기본 콘텐츠로 건너뛰기

Ollama Not Working? Fixes for Common Errors and GPU Issues

Quick summary: Most "Ollama not working" problems come down to five causes: the background server isn't running, the model name doesn't match what you downloaded, the GPU isn't being detected so everything runs slowly on the CPU, the model doesn't fit in memory, or a network setting is blocking downloads or connections. This guide walks through each one with the exact messages Ollama prints, the command that confirms the cause, and the fix from Ollama's own documentation.

What's in this guide

  1. Step 0: read the server log first
  2. "Could not connect to ollama server"
  3. "Model not found, try pulling it first"
  4. Ollama is slow: the GPU isn't being used
  5. Out of memory, crashes, or a sluggish PC
  6. Downloads fail or other apps can't connect
  7. Windows-specific glitches
  8. Ollama vs LM Studio vs Jan: when switching makes sense
  9. Editor's note, FAQ, and related guides

Step 0: Read the server log first

Ollama is really two pieces: a server that loads models and answers requests on your machine, and the ollama command or desktop app that talks to it. When something breaks, the terminal often shows only a short error, while the server log explains why. Ollama's troubleshooting page lists where that log lives on each system:

  • macOS: run cat ~/.ollama/logs/server.log
  • Linux (systemd): run journalctl -u ollama --no-pager --follow --pager-end
  • Docker: run docker logs <container-name> (find the name with docker ps)
  • Windows: press Win+R and open explorer %LOCALAPPDATA%\Ollama. The newest log is server.log, older ones are numbered server-#.log files, and app.log covers the tray app.

If the log is too quiet, turn on debug output. On Windows, quit Ollama from the tray icon first, then in PowerShell run $env:OLLAMA_DEBUG="1" followed by & "ollama app.exe". On macOS and Linux, set OLLAMA_DEBUG=1 as an environment variable for the server. Keep the log open in a second window while you work through the fixes below.

Fix 1: "Could not connect to ollama server"

On Linux, if you type ollama run gemma4 while no server is running, the CLI stops with could not connect to ollama server, run 'ollama serve' to start it. Running ollama -v in the same situation prints Warning: could not connect to a running Ollama instance instead of a server version. Both mean the same thing: nothing is answering on the default address, 127.0.0.1:11434.

How to fix it

  1. Linux with the install script: start the service with sudo systemctl start ollama, then check it with sudo systemctl status ollama. If the status shows it crashing, read the journalctl log from Step 0.
  2. Linux manual install or quick test: run ollama serve in one terminal and use a second terminal for ollama run. Closing the first terminal stops the server.
  3. macOS and Windows: the desktop app runs the server in the background. Open Ollama from Applications or the Start menu. In those versions, the CLI tries to launch the app for you when it can't reach the server, so a repeated failure usually means the app itself is failing to start. Check app.log.
  4. You changed OLLAMA_HOST: the CLI and the server must agree on the address. If you moved the server to another port or IP, set the same OLLAMA_HOST value in the shell where you run the CLI.

One more Linux trap: if your system mounts its temp folder with the noexec flag, Ollama can't run its helper files from there. The docs suggest pointing OLLAMA_TMPDIR at a folder the Ollama user can write to, such as /usr/share/ollama/.

Fix 2: "Model not found, try pulling it first"

When an API call or app asks for a model you don't have, the server replies with HTTP 404 and the message model "name" not found, try pulling it first. This often happens with apps that connect to Ollama, because they ask for a model by a name you never downloaded, or by a slightly different tag.

How to fix it

  1. Run ollama ls to see the exact names and tags already on your machine.
  2. Download the missing one with ollama pull gemma4 (use the name your app expects).
  3. In the connecting app, pick the model from its list rather than typing the name, or copy the name from ollama ls character for character, including the tag after the colon.

Ollama's API error page also lists the other codes you may see: 400 for a malformed request, 429 when a rate limit is hit, 500 for a server error, and 502 when a cloud model can't be reached. If an error happens partway through a streamed reply, the status code stays 200 and the error arrives as a final JSON line with an error field, so check the end of the stream, not just the status code.

Fix 3: Ollama is slow because the GPU isn't used

The fastest check is ollama ps while a model is loaded. The PROCESSOR column tells you where it is running: 100% GPU is what you want, 100% CPU means the GPU was not used, and a split such as 48%/52% CPU/GPU means the model only partly fit in video memory (VRAM).

NVIDIA cards

  • Ollama's hardware page requires compute capability 5.0 or higher and driver 550 or newer. Older cards with compute capability 5.0 to 6.2 (for example the GTX 10-series) need driver 570 or newer. On Windows, the stated minimum is driver 551.61.
  • On Linux, if the GPU vanishes after sleep (suspend and resume), reload the driver module with sudo rmmod nvidia_uvm && sudo modprobe nvidia_uvm, or reboot.
  • Server logs may show GPU initialization error codes such as 3 (not initialized), 46 (device unavailable), 100 (no device), or 999 (unknown). The documented steps are: run sudo nvidia-modprobe -u, reload nvidia_uvm, reboot, and update the driver.
  • In Docker, first confirm the container runtime can see the card with docker run --gpus all ubuntu nvidia-smi. If Ollama starts on the GPU and later drops to the CPU with discovery errors, the docs fix is adding "exec-opts": ["native.cgroupdriver=cgroupfs"] to /etc/docker/daemon.json.

AMD cards

  • Ollama now bundles ROCm 7 libraries. On Linux, an older ROCm 6.x driver can make GPU discovery hang. The log then shows failure during GPU discovery ... failed to finish discovery before timeout, and Ollama falls back to the CPU. Upgrade to the ROCm v7 driver with AMD's amdgpu-install tool, reboot, and restart Ollama.
  • On Linux, your user also needs access to /dev/kfd, which usually means membership in the video and/or render group.
  • On Windows, the ROCm list is limited to RX 7000-series and Radeon PRO W7000 cards. Some RX 6000-series systems don't get ROCm v7 on current drivers. The docs say Vulkan is on by default and is the recommended fallback for them.
  • For cards ROCm doesn't officially support on Linux, HSA_OVERRIDE_GFX_VERSION can make Ollama treat the card as a close supported target. The docs' example is 10.3.0 for an RX 5400. This is experimental, so expect possible instability.

Fix 4: Out of memory, crashes, or a sluggish computer

A model that barely fits will load partly onto the CPU, which slows it down sharply, or fail to load at all. Before you buy hardware, try these levers from Ollama's FAQ:

  • Context length. The default context window is 4,096 tokens. Raising it (with OLLAMA_CONTEXT_LENGTH, /set parameter num_ctx in a chat, or num_ctx in the API) uses more memory. If you raised it and things got slow, lower it again.
  • Parallel requests. Memory needed grows with OLLAMA_NUM_PARALLEL times the context length. The default is 1. If you raised it for an app, that may explain the jump.
  • K/V cache quantization. When Flash Attention is active, setting OLLAMA_KV_CACHE_TYPE=q8_0 uses roughly half the cache memory of the default f16, with a very small precision loss according to the docs. q4_0 saves more but costs more quality.
  • Models staying loaded. By default, a model stays in memory for 5 minutes after its last use. Free it right away with ollama stop gemma4, or change the timeout with OLLAMA_KEEP_ALIVE or the API's keep_alive setting (0 unloads right after the reply).
  • Too many requests. If you get a 503 "server overloaded" error, the request queue is full. Its size is OLLAMA_MAX_QUEUE, 512 by default.

If crashes started right after a GPU driver change, you can force CPU mode as a test. On NVIDIA, set CUDA_VISIBLE_DEVICES=-1. More generally, OLLAMA_LLM_LIBRARY="cpu_avx2" skips GPU auto-detection. If the crash disappears, the GPU stack is the problem.

Fix 5: Downloads fail or other apps can't connect

  • Behind a corporate proxy: set HTTPS_PROXY and install the proxy's certificate as a system certificate. The docs specifically warn against setting HTTP_PROXY: model pulls only use HTTPS, and an HTTP proxy setting can break connections between your apps and the server.
  • Slow downloads inside WSL2 on Windows 10: in the vEthernet (WSL) network adapter's advanced properties, disable both "Large Send Offload Version 2" entries (IPv4 and IPv6).
  • Another device on your network can't reach Ollama: by default it only listens on 127.0.0.1, which is this computer only. Set OLLAMA_HOST=0.0.0.0:11434 to listen on the network, and only do this on a network you trust.
  • A browser extension or web app gets blocked: add its origin to OLLAMA_ORIGINS, for example chrome-extension://*.
  • The disk is full: models can take tens to hundreds of GB. Move them with OLLAMA_MODELS. On Linux, give the ollama user ownership of the new folder (sudo chown -R ollama:ollama <directory>), or the server can't write to it.

Where you set these variables matters. On Linux, add Environment= lines with systemctl edit ollama.service, then run systemctl daemon-reload and systemctl restart ollama. On macOS, use launchctl setenv and restart the app. On Windows, quit Ollama from the taskbar, edit your account's environment variables in Settings, and relaunch from the Start menu. A variable typed into a random terminal won't reach the background app.

Fix 6: Windows-specific glitches

  • Strange text like ←[?25h←[?25l or "The parameter is incorrect": older Windows 10 builds don't display control characters correctly. Update Windows. Ollama's stated requirement is Windows 10 22H2 or newer.
  • Squares instead of progress bars: some older terminal fonts can't show the Unicode characters Ollama uses for progress. Switch to a different terminal font.
  • Ollama starts every time you log in: turn it off in Task Manager under Startup apps. On macOS, use Login Items in System Settings.

Ollama vs LM Studio vs Jan: when switching makes sense

If Ollama keeps fighting your hardware, a different local app won't magically add VRAM, but a graphical app can make model choice and settings easier to see. Here is how the three compare, based on each official site checked on 2026-10-24:

  Ollama LM Studio Jan
Local use Free plan: run models locally, starter cloud credits included Free plan: local LLMs via llama.cpp and MLX, offline voice transcription Free download; local models are the default
Paid options Pro $20/month or $200/year ($60 usage credits per month); Team $500/month Bionic+ $20/month (US-hosted open models); Pro $100/month Cloud models need your own provider API key (e.g. OpenAI, Anthropic)
Main interface CLI + local API on port 11434, desktop app on macOS/Windows Desktop app, plus the lms CLI Desktop app with a model Hub and local API server
Platforms macOS 14+, Windows 10 22H2+, Linux Apple Silicon Macs (macOS 14+), Windows x64 with AVX2 or ARM, Linux AppImage macOS, Windows, Linux
Open source Yes (MIT license) Not advertised as open source on lmstudio.ai Yes (Apache 2.0)
Helps with hardware fit ollama ps shows CPU/GPU split Docs recommend 16GB+ RAM, 4GB+ VRAM on Windows Hub shows whether each model fits your hardware

In practice, keep Ollama if other tools (coding assistants, scripts, browser extensions) need its API. Try Jan if you mainly want a chat window and like being told up front which models fit your computer. LM Studio suits people who want a polished app on Apple Silicon or Windows. Note that it doesn't support Intel Macs, and on x64 PCs it needs a CPU with AVX2.

Editor's note

Who this is for: anyone running Ollama on their own Mac, Windows PC, or Linux box, including people who installed it only so another app (a coding assistant, a notes app, a browser extension) could use local models.

Who it's not for: if you only use Ollama's cloud models, local GPU and memory fixes won't apply. Check your plan and the 429/502 errors instead.

Gotchas I'd flag:

  • Environment variables set in a terminal don't reach the background app on macOS or Windows. Set them the way the docs describe, then fully quit and reopen Ollama.
  • A model that runs "fine but slowly" usually doesn't fit in VRAM. Check ollama ps before blaming drivers.
  • Raising the context length is the most common self-inflicted slowdown.
  • Setting OLLAMA_HOST=0.0.0.0 exposes the API to your network. Only do this on a network you trust, never on public Wi-Fi.

Quick error-to-fix map: "could not connect" → start the service or app. "not found, try pulling it first" → ollama pull with the exact name. 100% CPU → driver or ROCm version. 503 → too many requests queued (OLLAMA_MAX_QUEUE). Garbled Windows terminal → update Windows 10.

FAQ

How do I update Ollama to rule out an old bug?

On macOS and Windows, Ollama downloads updates by itself. Click the menu bar or taskbar icon and choose "Restart to update". On Linux, re-run the install script: curl -fsSL https://ollama.com/install.sh | sh. If a new release breaks something on Linux, the same script accepts an OLLAMA_VERSION value to install a specific older version.

Does Ollama send my prompts to the internet?

According to Ollama's FAQ, prompts to local models stay on your machine. Cloud models are processed on Ollama's servers, and the company says it doesn't store or train on that content. To go local-only, set OLLAMA_NO_CLOUD=1 or add "disable_ollama_cloud": true to ~/.ollama/server.json, then restart. The log will show Ollama cloud disabled: true.

My GPU is in the list, but ollama ps still shows CPU. What next?

First confirm the model actually fits: a split like 48%/52% means VRAM ran out, not that detection failed. If it shows 100% CPU even with a small model, update the driver (NVIDIA 550+ or 570+ for older cards, AMD ROCm v7), restart Ollama, and look for GPU discovery messages in the server log with OLLAMA_DEBUG=1 turned on.

Related guides on lightmind24:

Start with the server log and ollama ps. Between them, they explain most Ollama failures in under a minute.

Images: screenshots of Ollama's official documentation pages (docs.ollama.com: Troubleshooting, FAQ, Hardware support, Windows), captured 2026-10-24. See the official screenshots at the end of this post.

Last verified: 2026-10-24
Sources:
Ollama docs: Troubleshooting
Ollama docs: FAQ
Ollama docs: Hardware support
Ollama docs: Windows
Ollama docs: Linux
Ollama docs: macOS
Ollama docs: API errors
Ollama docs: CLI reference
Ollama on GitHub (source code and license)
Ollama pricing
LM Studio pricing
LM Studio system requirements
Jan Desktop quickstart
Jan on GitHub (license)

Uploaded Image Uploaded Image Uploaded Image Uploaded Image

댓글

이 블로그의 인기 게시물

브루(Vrew) 요금제 정리: 무료 200크레딧으로 어디까지 쓸 수 있나

요약: 영상 자막 편집기 브루(Vrew)는 2026년 4월 22일부터 기능별 사용량 대신 '통합 크레딧' 하나로 요금을 매겨요. 무료 플랜은 매달 200크레딧이 들어오는데, 음성 분석(자동 자막)만 쓰면 최대 200분 분량이라 짧은 영상 몇 편은 충분하지만, 영상 초반의 Vrew 마크를 지우려면 유료 플랜이 필요해요. 이 글은 브루 공식 가격 페이지와 고객센터 공지를 2026년 10월 23일에 다시 확인해, 크레딧 계산법과 플랜 고르는 기준을 정리했어요. 목차 브루 요금 구조: 통합 크레딧이란 1크레딧으로 할 수 있는 일 무료 플랜으로 되는 것과 안 되는 것 유료 플랜 4종 비교표 월 구독·연 구독·기간 이용권, 무엇이 이득일까 상황별 추천: 학생·직장인·유튜버 크레딧이 모자랄 때: 사용 기록 지우개 업그레이드·환불 전에 알아둘 규칙 운영자 메모, 자주 묻는 질문 1. 브루 요금 구조: 통합 크레딧이란 예전 브루는 플랜마다 음성 분석 몇 분, AI 목소리 몇 자, 번역 몇 자처럼 기능별 한도를 따로 줬어요. 그런데 보이저엑스(브루 운영사)가 실제 이용 데이터를 보니 한두 가지 기능만 집중해서 쓰는 사람이 훨씬 많았다고 해요. 그래서 2026년 4월 22일 업데이트부터는 플랜마다 정해진 크레딧을 받고, 어떤 기능에 쓸지는 내가 정하는 방식으로 바뀌었어요. 플랜별 매달 지급량은 Free 200 · Light 2,000 · Standard 10,000 · Business 12,000크레딧 이에요. 공식 공지에 따르면 무료 사용자는 매달 1일에, 유료 사용자는 매월 결제일에 충전되고, 남은 크레딧은 다음 달로 넘어가지 않아요. 바뀐 뒤 무엇이 좋아졌나 공지에 나온 예시가 이해하기 쉬워요. Standard 플랜에서 AI 목소리만 쓰면 예전에는 최대 50만 자였지만, 지금은 1만 크레딧을 모두 AI 목소리에 몰아 쓸 수 있어 최대 100만 자가 돼요. 반대로 여러 기능을 골고루 쓰는 사람은 크레딧이 기능끼리 ...

에이닷 전화 vs 익시오: 통화 녹음·AI 요약, 내 통신사·폰에 맞는 앱 고르기

요약: 통화 녹음과 AI 요약을 쓰고 싶다면 먼저 내 통신사와 휴대폰 종류부터 확인해야 해요. LG유플러스 고객은 무료 앱 익시오, SKT 고객은 에이닷 전화가 가장 기능이 많고, 다른 통신사의 아이폰 사용자는 iOS에 들어 있는 기본 통화 녹음·전사 기능이 현실적인 선택이에요. 이 글은 2026-10-30에 SKT·LG유플러스·Apple 공식 페이지와 도움말을 직접 확인해 세 가지를 비교했어요. 왜 '앱 기능'보다 '내 조건'을 먼저 봐야 할까요 통화 요약 앱 비교 글을 보면 기능 목록부터 나열하는 경우가 많아요. 그런데 실제로 써 보려는 사람에게 가장 먼저 걸리는 건 기능이 아니라 이용 조건 이에요. 익시오는 LG유플러스 고객 전용이고, 에이닷 전화는 누구나 설치할 수 있지만 통역콜·baro 같은 기능은 SKT 고객에게만 열려 있어요. 아이폰은 운영체제 구조상 앱이 통화를 마음대로 녹음하기 어려워서, 통신사 앱이 안 되면 iOS 기본 기능에 기대야 해요. 그래서 이 글은 "어떤 앱이 제일 좋은가"보다 "내 통신사·내 폰에서 무엇을 쓸 수 있나" 순서로 정리했어요. 기능 차이는 그다음이에요. 한눈에 비교: 에이닷 전화 vs 익시오 vs 아이폰 기본 통화 녹음 구분 에이닷 전화 (SKT) 익시오 ixi-O (LG유플러스) 아이폰 기본 통화 녹음 (Apple) 누가 쓸 수 있나 통신사 구분 없이 앱 설치 가능. 단, 통역콜·baro·그룹통화 발신 등은 SKT 고객 전용 U+ 모바일 고객. 타 통신사, 만 14세 미만, 선불요금제 고객은 제한 (공식 사이트 FAQ상 알뜰폰 중 KB리브엠은 예외) 지원 국가·언어의 iPhone 사용자. 1명과 통화할 때만 녹음 가능 지원 OS·기기 Android 10 이상, iOS 16 이상 iOS 17 이상 iPhone 12 이상 / Android...

타입캐스트 vs 클로바더빙: 수익 나는 채널에 쓸 때 무료·유료 조건 비교

요약 타입캐스트와 클로바더빙은 둘 다 글만 입력하면 한국어 AI 목소리를 만들어 주는 서비스지만, 무료로 쓸 수 있는 범위가 꽤 다릅니다. 클로바더빙 무료는 매달 15,000자까지 쓸 수 있는 대신 수익이 생기는 채널에는 쓸 수 없고, 타입캐스트 무료는 처음에 주는 약 5분 분량의 다운로드로 끝나는 체험판에 가깝습니다. 광고 수익이 붙은 유튜브나 쇼핑몰 영상에 넣을 목소리라면 처음부터 유료 플랜 조건을 보고 고르는 편이 안전합니다. 왜 이 두 서비스를 같이 비교하나 쇼츠 내레이션, 수업 자료 더빙, 상품 소개 영상처럼 "내 목소리 대신 AI 목소리"가 필요할 때 한국 사용자들이 가장 많이 찾는 곳이 네오사피엔스의 타입캐스트 와 네이버의 클로바더빙 입니다. 기능 소개만 보면 비슷해 보이지만, 실제로 문제가 되는 건 음질보다 "이 목소리를 내 채널에 써도 되나" 입니다. 그래서 이 글은 기능 나열보다 무료 사용 범위, 출처 표기 의무, 유료 플랜 가격과 다운로드 한도를 중심으로 정리했습니다. 모든 수치는 2026년 10월 25일에 두 서비스의 공식 요금 페이지와 고객센터 도움말에서 직접 다시 확인한 내용입니다. 한눈에 보는 비교표 항목 타입캐스트 무료 타입캐스트 베이직 클로바더빙 Free 클로바더빙 Standard 월 요금 0원 9,900원 (연간 결제 시 10% 할인) 0원 19,900원 (네이버 클라우드 플랫폼에서 결제) 무료·포함 사용량 음성 생성·재생 무제한, 다운로드는 최초 3,000크레딧(약 5분) 1회 지급 월 30,000크레딧(약 60분) 월 15,000자, 다운로드 월 20회, 프로젝트 누적 5개 (매월 1일 갱신) 월 30,000자, 다운로드 월 40회, 프로젝트 누적 20개 쓸 수 있...