KoboldCpp release history
The latest KoboldCpp is 1.124.2, released on 2 November 2026. New feature versions arrive about every two weeks, usually followed by quick hotfixes. Below, every release is summarised in our own words, with the changes that break old guides listed separately.
Checked against KoboldCpp 1.124.2 · updated 13 Nov 2026
- lab-lnx: Linux test VM · 4 vCPU · 15 GB · no GPU
- gha-w25: Windows Server 2025 test machine · 4 vCPU · 16 GB · no GPU
- gha-w22: Windows Server 2022 test machine · 4 vCPU · 16 GB · no GPU
- gha-u24: Ubuntu 24.04 test machine · 4 vCPU · 16 GB · no GPU
- gha-mac15: macOS 15 test VM (Apple Silicon) · 3 vCPU · 7 GB · paravirtual Metal
On this page
Latest release
KoboldCpp 1.124.2
Adds music generation with YuE2 and a batch of extra TTS models ported from audio.cpp, plus MOSS Image, Fibo 2 Edit and Flux 3 Klein via a stable-diffusion.cpp sync. Introduces --lazy-mode, --moe-cache-mib, --paralleltotalcontext and --rpccache, lets AI Horde workers serve chat-completions jobs, and turns on CUDA graphs for every architecture. Hotfixes trimmed the PyInstaller bundle and added SheetSage2 music-to-ABC transcription.
New flags in 1.124: --lazy-mode, --moe-cache-mib, --paralleltotalcontext, --rpccache. Nothing you already use changed name in this release.
tested We ran this version in our lab on 13 November 2026: how we test. Release page on GitHub
KoboldCpp releases: the timeline
Every release that has its own GitHub page, newest first, summarised in our own words. Filter by tag to find new backends, breaking changes or renamed flags. Each release has an anchor such as #v1-124-2.
- 1.124.2latest
Adds music generation with YuE2 and a batch of extra TTS models ported from audio.cpp, plus MOSS Image, Fibo 2 Edit and Flux 3 Klein via a stable-diffusion.cpp sync. Introduces --lazy-mode, --moe-cache-mib, --paralleltotalcontext and --rpccache, lets AI Horde workers serve chat-completions jobs, and turns on CUDA graphs for every architecture. Hotfixes trimmed the PyInstaller bundle and added SheetSage2 music-to-ABC transcription.
new backendnew featureGitHub release
- 1.123.1
Supports Qwen Image 2.1 and greatly extends the built-in Agent (approval modes, session save/load via --agentload, usage tracking). Adds --queuelimit, lets parallel (batched) mode handle more features, and marks --multiuser and the separate --jinjatools switch as deprecated.
breakingnew featureGitHub release
- 1.122.1
First release of the bundled KoboldCpp Agent (--agent), a small agentic harness with built-in tools and MCP support. Splits the physical micro-batch (--ubatchsize) from --batchsize, adds --autoswapthreshold, drops the pipeline-parallel switch, and lets koboldcpp.sh build portable binaries with KCPP_PORTABLE=1.
new featureGitHub release
- 1.121
Adds reference media for MiniMax H3 video, video LoRAs and a reworked MusicUI. New --ffncpu keeps dense FFN weights on the CPU, --analyze now also reads safetensors, and the row split mode is gone.
new featureGitHub release
- 1.120
Adds --usedirectio for direct-I/O model loading (combinable with mlock/mmap) and support for Qwen3.8-Flash-Next and Ling-3.0-flash. Kobold Lite can now run user-defined JavaScript tools; several image-generation fixes.
new featureGitHub release
- 1.119
Brings MiniMax H3 text-to-video and image-to-video, Muse Glimmer support and speculative decoding via DSpark/Dflash. Raises attachment limits and adds image preview/edit options in the UI.
new featureGitHub release
- 1.118.1
Removes CUDA row-split and ROCm WMMA support and aligns RPC behaviour with upstream llama.cpp. Image generation is now aborted when the client disconnects (opt-out request field), --jinjathink switches Jinja on automatically, plus multi-GPU and server-thread fixes.
new backendbreakingGitHub release
- 1.117.1
Adds Ollama-style embeddings plus streaming and tool calls on the Ollama endpoints so Ollama-only apps can talk to KoboldCpp. Multimodal images are now inserted inline in the prompt and image LoRA metadata is displayed; last version to offer row split.
new featureGitHub release
- 1.116.1
Adds Krea 2 Turbo, Ideogram 4 and Boogu Edit image models and MCP support in the bundled llama.cpp web UI. Ships a Metal-free macOS failsafe build, MP3 output for TTS, and --allow-config-onready to control whether configs may set --onready.
new backendnew featureGitHub release
- 1.115.2
Completes the Anthropic /v1/messages API and adds MTP-based drafting (--usemtp) and Gemma assistant drafting. Adds --reasoningeffort, --sdvramlimit and Gemma4 UV support; the CLI help was reordered alphabetically and the clip_quantize tool retired.
new featureGitHub release
- 1.114.1
Experimental parallel generation (--parallelrequests) and RPC GPU sharing (--rpcmode/--rpctargets/--rpcport) arrive, alongside LTX 2.3 video and new image models. Sliding-window attention is now on by default (--noswa to opt out) and --jinjathink is added.
new featureGitHub release
- 1.113.2
Moves multi-GPU split mode into its own --splitmode flag (adding experimental tensor split), which changes the old --usecuda rowsplit syntax. Adds --reqtimeout, a runtime image-LoRA directory, --musiclowvram, and memory/config-loading fixes.
new featureGitHub release
- 1.112.2
Adds AceStep XL music models and API reasoning-effort control. Replaces numeric --quantkv levels with named cache types (breaking), adds --swapadding, --baseconfig, --jinjatemplate, vision min/max token overrides and streamed Jinja tool calls.
breakingnew featureGitHub release
- 1.111.2
Adds Gemma 4 support, upstream rotation-based KV improvements, Qwen3-TTS voice design, first /v1/responses and /v1/messages endpoints, and --autoswapmode for multi-model configs. Reasoning text is now returned in a separate field (breaking for some clients); adds --jinja-kwargs and --sdmaingpu.
breakingnew featureGitHub release
- 1.110
Introduces router mode (--routermode), an OpenAI-API reverse proxy that hot-swaps models/configs per request, and --adminunloadtimeout for idle unloading. Adds Qwen3-TTS 1.7B, AceStep 1.5 improvements with MP3 output and multiple image-LoRA directories.
new featureGitHub release
- 1.109.2
Experimental ACE-Step music generation with its own Music UI (--musicllm/--musicdiffusion/--musicembeddings/--musicvae) and Qwen3-TTS voice cloning from a folder (--ttsdir). Also improves SmartCache for RNN/hybrid models and adds SillyTavern JSONL import and ComfyUI auth tokens.
new featureGitHub release
- 1.108.2
Autofit is improved and now engages automatically when GPU layers stay at -1 (the layer box is hidden in that case); adds --autofitpadding. Adds Ollama /api/show emulation, an SDUI reference-image toggle and an optional 2048x2048 image limit.
new featureGitHub release
- 1.107.3
Removes the CLBlast backend and adds a Vulkan path for old CPUs without AVX. Flash attention becomes the CLI default (--noflashattention to disable), adds --downloaddir, SDXS, ESRGAN upscaling (--sdupscaler) and Flux Klein.
new backendbreakingnew featureGitHub release
- 1.106.2
Adds MCP client/server support using a Claude-Desktop-style mcp.json (--mcpfile), plus llama.cpp-style --device selection. Flash attention becomes the default in the GUI launcher; tool parsing and the Lite settings UI are reworked.
new featureGitHub release
- 1.105.4
Adds --gendefaults (and --gendefaultsoverwrite) to inject or force API payload fields, replacing --sdgendefaults, plus an Adaptive-P sampler. Adds Ovis Image, newer Qwen Image Edit, TAEHV for WAN and SmartCache snapshot improvements.
new featureGitHub release
- 1.104
Adds SmartCache (--smartcache), which snapshots KV state to RAM for quick context switches, and --autofit for automatic GPU fitting. Pipeline parallelism becomes opt-in, --forceversion is deprecated, and a vision mrope fix improves Qwen3-VL.
breakingnew featureGitHub release
- 1.103
Adds Flux 2 and Z-Image Turbo image generation and merges Mistral/Ministral 3 support. The bundled llama.cpp UI gains better continue behaviour and an assistant-prefill option under /lcpp; Linux gets a secondary console viewer.
new featureGitHub release
- 1.102.3
Bundles the llama.cpp web UI at /lcpp and adds Jinja chat-template support (--jinja, --jinja_tools). Improves universal tool calling, syncs stable-diffusion.cpp and, in a hotfix, adds Qwen3-Next.
new featureGitHub release
- 1.101.1
Adds Qwen3-VL vision and Qwen Image / Qwen Image Edit generation. Renames --blasbatchsize to --batchsize (old name kept as alias) and lets --overridekv take several comma-separated values; the ROCm Linux build moves to ROCm 7.1.
new backendnew featureGitHub release
- 1.100.1
Adds WAN video generation and new image offload controls (--sdoffloadcpu, VAE/CLIP placement). --sdclipl/--sdclipg become --sdclip1/--sdclip2, low-VRAM mode becomes its own --lowvram flag, and GLM 4.6 and Granite 4 are supported.
new featureGitHub release
- 1.99.4
Redesigned Kobold Lite and a large set of llama.cpp-style flag aliases (-m, -c, -ngl, -ts, -fa, -ot, etc.) for easy migration. Adds per-IP --ratelimit and renames --promptlimit to --genlimit.
new featureGitHub release
- 1.98.1
Integrates TTS.cpp with Kokoro, Parler and Dia voices. Image gen gains --sdflashattention, --sdconvdirect and q8 quantization; adds --overridenativecontext for RoPE and Seed-OSS support.
new featureGitHub release
- 1.97.4
Adds GLM 4.5, GPT-OSS and Voxtral support, a /ping endpoint for serverless hosts and --moecpu to keep MoE expert weights on the CPU.
new featureGitHub release
- 1.96.2
Adds audio input for omni models (e.g. Qwen2.5-Omni) with wav/mp3/flac, raises the default context to 8192, and adds an OpenAI-style image endpoint plus image flash attention. --usecublas is renamed --usecuda (old name still accepted).
breakingflag renamednew featureGitHub release
- 1.95.1
Adds Flux Kontext natural-language image editing and multi-image PhotoMaker, Gemma 3n and ERNIE templates, and Python 3.13 fixes.
new featureGitHub release
- 1.94.2
Adds an unpacked mini-launcher, Chroma image models and PhotoMaker face cloning, --maingpu, soft/hard image resolution caps and --sdtiledvae (replacing --sdnotile), plus embeddings GPU/context options. First release using only the new asset names (koboldcpp-oldpc.exe etc.).
new featureGitHub release
- 1.93.2
Adds Windows file-association integration for .gguf, KV savestates to switch chats without reprocessing, admin-API model unload, --loramult and --singleinstance; the LoRA-base option is dropped. Binaries were published under both old and new names during this transition.
new featureGitHub release
- 1.92.1
Adds a sliding-window-attention mode (--useswa) that shrinks KV memory but conflicts with context shifting, more reliable tool calling, Llama 4 vision and a DDIM sampler. --skiplauncher now also disables --showgui.
new featureGitHub release
- 1.91
Adds a Hugging Face model search in the launcher, an embedded aria2c downloader on Windows, classifier-free guidance (--enableguidance) and makes KoboldCppAuto the default instruct preset.
new featureGitHub release
- 1.90.2
Adds Qwen3, Pixtral and GLM-4 fixes, OpenAI structured outputs and --mmprojcpu to keep the vision projector on the CPU. Batch-size behaviour changes and --onready is no longer honoured from .kcpps/.kcppt files.
breakingnew featureGitHub release
- 1.89
Reworks the no-JavaScript UI (/noscript) to cover chat and images, and adds --overridekv and --overridetensors. Vulkan gains coopmat2; YAD becomes the default Linux file picker.
new backendnew featureGitHub release
- 1.88
Adds inpainting in StableUI, Zenity/YAD file dialogs on Linux, Llama 4 support, a JSON-schema-to-GBNF endpoint and --maxrequestsize.
new featureGitHub release
- 1.87.4
Adds an embeddings endpoint (--embeddingsmodel), Qwen2.5-VL vision, automatic tool calling and a terminal chat mode (--cli). OuteTTS voice cloning and Vulkan DP4A arrive; quantized KV now works without flash attention (not recommended).
new backendnew featureGitHub release
- 1.86.2
Adds Gemma 3 text and vision, lets admin mode switch between raw GGUF files, and makes quantized KV compatible with context shifting. Hotfixes add --defaultgenamt and --nobostoken.
new featureGitHub release
- 1.85.1
Adds server-side save slots (--savedatafile), the Top-n-sigma sampler and --exportconfig / --exporttemplate.
new featureGitHub release
- 1.84.2
Mostly fixes: aria2c/wget downloads, automatic rollback when admin mode loads a broken config, Deepseek MoE override fixes; --model can take several URLs for split models.
- 1.83.1
Introduces admin mode (--admin, --admindir, --adminpassword) for runtime model/config switching by relaunch. Adds --visionmaxres and --ttsmaxlen; the notes flag known bugs.
new featureGitHub release
- 1.82.4
Adds OuteTTS text-to-speech with OpenAI- and XTTS-compatible endpoints (--ttsmodel, --ttswavtokenizer), a GGUF analyzer (--analyze) and TAESD for SD3/Flux. mmap becomes opt-in (--usemmap).
new featureGitHub release
- 1.81.1
Adds a DuckDuckGo-backed web search proxy (--websearch) and the AutoGuess chat adapter that detects the instruct template.
new featureGitHub release
- 1.80.3
Adds Qwen2-VL vision, Vulkan coopmat1, --moeexperts, --failsafe and separate draft-model GPU options. MMQ becomes the CLI default and trim_stop defaults on.
new backendnew featureGitHub release
- 1.79.1
Adds a live-synced multiplayer mode (--multiplayer), Ollama and ComfyUI API emulation, draft-model speculative decoding (--draftmodel) and GGUF text encoders for Flux/SD3.5.
new featureGitHub release
- 1.78
Adds Flux and SD 3.5 image generation with separately loaded components (--sdt5xxl, --sdclipl, --sdclipg) and --nofastforward. CLBlast is reduced to a compatibility mode.
new featureGitHub release
- 1.77
Adds token log-probabilities over the API, including /api/extra/last_logprobs, and experimental ComfyUI support in Lite.
new featureGitHub release
- 1.76
Adds anti-slop phrase banning (banned_strings) that backtracks and regenerates, and raises stop-sequence and ban-list limits.
new featureGitHub release
- 1.75.2
CLI launches now auto-pick a GPU backend; OpenBLAS is folded into the CPU option and --noblas is deprecated in favour of --usecpu. Adds RWKV v6 and --showgui.
new backendbreakingnew featureGitHub release
- 1.74
Adds the XTC sampler and --nomodel (run the UI without a local model); Horde workers gain timeouts and SSL options. macOS defaults to full GPU offload.
new featureGitHub release
- 1.73.1
First official macOS (ARM64) binaries and dual-stack IPv6. Adds one-shot --prompt mode (works with --benchmark), --promptlimit, and multiuser mode on by default.
new featureGitHub release
- 1.72
Stable Diffusion now runs on Vulkan, CUDA device ordering is fixed, and the GUI defaults GPU layers to -1 with estimates that account for image/whisper models.
new backendGitHub release
- 1.71.1
Improves automatic layer estimation (--gpulayers -1) and introduces .kcppt launch templates loadable with --config.
new featureGitHub release
- 1.70.1
Adds DRY anti-repetition sampling, experimental OpenAI tool calling and --unpack. Default context rises to 4096 and --chatcompletionsadapter accepts bundled adapter names.
new featureGitHub release
- 1.69.1
Adds bundled chat-completions adapters selectable in the GUI, multi-image separators for LLaVA, and a Vulkan tensor-split field.
new backendnew featureGitHub release
- 1.68
Adds automatic RoPE scaling for GradientAI long-context models, a Vulkan 'All GPUs' option and custom --sdclamped sizes.
new backendnew featureGitHub release
- 1.67
Integrates whisper.cpp speech-to-text (--whispermodel, /api/extra/transcribe, /v1/audio/transcriptions) and adds quantized KV cache (--quantkv).
new featureGitHub release
- 1.66
Adds image LoRAs and custom VAEs (--sdlora, --sdvae, --sdvaeauto with TAESD) and lets --model take a download URL; --sdconfig deprecated.
breakingnew featureGitHub release
- 1.65
Adds the standalone StableUI at /sdui, official CUDA 12 binaries and dedicated --sdmodel/--horde* flags that replace --sdconfig/--hordeconfig; --smartcontext deprecated in favour of context shift.
new backendbreakingnew featureGitHub release
- 1.64.1
Llama 3 tokenizer/EOT fixes, experimental --flashattention, and removal of --bantokens in favour of per-request bans.
- 1.63
Allows special tokens as stop sequences and reworks automatic RoPE scaling for Llama 3.
- 1.62.2
Adds img2img via the emulated A1111 API, chat-completions adapters (--chatcompletionsadapter), usage stats and a /docs page.
new featureGitHub release
- 1.61.2
Adds LLaVA-style vision via --mmproj, a GPT-4V-style endpoint, API-key auth (--password), Mamba on CPU and --ignoremissing.
new featureGitHub release
- 1.60.1
First local image generation via stable-diffusion.cpp with an A1111-compatible txt2img endpoint (--sdconfig); arbitrary --contextsize values and macOS --remotetunnel.
- 1.59.1
Bugfix release: CLBlast rollback, a NoAVX2 Vulkan target and --nocertify for the Horde worker.
new backendGitHub release
- 1.58
Adds a CUDA row-split toggle (layer split becomes default) and more forgiving context shifting.
new backendnew featureGitHub release
- 1.57.1
Adds --benchmark, quadratic sampling (smoothing_factor) and multi-GPU Vulkan.
new backendnew featureGitHub release
- 1.56
First Vulkan backend (--usevulkan) in Windows and Linux binaries, after a large upstream ggml backend rework; old non-GGUF models kept via a legacy library.
new backendGitHub release
- 1.55.1
Adds dynamic temperature sampling and CUDA pool-allocator changes.
new backendnew featureGitHub release
- 1.54
Adds logit_bias on OpenAI and Kobold APIs, GUI tooltips and a no-CUDA Linux binary.
new backendnew featureGitHub release
- 1.53
Adds HTTPS serving with your own cert/key (--ssl) and presence penalty.
new featureGitHub release
- 1.52.2
Adds the no-JavaScript /noscript UI and CUDA lowvram (no KV offload); Mixtral and Qwen support.
new backendnew featureGitHub release
- 1.51.1
Adds --quiet and an optional queue size for --multiuser; context shift reserves headroom.
new featureGitHub release
- 1.50.1
Better GUI layer auto-selection and extra sampler params plus SSE streaming on /v1/completions.
- 1.49
Adds memory and trim_stop API fields and --preloadstory; adds the LLAMA_PORTABLE make flag.
new featureGitHub release
- 1.48.1
Introduces ContextShift (KV-cache shifting to avoid reprocessing), GPU detection with layer suggestions, the Min-P sampler and --remotetunnel.
new featureGitHub release
- 1.47.2
Adds chat-completions instruct adapters and Horde worker resilience; Lite gains streaming and A1111 image integration.
new featureGitHub release
- 1.46.1
Removes several old flags and the old tkinter GUI, adds /v1/chat/completions and API docs, and raises GGUF context to 64k.
breakingnew featureGitHub release
- 1.45.2
Adds an OpenAI-style /v1/completions endpoint, horde worker stats and --foreground; GGUF context up to 32k.
new featureGitHub release
- 1.44.2
Adds --multiuser request queuing, grammar sampling for all architectures and --onready.
new featureGitHub release
- 1.43
Restores automatic RoPE scaling from the trained context and Windows 7 support.
- 1.42.1
Adds GGUFv2 and GGUF Falcon support and merges ROCm/hipBLAS build support.
new backendnew featureGitHub release
- 1.41
First GGUF (Llama) support, a token-count endpoint and --config for settings files.
- 1.40.1a
Mostly fixes; the Horde worker moves to daemon mode.
- 1.39.1
SSE abort fixes, experimental 2048 batch and 12k/16k context options.
- 1.38
Adds CUDA quantized matmul (mmq) as an option.
new backendnew featureGitHub release
- 1.37.1a
Adds the embedded AI Horde worker (--hordeconfig) and Llama 2 70B support.
new featureGitHub release
- 1.36
Consolidates RoPE settings into --ropeconfig (removing --linearrope).
- 1.35
CUDA 8-bit MMV speedups, linear RoPE option and GUI settings save/load.
new backendGitHub release
- 1.34.2
New customtkinter launcher GUI, NTK-aware RoPE up to 8K context and token banning.
new featureGitHub release
- 1.33
'Ultimate' build combining CUDA and CLBlast/OpenBLAS; scaled RoPE beyond 2048.
new backendGitHub release
- 1.32.3
K-quant OpenCL kernels and GPU offload for MPT, GPT-2, GPT-J and NeoX.
- 1.31.2
Bugfixes, token-rate console output, RWKV sequence mode disabled.
- 1.30.3
CLBlast K-quants, RWKV World tokenizer, streaming/abort endpoints and LoRA base model.
- 1.29
BLAS batch size in the launcher and K-quants on OpenBLAS.
- 1.28
MPT support; --hordeconfig replaces --renamemodel.
- 1.27
CLBlast full-VRAM layers and RWKV/OpenLlama support.
- 1.26b
Token-probability debug output, Top-A sampler and Starcoder.
- 1.25.1
Failsafe mode for very old CPUs and CLBlast 1.6.
- 1.24
GGJT v3 quantization formats with backward compatibility.
- 1.23.1
Quantization for GPT-2/GPT-J/NeoX and first CLBlast GPU layer offload (--gpulayers).
- 1.21.3
New quantization formats with backward compatibility.
new featureGitHub release
- 1.20
Adds --contextsize for contexts beyond 2048 and RedPajama.
new featureGitHub release
- 1.19
Mirostat, --forceversion and --blasthreads.
- 1.18
Lite group chat, --highpriority and --usemlock.
- 1.17
Removes Cloudflare Insights; EOS unbanning in the UI.
breakingGitHub release
- 1.16
Sampler rework adding Tail-Free and Typical sampling.
- 1.15
First 'Easy Mode' launcher GUI (--skiplauncher to bypass) and q5 formats.
- 1.14
Older NeoX quant compatibility and OOM fixes.
- 1.13.1
CLBlast threading fix and --unbantokens.
- 1.12
OOM fixes for GPT-2/GPT-J/NeoX with BLAS.
- 1.11
GPT-NeoX/Pythia/StableLM support and LoRA for llama (--lora).
- 1.10
RWKV support without PyTorch and a version endpoint.
- 1.9
Stop sequences in the API; BLAS for GPT-J/GPT-2.
- 1.8.1
CLBlast GPU dequantization speedup.
- 1.7
Adds SmartContext (--smartcontext).
new featureGitHub release
- 1.6
Crash fixes and batch size back to 512.
- 1.5
Merged AVX2/non-AVX2 Windows binary (--noavx2) and experimental CLBlast (--useclblast).
- 1.4
mmap becomes the default.
- 1.3
Endpoint fixes and older/newer GPT-J/GPT-2 auto-detection.
- 1.2
Newer GPT-2 (Cerebras) support.
- 1.1
Simplified version numbering and fixes.
- 1.0.9beta
GPT-2 support and Alpaca-style instruct mode in Lite.
- 1.0.8beta
Project renamed from llamacpp-for-kobold to KoboldCpp; GPT-J and GPT4All formats.
breakingflag renamedGitHub release
- 1.0.7
ggjt model format support (still named llamacpp-for-kobold).
- 1.0.6beta
OpenBLAS prompt processing, argparse flags and --noblas.
- 1.0.5
Upstream 65B fixes and a soft-prompts endpoint.
- 1.0.4
PyInstaller standalone build script and prompt token caching.
- 1.0.3
Upstream refactor and client-specified context lengths.
- 1.0.2
Embeds Kobold Lite and switches the licence to AGPL v3.
- 1.0.1
macOS fixes and KV caching for continuations.
- 1.0.0
Initial release as llamacpp-for-kobold: a Python script serving a llama.cpp model to KoboldAI clients.
Changes that break old guides
Old guides still rank high, and these changes make them wrong. Each row links to the flag in the flag reference.
| Default context size: 2048, then 4096, 8192, 16384, 12288 and 16384 again | 1.20, 1.70, 1.96, 1.115, 1.115.2, 1.122 | Check the value with --contextsize. See context size. |
| GPU layers default changed from 0 to -1 (automatic) | 1.75 | Leave --gpulayers alone: GPU layers. |
| Memory-mapping is off by default | 1.82 | --usemmap turns it on. |
| Flash attention is on by default (and the CLBlast backend was removed) | 1.107 | --noflashattention turns it off. Use Vulkan instead of CLBlast. |
| <code>--usecublas</code> renamed <code>--usecuda</code> (old name kept as an alias) | 1.96 | --usecuda |
| <code>--blasbatchsize</code> renamed <code>--batchsize</code> (old name kept) | 1.101 | --batchsize |
| <code>--multiuser</code> is always on | 1.123 | Nothing to set. The queue length is --queuelimit. |
| Row split removed; multi-GPU modes moved to <code>--splitmode</code> | 1.113, 1.118 | --splitmode |
| <code>--quantkv</code> takes names (q8_0, q4_0) instead of numbers (numbers still accepted) | 1.112 | --quantkv |
The full list of renamed and removed flags, with their aliases, is in the flag reference.
How often KoboldCpp releases
Data
| Item | Value (releases) |
|---|---|
| 2023 Q1 | 7 |
| 2023 Q2 | 35 |
| 2023 Q3 | 11 |
| 2023 Q4 | 9 |
| 2024 Q1 | 8 |
| 2024 Q2 | 7 |
| 2024 Q3 | 7 |
| 2024 Q4 | 5 |
| 2025 Q1 | 6 |
| 2025 Q2 | 9 |
| 2025 Q3 | 4 |
| 2025 Q4 | 5 |
| 2026 Q1 | 6 |
| 2026 Q2 | 6 |
| 2026 Q3 | 6 |
| 2026 Q4 | 2 |
Download CSV · Results: F01
Releases came every few days in 2023, then settled to about every two weeks: the median gap between releases with a GitHub page was 3 days in 2023, 13 in 2024, 16 in 2025 and 15 in 2026.
Should you update?
- A new model type you need. Yes: models for a new architecture only load in a version that supports it (model will not load).
- The latest version has a reported regression. Wait for the hotfix, which usually follows within a few days, or stay on the version that works.
- A stable setup that other people use. Pin a version and keep the old file. Update on purpose, not by habit.
- To go back, download the older file from GitHub: older versions.
After any update, run the same prompt with Runs a benchmark instead of starting the server; optionally appends results to a CSV file.Default noneLauncher Hardware tab: Run Benchmark (button)Since v1.57All details before and after if speed matters (benchmark).
Rolling builds
Next to the numbered releases the project publishes a rolling pre-release that always holds the newest code, and an experimental Linux ROCm build under rocm-rolling (AMD). They get the latest fixes and the least testing, and they are not tied to a version number. Use them to check whether a bug you hit is already fixed, not for a setup you rely on. We have not tested the rolling builds.
Frequently asked questions
What is the latest version of KoboldCpp?
1.124.2, released on 2 November 2026. This page and the download page are rebuilt from the project's release list, so they show the newest version we know about.
How often does KoboldCpp update?
Feature releases come about every two weeks in 2025 and 2026 (the median gap between releases with a GitHub page was 16 days in 2025 and 15 in 2026), often followed by a hotfix within days. In 2023 releases came every few days.
How do I go back to an older version?
Every release stays on the project's GitHub releases page. Download the older file from there and keep your current one next to it. See the download page for the steps.
What changed in the default context size?
It was 2048 (1.20), then 4096 (1.70), 8192 (1.96), 16384 (1.115), 12288 (1.115.2) and 16384 again (1.122). That is why older guides give different numbers.
What is the rolling build?
A pre-release that the project updates continuously with the newest code, next to the numbered releases. It has the latest fixes and the least testing. There is also an experimental Linux ROCm build under rocm-rolling.