VibeServe demonstrates that AI agents can synthesize bespoke LLM serving systems end-to-end, remaining competitive with vLLM in standard settings while outperforming it in six non-standard scenarios involving unusual models, workloads, or hardware.
hub Canonical reference
Jamba: A Hybrid Transformer-Mamba Language Model
Canonical reference. 79% of citing Pith papers cite this work as background.
68
Pith papers citing it
41
external citations · Pith
Background
79% of classified citations
abstract
We present Jamba, a new base large language model based on a novel hybrid Transformer-Mamba mixture-of-experts (MoE) architecture. Specifically, Jamba interleaves blocks of Transformer and Mamba layers, enjoying the benefits of both model families. MoE is added in some of these layers to increase model capacity while keeping active parameter usage manageable. This flexible architecture allows resource- and objective-specific configurations. In the particular configuration we have implemented, we end up with a powerful model that fits in a single 80GB GPU. Built at large scale, Jamba provides high throughput and small memory footprint compared to vanilla Transformers, and at the same time state-of-the-art performance on standard language model benchmarks and long-context evaluations. Remarkably, the model presents strong results for up to 256K tokens context length. We study various architectural decisions, such as how to combine Transformer and Mamba layers, and how to mix experts, and show that some of them are crucial in large scale modeling. We also describe several interesting properties of these architectures which the training and evaluation of Jamba have revealed, and plan to release checkpoints from various ablation runs, to encourage further exploration of this novel architecture. We make the weights of our implementation of Jamba publicly available under a permissive license.
hub tools
citation-role summary
background 12
method 2
citation-polarity summary
representative citing papers
Content-based routing succeeds only when models provide bidirectional context and perform pairwise comparisons, with bidirectional Mamba plus rank-1 projection reaching 99.7% precision at linear inference cost.
Multi-agent LLMs classify USPTO reactions and write verified SMIRKS rules, expanding a reaction taxonomy from 68 to 14,073 classes and matching proprietary classifiers on held-out and out-of-distribution data.
Hypic enables position-independent KV/state reuse on hybrid-attention LLMs via cached segment transitions, seam-window repair, and inter-instance segment parallelism, cutting TTFT 3.25x over prefix caching.
FlashMorph formulates hybrid layer selection as budget-constrained optimization, trains per-layer gates on synthetic retrieval data with linearization regularization, then discretizes and distills to produce efficient hybrid architectures.
HRM adapters via Hankel reduced-order modeling outperform LoRA on long-context tasks in Mistral-7B when used as SSM residual modules with FFT-based parallel scan.
LongSpike integrates fractional-order state-space modeling into spiking neural networks, enabling better long-sequence performance than prior SNNs on LRA, WikiText-103, and Speech Commands benchmarks while retaining sparse computation.
Tangram makes non-uniform KV cache compression practical for LLM serving with deterministic budget allocation, head group paging, and ahead-of-time load balancing, achieving up to 2.6x throughput gains.
SISA adds an SSM importance term inside the attention score and runs the full operation as one SDPA call on augmented Q/K vectors, reporting better LAMBADA and perfect NIAH at small scale.
AVMP separates KV and SSM cache pools behind unified virtual addressing with failure-triggered migration, cutting OOM events 7.6% and raising throughput 1.83-13.3x on synthetic loads and 2.36x on ShareGPT traces.
Tensor Cache augments sliding-window attention with an eviction-fed outer-product associative memory and a training correction to improve long-context performance under bounded memory.
FRACTAL integrates fractional recurrent architecture into SSMs using a tunable singularity index to capture multi-scale temporal features, reporting 87.11% average on Long Range Arena and outperforming S5.
Star Elastic trains N nested submodels in a single post-training job on a parent reasoning LLM, supporting elastic budget control that matches or exceeds independent baselines while cutting training compute by up to 360x.
Component-aware self-speculative decoding achieves high acceptance rates in parallel hybrid models like Falcon-H1 but fails in sequential ones like Qwen3.5, with the gap tied to how components are integrated.
Short input phrases can irreversibly overwrite hidden states in Mamba models, impairing information retrieval on a new benchmark while leaving pure Transformer models unaffected.
MbaGCN combines message aggregation, selective state space transitions, and node state prediction to create a more scalable deep graph convolutional network.
Transformers and SSMs are unified through structured state space duality, producing a 2-8X faster Mamba-2 model that remains competitive with Transformers.
KV-cache eviction, prompt compression, recurrent state bounding, and agent memory consolidation are unified as one rate-distortion problem with a shared lower bound, shared failure mode, and transferable mechanisms.
HOLA pairs a compressive delta-rule recurrent state with a residual-selected exact KV cache and decoupled RMSNorm-gamma read, yielding lower perplexity than both standard linear attention and full-attention baselines on Wikitext and LAMBADA plus stronger needle-in-haystack recall.
R2LM combines causal attention with a reverse Mamba SSM sidecar to supply right-side context in dLLMs, claiming 2.4x-12.9x throughput gains over bidirectional dLLMs and 1.9x-2.9x over AR baselines while matching or exceeding quality.
Key-axis-only content-aware erase plus a scalar write gate restores WY-form chunk training for delta-rule models and improves 1.3B language-model quality over GDN-2.
SARA aligns internal routing distributions in MoE layers to high-resource semantic anchors via symmetric JS divergence, improving low-resource language performance by 0.8-1.2% over standard instruction tuning on Global-MMLU.
WiSP achieves up to 1.95x decode throughput on low-resource MoE serving by dynamically paging reused experts and using MV-WSA to allocate VRAM between experts and KV cache, with the offline policy performing well on both prefill and decode.
Reinforcement learning after SFT conversion narrows the performance gap between sliding-window attention and full self-attention on math reasoning benchmarks while preserving linear complexity.
citing papers explorer
-
VibeServe: Can AI Agents Build Bespoke LLM Serving Systems?
VibeServe demonstrates that AI agents can synthesize bespoke LLM serving systems end-to-end, remaining competitive with vLLM in standard settings while outperforming it in six non-standard scenarios involving unusual models, workloads, or hardware.
-
When Does Content-Based Routing Work? Representation Requirements for Selective Attention in Hybrid Sequence Models
Content-based routing succeeds only when models provide bidirectional context and perform pairwise comparisons, with bidirectional Mamba plus rank-1 projection reaching 99.7% precision at linear inference cost.
-
Agentic generation of verifiable rules for deterministic, self-expanding reaction classification
Multi-agent LLMs classify USPTO reactions and write verified SMIRKS rules, expanding a reaction taxonomy from 68 to 14,073 classes and matching proprietary classifiers on held-out and out-of-distribution data.
-
HYPIC: Accelerating Hybrid-Attention LLM Serving with Position-Independent Caching
Hypic enables position-independent KV/state reuse on hybrid-attention LLMs via cached segment transitions, seam-window repair, and inter-instance segment parallelism, cutting TTFT 3.25x over prefix caching.
-
Morphing into Hybrid Attention Models
FlashMorph formulates hybrid layer selection as budget-constrained optimization, trains per-layer gates on synthetic retrieval data with linearization regularization, then discretizes and distills to produce efficient hybrid architectures.
-
SSM Adapters via Hankel Reduced-order Modeling: Injection Site Determines Task Suitability in Long-Context Fine-Tuning
HRM adapters via Hankel reduced-order modeling outperform LoRA on long-context tasks in Mistral-7B when used as SSM residual modules with FFT-based parallel scan.
-
LongSpike: Fractional Order Spiking State Space Models for Efficient Long Sequence Learning
LongSpike integrates fractional-order state-space modeling into spiking neural networks, enabling better long-sequence performance than prior SNNs on LRA, WikiText-103, and Speech Commands benchmarks while retaining sparse computation.
-
Tangram: Unlocking Non-Uniform KV Cache for Efficient Multi-turn LLM Serving
Tangram makes non-uniform KV cache compression practical for LLM serving with deterministic budget allocation, head group paging, and ahead-of-time load balancing, achieving up to 2.6x throughput gains.
-
Forget Attention: Importance-Aware Attention Is All You Need
SISA adds an SSM importance term inside the attention score and runs the full operation as one SDPA call on augmented Q/K vectors, reporting better LAMBADA and perfect NIAH at small scale.
-
Asymmetric Virtual Memory Paging for Hybrid Mamba-Transformer Inference
AVMP separates KV and SSM cache pools behind unified virtual addressing with failure-triggered migration, cutting OOM events 7.6% and raising throughput 1.83-13.3x on synthetic loads and 2.36x on ShareGPT traces.
-
Tensor Cache: Eviction-conditioned Associative Memory for Transformers
Tensor Cache augments sliding-window attention with an eviction-fed outer-product associative memory and a training correction to improve long-context performance under bounded memory.
-
FRACTAL: SSM with Fractional Recurrent Architecture for Computational Temporal Analysis of Long Sequences
FRACTAL integrates fractional recurrent architecture into SSMs using a tunable singularity index to capture multi-scale temporal features, reporting 87.11% average on Long Range Arena and outperforming S5.
-
Star Elastic: Many-in-One Reasoning LLMs with Efficient Budget Control
Star Elastic trains N nested submodels in a single post-training job on a parent reasoning LLM, supporting elastic budget control that matches or exceeds independent baselines while cutting training compute by up to 360x.
-
Component-Aware Self-Speculative Decoding in Hybrid Language Models
Component-aware self-speculative decoding achieves high acceptance rates in parallel hybrid models like Falcon-H1 but fails in sequential ones like Qwen3.5, with the gap tied to how components are integrated.
-
Hidden State Poisoning Attacks against Mamba-based Language Models
Short input phrases can irreversibly overwrite hidden states in Mamba models, impairing information retrieval on a new benchmark while leaving pure Transformer models unaffected.
-
Mamba-Based Graph Convolutional Networks: Tackling Over-smoothing with Selective State Space
MbaGCN combines message aggregation, selective state space transitions, and node state prediction to create a more scalable deep graph convolutional network.
-
Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality
Transformers and SSMs are unified through structured state space duality, producing a 2-8X faster Mamba-2 model that remains competitive with Transformers.
-
What to Keep, What to Forget: A Rate--Distortion View of Memory Compaction in LLMs and Agents
KV-cache eviction, prompt compression, recurrent state bounding, and agent memory consolidation are unified as one rate-distortion problem with a shared lower bound, shared failure mode, and transferable mechanisms.
-
A Hippocampus for Linear Attention: An Exact Memory for What the Recurrent State Forgets
HOLA pairs a compressive delta-rule recurrent state with a residual-selected exact KV cache and decoupled RMSNorm-gamma read, yielding lower perplexity than both standard linear attention and full-attention baselines on Wikitext and LAMBADA plus stronger needle-in-haystack recall.
-
Bifocal Diffusion Language Models: Asymmetric Bidirectional Context for Parallel Generation
R2LM combines causal attention with a reverse Mamba SSM sidecar to supply right-side context in dLLMs, claiming 2.4x-12.9x throughput gains over bidirectional dLLMs and 1.9x-2.9x over AR baselines while matching or exceeding quality.
-
CARVE: Content-Aware Recurrent with Value Efficiency for Chunk-Parallel Linear Attention
Key-axis-only content-aware erase plus a scalar write gate restores WY-form chunk training for delta-rule models and improves 1.3B language-model quality over GDN-2.
-
SARA: Unlocking Multilingual Knowledge in Mixture-of-Experts via Semantically Anchored Routing Alignment
SARA aligns internal routing distributions in MoE layers to high-resource semantic anchors via symmetric JS divergence, improving low-resource language performance by 0.8-1.2% over standard instruction tuning on Global-MMLU.
-
WiSP: A Working-Set View of Mixture-of-Experts Serving on Extremely Low-Resource Hardware
WiSP achieves up to 1.95x decode throughput on low-resource MoE serving by dynamically paging reused experts and using MV-WSA to allocate VRAM between experts and KV cache, with the offline policy performing well on both prefill and decode.
-
Architecture-Aware Reinforcement Learning Makes Sliding-Window Attention Competitive in Math Reasoning
Reinforcement learning after SFT conversion narrows the performance gap between sliding-window attention and full self-attention on math reasoning benchmarks while preserving linear complexity.
-
Titans-as-a-Layer: Test-Time Memory for Conversational Speech Emotion Recognition
Introduces a Memory-as-a-Layer adapter that writes dialogue history into neural memory and reads it as a residual update to improve conversational speech emotion recognition on audio LLMs.
-
You Only Index Once: Cross-Layer Sparse Attention with Shared Routing
CLSA shares both KV cache and routing indices across decoder layers to amortize top-k selection, delivering up to 7.6x decoding speedup and 17.1x throughput at 128K context while preserving accuracy.
-
Zamba2-VL Technical Report
Zamba2-VL is a family of 1.2B–7B hybrid Mamba2-transformer vision-language models that match leading transformer VLMs on image, reasoning, OCR, grounding and counting benchmarks while delivering roughly 10x lower time-to-first-token.
-
The Routing and Filtering Structure of Attention
Attention decomposes into low-rank routing and symmetric filtering; disentangled S-D attention reveals a spectral cascade allowing early-layer linearization at under 5% perplexity cost.
-
Freeze Deep, Train Shallow: Interpretable Layer Allocation for Continued Pre-Training
LayerTracer analysis identifies deep LLM layers as stable task-critical regions, leading to a shallow-train deep-freeze strategy that outperforms full fine-tuning on C-Eval and CMMLU.
-
Priming: Hybrid State Space Models From Pre-trained Transformers
Priming transfers knowledge from pre-trained Transformers to hybrid SSM-attention models, recovering performance with minimal additional tokens and showing Gated KalmaNet outperforming Mamba-2 on long-context reasoning at 32B scale.
-
Echo: KV-Cache-Free Associative Recall with Spectral Koopman Operators
Spectral Koopman operators let SSMs achieve 100% accuracy on long-gap multi-query associative recall with fixed memory, where pure Mamba fails.
-
Retrieval from Within: An Intrinsic Capability of Attention-Based Models
Attention-based models can retrieve evidence intrinsically by using decoder attention to score and reuse their own pre-encoded chunks, outperforming separate retrieval pipelines on QA benchmarks.
-
Rhamba: Region-Aware Hybrid Attention-Mamba Framework for Self-Supervised Learning in Resting-State fMRI
Rhamba uses region-aware masking strategies and hybrid Attention-Mamba models pretrained on ABIDE fMRI data to achieve top AUROC on schizophrenia and ADHD classification tasks while outperforming prior methods.
-
Lost in State Space: Probing Frozen Mamba Representations
Frozen Mamba patch-boundary readouts do not outperform mean pooling for sentence representations on SST-2, CoLA, MRPC, STS-B, and IMDb due to anisotropy (cosine similarity ~0.9999) and representational collapse (MCC=0 on CoLA).
-
Long-Context Aware Upcycling: A New Frontier for Hybrid LLM Scaling
HyLo upcycles Transformer LLMs into hybrids with MLA and Mamba2/Gated DeltaNet blocks via staged training and distillation, extending context to 2M tokens and outperforming prior upcycled hybrids on long-context benchmarks.
-
HubRouter: A Pluggable Sub-Quadratic Routing Primitive for Hybrid Sequence Models
HubRouter is a sub-quadratic routing primitive using learned hubs that replaces attention layers in hybrid models while delivering competitive perplexity and large throughput gains.
-
MP-ISMoE: Mixed-Precision Interactive Side Mixture-of-Experts for Efficient Transfer Learning
MP-ISMoE uses Gaussian noise perturbed iterative quantization and interactive side mixture-of-experts to deliver higher accuracy than prior memory-efficient transfer learning methods while keeping similar parameter and memory usage.
-
Safety, Security, and Cognitive Risks in State-Space Models: A Systematic Threat Analysis with Spectral, Stateful, and Capacity Attacks
State-space models are vulnerable to three new attack types that corrupt state integrity, with experiments showing up to 156x output changes and 6x higher targeted corruption than random inputs.
-
Why Attend to Everything? Focus is the Key
Focus learns a few centroids to gate long-range token attention, producing sparse attention that matches or beats full attention quality with up to 8.6x speedup at million-token lengths.
-
When to Think Fast and Slow? AMOR: Adaptive Entropy Gate for Hybrid Models
AMOR uses output entropy to gate attention in recurrent hybrids, matching full attention performance at roughly 22% attention invocations across 180M-1.5B models.
-
DeMa: Dual-Path Delay-Aware Mamba for Efficient Multivariate Time Series Analysis
DeMa is a dual-path delay-aware Mamba architecture that decomposes MTS into intra-series temporal and inter-series variate paths to achieve SOTA performance with linear complexity on forecasting, imputation, anomaly detection, and classification.
-
Gated KalmaNet: A Fading Memory Layer Through Test-Time Ridge Regression
Gated KalmaNet uses exact Kalman gain computation with adaptive gating and Chebyshev iteration to improve SSM performance on long-context tasks over prior approximations like DeltaNet.
-
Next-Latent Prediction Transformers Learn Compact World Models
Adding a next-latent prediction loss to next-token training makes transformer hidden states more predictive of future tokens and improves planning/reasoning on small benchmarks.
-
SpikingBrain: Spiking Brain-inspired Large Models
SpikingBrain-7B and SpikingBrain-76B achieve Transformer-comparable performance after continual pre-training on 150B tokens, with over 100x TTFT speedup on 4M-token sequences and 69.15% sparsity from event-driven spiking.
-
eLLM: Elastic Memory Management Framework for Efficient LLM Serving
eLLM unifies LLM memory management with virtual tensors and elastic ballooning to CPU memory, reporting 2.32x higher decoding throughput and 3x larger batch sizes for 128K inputs.
-
When Attention Sink Emerges in Language Models: An Empirical View
Attention sinks emerge in language models from softmax-induced token dependence on attention scores and do not appear when using sigmoid attention without normalization in models up to 1B parameters.
-
An Empirical Study of Mamba-based Language Models
An 8B Mamba-2-Hybrid with 43% Mamba-2, 7% attention, and 50% MLP layers exceeds an 8B Transformer by 2.65 points on average across 12 tasks and matches it on 23 long-context tasks while enabling up to 8x faster inference.
-
How Much Dense Attention is Necessary? Oracle-Guided Sparse Prefill for Full/GQA Layers in Hybrid Long-Context Models
An oracle shows sparse token support preserves near-dense performance on Qwen retrieval tasks, and a KL-distilled head-collapsed indexer delivers 1.7-1.9x speedups with small quality gaps.
-
MOSAIC: Efficient Mixture-of-Agent Scheduling via Adaptive Aggregation and Inference Concurrency
MOSAIC uses an Integer Linear Program scheduler for expert placement and prompt assignment plus adaptive aggregation to achieve 1.7-2.3x end-to-end speedup on 4-GPU MoA workloads while keeping accuracy within 0.1pp.
-
Latent Recurrent Transformer: Architecture Exploration, Training Strategies, and Scaling Behavior
Latent Recurrent Transformer augments autoregressive transformers with a cross-layer recurrent latent pathway from prior hidden states and uses interleaved parallel training to improve loss and in-context learning at ~0.3% extra parameters.