long-context repositories

Trending repositories tagged long-context, ranked by trending score. Updated daily.

  1. 1
    JakeATX/llamAmpere

    llama.cpp fork for significantly improved performance on Ampere generation Nvidia cards (especially RTX 3090 / 3090 Ti): example: 110+ tok/s over a 100K-token generation at temperature 1 for qwen3.8, supporting full 262K ctx. Supports low memory (12gb) configs for 200k+ ctx. 120+ TPS, 3300+ prefill for QFN 3.8.

    GitHub · C++ · 403 stars · 50 forks · +87 stars in 24 hours · Trending score 3.35

  2. 2
    peonist-ai/halogen-flash-server

    Local inference that behaves like a frontier API: Qwen3.8-Flash-Next on AMD Strix Halo plus small NPU models, over the OpenAI and Anthropic APIs on one port.

    GitHub · Shell · 1,403 stars · 101 forks · +17 stars in 24 hours · Trending score 2.66

  3. 3
    DevDock-AI/Speedy

    Qwen3.8-Flash-Next (125B MoE) up to 1M context on one or two RTX cards (tuned on 3080 10gb). A measured, tuned and benchmarked fork of Strata.

    GitHub · C++ · 125 stars · 8 forks · +26 stars in 24 hours · Trending score 2.59

  4. 4
    snap-stanford/memorilla

    🦍 Memorilla: Latent Semantic Memory for LLMs

    GitHub · Python · 1,283 stars · 85 forks · 0 stars in 24 hours · Trending score 2.11

  5. 5
    MemTensor/MemT-1.0

    Official repository for MemT-1.0

    GitHub · Python · 545 stars · 44 forks · +3 stars in 24 hours · Trending score 1.41

  6. 6
    ARahim3/kaggle-tpu-lab

    Frontier-class open models on a free Kaggle TPU v5e-8: GLM-5.3-Flash 320B MoE (~165 tok/s, our own JAX engine) and Qwen3.8-27B bf16 (~130 tok/s), 262k context, prefix caching. Works with Claude Code, Codex, opencode and pi.

    GitHub · Python · 813 stars · 110 forks · +2 stars in 24 hours · Trending score 1.32

  7. 7
    huawei-csl/KVarN

    KVarN is a native vLLM KV-cache quantization backend for your agents: 3-5x more context, throughput above FP16, and FP16-level accuracy. Calibration-free, one flag.

    GitHub · Python · 521 stars · 38 forks · 0 stars in 24 hours · Trending score 1.27

  8. 8
    LilMGenius/paperthin

    The skillbook for recursive self-improvement (RSI). NVIDIA's latest proof, unpacked with timeless engineering. Distilled from the OpenAI harness engineering hackathon winner's zero-touch method.

    GitHub · JavaScript · 1,448 stars · 123 forks · 0 stars in 24 hours · Trending score 1.23

  9. 9
    tuan3w/qwen-warp

    Fast Qwen3.8-27B for one RTX 3090. 200k context, 100+ tok/s. Custom CUDA decode engine, OpenAI-compatible API.

    GitHub · Python · 49 stars · 6 forks · +3 stars in 24 hours · Trending score 1.16

  10. 10
    NVIDIA/kvpress

    LLM KV cache compression made easy

    GitHub · Python · 1,255 stars · 201 forks · 0 stars in 24 hours · Trending score 0.76

  11. 11
    Veloresearch/impulse

    Impulse by Velocity: run AI models on your own computer, on Windows and Linux. Load any GGUF model and chat with web search, files, charts and voice, let it use your computer, and serve it to any app through an OpenAI-compatible API and MCP. Momentum memory up to 32M tokens. No cloud.

    GitHub · 37 stars · 4 forks · +1 stars in 24 hours · Trending score 0.55

  12. 12
    CSUNX233/Sparkle-Qwen27b-Kvmem-ninfer

    RTX 5070 Ti Laptop 12GB 跑 Qwen3.8-27B:NInfer + KVMem、256K 逻辑上下文、RAM KV、MTP、快速缓存续接与 dsh 本地 Agent / 搜索 / 浏览器。含源码、Windows 部署及实测。

    GitHub · C++ · 101 stars · 5 forks · 0 stars in 24 hours · Trending score 0.48

  13. 13
    pablo-reyes8/deepseek-v4-mini-pytorch

    From-scratch, paper-faithful PyTorch implementation of DeepSeek-V4 architecture for transparent study, testing, ablation, and mini-scale training.

    GitHub · Python · 13 stars · 1 forks · 0 stars in 24 hours · Trending score 0.46

  14. 14
    pepuscz/strix-halo-llm

    DeepSeek-V4.1-Flash (1M context) and GLM-5.3-Flash on a BOSGAME M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GiB) with two RX 7900 XTX over OCuLink: pinned Ansible recipes and measured results. Also the M5-only Qwen and DeepSeek setups.

    GitHub · Python · 31 stars · 2 forks · 0 stars in 24 hours · Trending score 0.43

  15. 15
    qflen/nsa-from-scratch

    From-scratch reimplementation of DeepSeek's Native Sparse Attention (arXiv:2502.11089) in Triton + CUDA Hopper WGMMA. 7.07x faster than FlashAttention-3 at 64k context. Five-model training fleet, perplexity sweep, LongBench v2, MoBA comparison.

    GitHub · Python · 15 stars · 0 forks · 0 stars in 24 hours · Trending score 0.36

  16. 16
    massimodeluisa/recursive-decomposition-skill

    Claude Code skill for handling long-context tasks through recursive decomposition

    GitHub · HTML · 55 stars · 2 forks · 0 stars in 24 hours · Trending score 0.36

  17. 17
    BobClawblaw/deepseek-v41-flash-mlx-m5-ultra

    DeepSeek-V4.1-Flash-1017 (2/3-bit experts) on one 256 GB M5 Ultra: the full 1M context, DSpark drafts, int8 x int8 prompt experts, up to four streams, images, tools, JSON. TensorFold native Metal.

    GitHub · Python · 16 stars · 1 forks · 0 stars in 24 hours · Trending score 0.30

  18. 18
    skylight-org/skylight-research

    Advancing the frontier of efficient AI

    GitHub · Python · 70 stars · 16 forks · 0 stars in 24 hours · Trending score 0.29

  19. 19
    wanqili857-byte/fictionforge

    多智能体小说写作框架 | Multi-agent novel-writing framework — 每个角色都有自己的脑子,每一章都有质量门禁

    GitHub · Python · 56 stars · 3 forks · 0 stars in 24 hours · Trending score 0.22

  20. 20
    Tencent-Hunyuan/FlashTopK

    Fast, exact fused top-k selection for learned sparse attention training and inference, from HyScale (Tencent Hy & University of Edinburgh).

    GitHub · Python · 16 stars · 2 forks · 0 stars in 24 hours · Trending score 0.16

  21. 21
    ZipengWu365/TOMC

    TOMC Memory: CPU context compression and task memory for Codex, Claude Code, Cursor, TRAE, WorkBuddy and other MCP clients.

    GitHub · Python · 264 stars · 26 forks · 0 stars in 24 hours · Trending score 0.10

  22. 22
    yuuki-net/FreeToken-Kai

    Unofficial FreeToken fork: Flash-Next 125B at 38 tok/s on two RTX 3060s with 2-bit experts, gpt-oss-120b on one; a benchmark tunes your settings. Qwen3.8-Flash-Next, 12 GB, MTP, 250k context, image input, WSL2. Runs on Turing: RTX 2060, RTX 20 series, sm_75.

    GitHub · Python · 11 stars · 2 forks · 0 stars in 24 hours · Trending score 0.09

  23. 23
    Clare-Nie/MemOnce

    Official implementation of "MemOnce: A Scalable Memory Architecture for Large Language Models".

    GitHub · Python · 10 stars · 1 forks · 0 stars in 24 hours · Trending score 0.08

  24. 24
    IonDen/mlx-train-perf

    Fused, logit-free linear-cross-entropy loss, RAM-fit planner, and benchmark harness for MLX fine-tuning on Apple Silicon

    GitHub · Python · 6 stars · 1 forks · 0 stars in 24 hours · Trending score 0.07