LongCat-Flash-Exp: Efficient Long-Context Modeling

• LongCat-Flash-Exp is a suite for efficient long-context language modeling that integrates a Mixture-of-Experts architecture with zero-compute experts and dynamic routing.
• It leverages LongCat ZigZag Attention (LoZA) for block-sparse, streaming topologies, scaling performance to 1M tokens in agentic reasoning tasks.
• The framework combines advanced training regimens, overlapping compute strategies, and best practices for optimal throughput and cost efficiency.

LongCat-Flash-Exp refers to a comprehensive suite of experiments and model designs centered on efficient long-context language modeling. It encompasses the evaluation and extension of the LongCat-Flash foundation model, its mid-training modifications via LoZA (LongCat ZigZag Attention), and rigorous benchmarking—both as a Mixture-of-Experts (MoE) Transformer (560B→1.2T param) and versus state-of-the-art linear attention alternatives (e.g., LAWCAT). LongCat-Flash-Exp’s primary emphasis is on agentic reasoning, throughput, and efficient inference at extreme sequence lengths, notably scaling to 1 million tokens with competitive or superior task performance.

1. Model Architecture and Innovations

LongCat-Flash-Exp is based on variants of the LongCat-Flash architecture, a Mixture-of-Experts (MoE) Transformer developed for scalable, cost-effective, and adaptable compute allocation. The core innovations are:

• Zero-Compute Experts:

Many-token inference exhibits high per-token variability in computational complexity. LongCat-Flash introduces $Z$ zero-compute experts, which simply forward the input for "easy" tokens, and $N$ standard FFN experts. A softmax router with adaptive biasing (PID-controlled) ensures a target number $K_e$ of FFN experts versus $K-K_e$ zero experts is selected per token:

$$\text{MoE}(x_t) = \sum_{i=1}^{N+Z} g_i E_i(x_t), \quad E_i = \begin{cases} \mathrm{FFN}_i(x_t), & i \le N \ x_t, & N < i \le N+Z \end{cases}$$

Router bias $b$ dynamically adapts to maintain $K_e$ real experts averaged over time.

• Shortcut-Connected MoE (ScMoE):

The model introduces cross-layer shortcuts, allowing communication (dispatch/combine) and computation between experts and dense blocks to be efficiently overlapped, resulting in a significant reduction in time-per-output token (TPOT). For instance, TPOT is 16 ms in LongCat-Flash (SBO), versus 30 ms in DeepSeek V3 (TBO) (Team et al., 1 Sep 2025).

• Layer and Width Scaling:

To achieve stable scaling to 560B–1.2T parameters, the model leverages a scheme where proxy hyperparameters are transferred to the final model via "Adam LR Full Align" rules, and layer stacking (half-depth student checkpoint duplicated) optimizes convergence.

• Stability Suite:

The training pipeline includes mechanisms such as router-vs-LM gradient balancing ($R_g < 0.1$), a hidden $z$-loss to control activation spikes, and extremely small Adam $\epsilon$ initialization ($N$0) to manage adaptive step-sizes at scale.

2. LongCat ZigZag Attention (LoZA) and Sparse Extension

LongCat-Flash-Exp incorporates LoZA for efficient sparse attention at extreme context lengths (Zhang et al., 30 Dec 2025):

• Sparse Block Attention Pattern:

Each query attends to $N$1 "sink blocks" (coarse-grained, e.g., one global per $N$2 tokens) and $N$3 local blocks of size $N$4; the total number of attended blocks for each token is $N$5. The block-sparse binary mask $N$6 is defined as:

$N$7

This forms a streaming-sparse attention topology.

• Lottery-Ticket Layer Calibration:

During calibration, all weights are frozen, learnable scalars $N$8 interpolate between full and block-sparse outputs in each MLA layer. After optimizing $N$9 on out-of-distribution calibration data, 50% of layers with lowest $K_e$0 are converted to pure block-sparse. The remaining model undergoes further mid-training to recover potential performance gaps.

• 1M Token Contexts and YaRN Extrapolation:

A staged context expansion (32K → 128K → 256K) is followed by YaRN-based linear scaling of attention dropout and rotary position embeddings to extrapolate generalization to 1M tokens.

3. Training Regimen and Data Mixture

• General Pretraining:
  • Data decontamination: Test overlap (≥13-grams) and semantic filtering (BGE-m3, threshold 0.7–0.9).
  • Stage-2/3 target mix: 70% STEM/code (stage 2), 25% long-context data (stage 3).
• Long Context and Agentic Data:

Mid- and post-training focuses on multi-turn reasoning, agentic tool-use, and lengthy context documents (up to 1M).

• Curriculum Transfer and Layer Stacking:

A half-depth checkpoint is duplicated to form full-depth initialization, preserving optimizer state and accelerating convergence.

4. Empirical Performance and Benchmarks

The following tables summarize key throughput and task results obtained in LongCat-Flash-Exp (Team et al., 1 Sep 2025, Zhang et al., 30 Dec 2025):

Inference Throughput and Cost (LongCat-Flash):

Precision	Context	GPUs	TPS/u	Cost ($/M tokens)
bf16	5,000	128	100.5	0.7
fp8	8,192	128	33.8	0.7

End-to-End Prefill/Decode on LoZA-augmented Model (H20 cluster):

Mode	LongCat-Flash	LongCat-Flash-Exp
Prefill	12 tokens/ms	25 tokens/ms
Decode	0.8 tokens/ms	1.1 tokens/ms

Selected Leaderboard Scores—LongCat-Flash (27B active):

Benchmark	DeepSeek	Qwen3	Kimi-K2	LongCat-Flash
MMLU (%)	90.96	90.23	89.86	89.71
ArenaHard-V2	84.10	88.20	85.70	86.50
TerminalBench	31.30	17.28	25.93	39.51
τ²-Bench avg@4	49.13	43.01	64.17	67.65
VitaBench	20.3	8.5	18.2	24.3

LongBench and Agentic Benchmarks (LoZA extension):

• LongEval: Exp-Base 99.3% (vs Flash-Base 95.7%)
• SWE-Bench: 63.2 (vs 60.4)
• Terminal-Bench: 42.5 (vs 39.5)
• MRCR AUC: +5% over Qwen-3 (at 1M tokens)
• HELMET: 64.7% (vs 59.1% baseline)
• Ablation: hand-crafted interleaved sparsity on LongEval drops from 95.7 → 54.1; calibrated LoZA sparsity retains 89.6.

5. Linear Attention Alternatives and Comparative Analysis

LAWCAT provides an O(N) kernel with favorable properties relative to quadratic baselines (Liu et al., 22 Sep 2025):

• Causal Conv1D and Normalized Gated Linear Attention (GLA):

After a Conv1D smoothing over queries and keys (kernel size 4), outputs are fed into a normalized, gated linear recurrence:

$K_e$1

with explicit normalization critical for long-range stability.

• Distillation and Generalization:
  • Passkey retrieval: LAWCAT maintains 95–91% accuracy up to 16–22K, where FlashAttention-2 performance collapses past 8K.
  • Throughput: Linear attention kernels surpass FA2 in tokens/sec at sequence lengths >8K; memory remains $K_e$2.
• Edge and Streaming Suitability:

Low memory footprint (<12GB at 32K), linear latency, and seamless streaming favor deployment on single GPUs or in edge contexts.

6. Analysis, Limitations, and Best Practices

• Dynamic Compute Scaling:

Zero-compute experts and ScMoE enable substantial resource savings. Only "hard" tokens invoke FFN experts, lowering average active parameter count to ∼27B for a 560B model.

• Scalability and Stability:

Layer stacking, hyperparameter transfer, and the stability suite enable reproducible scaling with high hardware availability (98.48%). Tuning of Adam's $K_e$3, z-loss, and router balance parameter $K_e$4 is required at massive scale.

• Long-Context Generalization:

LoZA’s layer-level calibration, streaming block-sparse attention, and selective sparsification preserve or outperform full-attention baselines across both retrieval-augmented (prefill) and decode-intensive (tool-integrated) tasks up to 1M tokens.

• Limitations:

Slightly reduced performance is observed on some extreme long-context structured reasoning tasks (e.g., GraphWalks-128K). The engineering burden (custom kernels, routing-loss tuning, distributed infra) is nontrivial.

• Best Practices:

1. Implement zero-expert routing with a bias PID controller.
2. Calibrate router-vs-LM gradients to $K_e$5.
3. Control rare activation spikes using $K_e$6-loss.
4. Use width scaling to accelerate hyperparameter search.
5. Stack layered checkpoints for fast and reliable initialization.
6. Minimize Adam $K_e$7 to be lower than observed grad RMS.
7. Invest in overlapping comm/compute (ScMoE/SBO) for maximal throughput.
8. Deploy speculative decoding in inference with MTP heads and SBO.

7. Implications and Context in Long-Context Language Modeling

LongCat-Flash-Exp exemplifies a convergent trend in large-scale language modeling: fusing Mixture-of-Experts, dynamic compute routing, and sparse or linear attention paradigms. The combination of LoZA-based sparse extension and LAWCAT-style linearization demonstrates that the quadratic barrier to context scaling can be mitigated via both architectural and training-process innovations.

The reported suite also provides an empirical roadmap: integrate data decontamination, robust agentic/long-form training, and efficient, reproducible hardware and software pipelines. As a result, LongCat-Flash-Exp provides a foundational technical reference for the design, scaling, and evaluation of long-context agentic LLMs targeting both cloud and edge deployment scenarios (Team et al., 1 Sep 2025, Zhang et al., 30 Dec 2025, Liu et al., 22 Sep 2025).