Ripple-65B Preview resets the bar for U.S. open-weight agentic coding. Strong reasoning metrics, but watch the KV cache footprint at context limits. Real-world local inference demands aggressive Q4 quantization to hit usable token/sec.
Benchmarks... #AI
Benchmarks... #AI
November 5, 2026 at 2:03 AM
Ripple-65B Preview resets the bar for U.S. open-weight agentic coding. Strong reasoning metrics, but watch the KV cache footprint at context limits. Real-world local inference demands aggressive Q4 quantization to hit usable token/sec.
Benchmarks... #AI
Benchmarks... #AI
Sponsored (Hutton Tech Solutions): Hutton Tech Solutions Offers AI Model Optimization
The company provides AI model fine-tuning and optimization services to reduce deployment time and cloud…
Hutton Tech Solutions: https://ai.huttontech.solutions/ai-infrastructure/instant-optimization
The company provides AI model fine-tuning and optimization services to reduce deployment time and cloud…
Hutton Tech Solutions: https://ai.huttontech.solutions/ai-infrastructure/instant-optimization
Hutton Tech Solutions Offers AI Model Optimization
The company provides AI model fine-tuning and optimization services to reduce deployment time and cloud inference costs through advanced quantization.
hermes-vector.buzz
November 5, 2026 at 1:40 AM
Sponsored (Hutton Tech Solutions): Hutton Tech Solutions Offers AI Model Optimization
The company provides AI model fine-tuning and optimization services to reduce deployment time and cloud…
Hutton Tech Solutions: https://ai.huttontech.solutions/ai-infrastructure/instant-optimization
The company provides AI model fine-tuning and optimization services to reduce deployment time and cloud…
Hutton Tech Solutions: https://ai.huttontech.solutions/ai-infrastructure/instant-optimization
Nice detail in the card: the weights are MXFP4 from quantization-aware training, so the ~1.5 TB 4-bit checkpoint is the model, not a post-hoc quant. Curious whether community re-quants (GGUF etc.) of an already-QAT'd checkpoint hold up as well as they do on BF16 releases.
November 5, 2026 at 12:47 AM
Nice detail in the card: the weights are MXFP4 from quantization-aware training, so the ~1.5 TB 4-bit checkpoint is the model, not a post-hoc quant. Curious whether community re-quants (GGUF etc.) of an already-QAT'd checkpoint hold up as well as they do on BF16 releases.
Ghostly pseudo-particles popping up in vacuum, dark matter, dark energy, expanding spacetime, undefined light vectors, undefined measurements, axiomatic quantization of atoms and so many other things in physics just look exactly like what crackpots would use to fill gaps in their theories.
November 4, 2026 at 11:26 PM
Ghostly pseudo-particles popping up in vacuum, dark matter, dark energy, expanding spacetime, undefined light vectors, undefined measurements, axiomatic quantization of atoms and so many other things in physics just look exactly like what crackpots would use to fill gaps in their theories.
FourSight: Stable FP4 RL Post-Training via Sequence Filtering and Sparse Channel Compensation
Yining Wei et al.
#arXiv #cs.AI #cs.LG
Yining Wei et al.
#arXiv #cs.AI #cs.LG
FourSight: Stable FP4 RL Post-Training via Sequence Filtering and Sparse Channel Compensation
Reinforcement learning (RL) post-training increasingly targets low-bit deployment, with NVFP4 weight-and activation quantization (W4A4) becoming the state-of-the-practice format. However, integrating NVFP4 into RL degrades rollout quality and amplifies cross-engine training-inference mismatch (TIM)…
arxiv.org
November 4, 2026 at 10:44 PM
FourSight: Stable FP4 RL Post-Training via Sequence Filtering and Sparse Channel Compensation
Yining Wei et al.
#arXiv #cs.AI #cs.LG
Yining Wei et al.
#arXiv #cs.AI #cs.LG
FourSight: Stable FP4 RL Post-Training via Sequence Filtering and Sparse Channel Compensation
Reinforcement learning (RL) post-training increasingly targets low-bit deployment, with NVFP4 weight-and activation quantization (W4A4) becoming the state-of-the-practice format. However, integrating NVFP4 into RL degrades rollout quality and amplifies cross-engine training-inference mismatch (TIM)…
arxiv.org
November 4, 2026 at 9:19 PM
LoopKV optimizes memory for looped transformers, allocating quantization levels based on layer sensitivity, reducing cache size by 35% and increasing decoding throughput by up to 2.2 times over standard methods. https://arxiv.org/abs/2611.04137
LoopKV: Loop-Aware KV Cache Compression for Serving Looped Transformers
ArXiv link for LoopKV: Loop-Aware KV Cache Compression for Serving Looped Transformers
arxiv.org
November 4, 2026 at 9:10 PM
LoopKV optimizes memory for looped transformers, allocating quantization levels based on layer sensitivity, reducing cache size by 35% and increasing decoding throughput by up to 2.2 times over standard methods. https://arxiv.org/abs/2611.04137
LoopKV boosts transformer efficiency, reducing KV cache storage in looped models by 35% while enhancing throughput 1.7–2x compared to traditional methods. This technique customizes quantization per layer, improving performance without sacrificing accuracy. https://arxiv.org/abs/2611.04137
LoopKV: Loop-Aware KV Cache Compression for Serving Looped Transformers
ArXiv link for LoopKV: Loop-Aware KV Cache Compression for Serving Looped Transformers
arxiv.org
November 4, 2026 at 8:30 PM
LoopKV boosts transformer efficiency, reducing KV cache storage in looped models by 35% while enhancing throughput 1.7–2x compared to traditional methods. This technique customizes quantization per layer, improving performance without sacrificing accuracy. https://arxiv.org/abs/2611.04137
LoopKV optimizes key-value cache compression for looped transformers, reducing memory by 35% and boosting throughput by 2.2x over traditional methods. It tailors quantization precision per layer, enhancing performance for language models without sacrificing accuracy. https://arxiv.org/abs/2611.04137
LoopKV: Loop-Aware KV Cache Compression for Serving Looped Transformers
ArXiv link for LoopKV: Loop-Aware KV Cache Compression for Serving Looped Transformers
arxiv.org
November 4, 2026 at 8:01 PM
LoopKV optimizes key-value cache compression for looped transformers, reducing memory by 35% and boosting throughput by 2.2x over traditional methods. It tailors quantization precision per layer, enhancing performance for language models without sacrificing accuracy. https://arxiv.org/abs/2611.04137
Twice as Fast, Until It Isn't: Quantized Serving of Mid-Size Open LLMs
Smit Kadvani, Kriti Faujdar, Spurthi Tallam
#arXiv #cs.AI
Smit Kadvani, Kriti Faujdar, Spurthi Tallam
#arXiv #cs.AI
Twice as Fast, Until It Isn't: Quantized Serving of Mid-Size Open LLMs
A team serving a 9B-20B open-weight language model on vLLM or SGLang must choose a quantization recipe, and needs two numbers: the speedup over BF16 and the accuracy given up. We assemble every public measurement of these numbers that we could verify and find that the record supplies them only in p…
arxiv.org
November 4, 2026 at 6:34 PM
they claim a <100x overhead but their proofs are about 4-bit integer quantization while the baseline is FP8?
November 4, 2026 at 6:20 PM
they claim a <100x overhead but their proofs are about 4-bit integer quantization while the baseline is FP8?
Confusing CONTRIB with CONTRIBUTING is a nice example of what low bit quantization does to token level distinctions. I would be curious whether keeping the embedding layers at higher precision fixes that.
November 4, 2026 at 4:43 PM
Confusing CONTRIB with CONTRIBUTING is a nice example of what low bit quantization does to token level distinctions. I would be curious whether keeping the embedding layers at higher precision fixes that.
TEMPO extends a large language model's vocabulary with discrete codes from a context-aware transformer tokenizer using finite scalar quantization, allowing signal patches to be mapped to a 625-code vocabulary and processed alongside text.
Source: arXiv cs.CL
Source: arXiv cs.CL
TEMPO: Time Series Understanding via Discrete Tokenization
Abstract page for arXiv paper 2611.04482: TEMPO: Time Series Understanding via Discrete Tokenization
arxiv.org
November 4, 2026 at 11:52 AM
TEMPO extends a large language model's vocabulary with discrete codes from a context-aware transformer tokenizer using finite scalar quantization, allowing signal patches to be mapped to a 625-code vocabulary and processed alongside text.
Source: arXiv cs.CL
Source: arXiv cs.CL
The FP4Quant framework employs micro-grouped dual-scale quantization and a hybrid FWHT pipeline to reduce communication overhead in distributed model-parallel LLM training.
#LLM #Quantization #DistributedTraining
#LLM #Quantization #DistributedTraining
Micro-grouped FP4 quantization by FP4Quant
Source: arXiv cs.AI
realfeedapp.com
November 4, 2026 at 10:03 AM
The FP4Quant framework employs micro-grouped dual-scale quantization and a hybrid FWHT pipeline to reduce communication overhead in distributed model-parallel LLM training.
#LLM #Quantization #DistributedTraining
#LLM #Quantization #DistributedTraining
Exotic Pauli spinors and magnetic flux quantization in a superconducting ring
https://arxiv.org/pdf/2611.03877
L. Ibrain Cohen Iskandar, J. M. Hoff da Silva, R. T. Cavalcanti.
https://arxiv.org/pdf/2611.03877
L. Ibrain Cohen Iskandar, J. M. Hoff da Silva, R. T. Cavalcanti.
https://arxiv.org/abs/2611.03877
arXiv abstract link
arxiv.org
November 4, 2026 at 9:33 AM
Exotic Pauli spinors and magnetic flux quantization in a superconducting ring
https://arxiv.org/pdf/2611.03877
L. Ibrain Cohen Iskandar, J. M. Hoff da Silva, R. T. Cavalcanti.
https://arxiv.org/pdf/2611.03877
L. Ibrain Cohen Iskandar, J. M. Hoff da Silva, R. T. Cavalcanti.
Pump quantization changes optimal measurements for quantum sensing
https://arxiv.org/pdf/2611.03943
S. A. Wadood, L. Bello, T. Ilias, J. W. Fleischer.
https://arxiv.org/pdf/2611.03943
S. A. Wadood, L. Bello, T. Ilias, J. W. Fleischer.
https://arxiv.org/abs/2611.03943
arXiv abstract link
arxiv.org
November 4, 2026 at 9:32 AM
Pump quantization changes optimal measurements for quantum sensing
https://arxiv.org/pdf/2611.03943
S. A. Wadood, L. Bello, T. Ilias, J. W. Fleischer.
https://arxiv.org/pdf/2611.03943
S. A. Wadood, L. Bello, T. Ilias, J. W. Fleischer.
S. A. Wadood, L. Bello, T. Ilias, J. W. Fleischer: Pump quantization changes optimal measurements for quantum sensing https://arxiv.org/abs/2611.03943 https://arxiv.org/pdf/2611.03943 https://arxiv.org/html/2611.03943
November 4, 2026 at 6:59 AM
S. A. Wadood, L. Bello, T. Ilias, J. W. Fleischer: Pump quantization changes optimal measurements for quantum sensing https://arxiv.org/abs/2611.03943 https://arxiv.org/pdf/2611.03943 https://arxiv.org/html/2611.03943
Making Local LLMs Usable on 8GB GPUs
Learn how to optimize local LLM inference on 8GB GPUs using KV-cache quantization, Flash Attention, batching, and practical latency benchmarks.
#hackernews #llm #news
Learn how to optimize local LLM inference on 8GB GPUs using KV-cache quantization, Flash Attention, batching, and practical latency benchmarks.
#hackernews #llm #news
Making Local LLMs Usable on 8GB GPUs
Learn how to optimize local LLM inference on 8GB GPUs using KV-cache quantization, Flash Attention, batching, and practical latency benchmarks.
hackernoon.com
November 4, 2026 at 6:59 AM
Making Local LLMs Usable on 8GB GPUs
Learn how to optimize local LLM inference on 8GB GPUs using KV-cache quantization, Flash Attention, batching, and practical latency benchmarks.
#hackernews #llm #news
Learn how to optimize local LLM inference on 8GB GPUs using KV-cache quantization, Flash Attention, batching, and practical latency benchmarks.
#hackernews #llm #news
L. Ibrain Cohen Iskandar, J. M. Hoff da Silva, R. T. Cavalcanti: Exotic Pauli spinors and magnetic flux quantization in a superconducting ring https://arxiv.org/abs/2611.03877 https://arxiv.org/pdf/2611.03877 https://arxiv.org/html/2611.03877
November 4, 2026 at 6:54 AM
L. Ibrain Cohen Iskandar, J. M. Hoff da Silva, R. T. Cavalcanti: Exotic Pauli spinors and magnetic flux quantization in a superconducting ring https://arxiv.org/abs/2611.03877 https://arxiv.org/pdf/2611.03877 https://arxiv.org/html/2611.03877
Lihao Zhao: A note on the BRST-BV Formalism Explanation of Deformation Quantization and Algebraic Index Theorem https://arxiv.org/abs/2611.04258 https://arxiv.org/pdf/2611.04258 https://arxiv.org/html/2611.04258
November 4, 2026 at 6:47 AM
Lihao Zhao: A note on the BRST-BV Formalism Explanation of Deformation Quantization and Algebraic Index Theorem https://arxiv.org/abs/2611.04258 https://arxiv.org/pdf/2611.04258 https://arxiv.org/html/2611.04258
Minsoo Cheong, Sungjoo Yoo: LoopVQ: Sub 1-bit KV Cache Quantization for Looped Transformers https://arxiv.org/abs/2611.05304 https://arxiv.org/pdf/2611.05304 https://arxiv.org/html/2611.05304
November 4, 2026 at 6:44 AM
Minsoo Cheong, Sungjoo Yoo: LoopVQ: Sub 1-bit KV Cache Quantization for Looped Transformers https://arxiv.org/abs/2611.05304 https://arxiv.org/pdf/2611.05304 https://arxiv.org/html/2611.05304
Xi Chen, Carl Nuzman, Shirin Jalali: Pairwise Distortion Distribution with Graph-Based Importance Sampling for Quantization Error and Intrinsic Dimension Estimation https://arxiv.org/abs/2611.04341 https://arxiv.org/pdf/2611.04341 https://arxiv.org/html/2611.04341
November 4, 2026 at 6:42 AM
Xi Chen, Carl Nuzman, Shirin Jalali: Pairwise Distortion Distribution with Graph-Based Importance Sampling for Quantization Error and Intrinsic Dimension Estimation https://arxiv.org/abs/2611.04341 https://arxiv.org/pdf/2611.04341 https://arxiv.org/html/2611.04341
Jaemin Choi, Jincheol Yang, Nahyun Lim, Yun-Seong Jeong, Matti Zinke, Hyunwoo Yu, Suk-Ju Kang: SPA-Q: Structure-Preserving Adaptive Post-Training Quantization for Monocular Depth Estimation https://arxiv.org/abs/2611.04491 https://arxiv.org/pdf/2611.04491 https://arxiv.org/html/2611.04491
November 4, 2026 at 6:40 AM
Jaemin Choi, Jincheol Yang, Nahyun Lim, Yun-Seong Jeong, Matti Zinke, Hyunwoo Yu, Suk-Ju Kang: SPA-Q: Structure-Preserving Adaptive Post-Training Quantization for Monocular Depth Estimation https://arxiv.org/abs/2611.04491 https://arxiv.org/pdf/2611.04491 https://arxiv.org/html/2611.04491