 52.1x and 10-second end-to-end latency by up to 11.9x, with code released on GitHub.

Hugging Face daily papers·Sep 1, 20:00 UTC · 14d agoAI research

Kinara and Arcturus Networks partner to provide AI solutions for smart city apps

Help Net Security·Sep 16, 00:00 UTC · Sep 16, 2022AI research