Connect Local Qwen 3.8 27B to DSH
Beginner guide to connecting local Qwen 3.8 27B to DSH: GPU memory table, LM Studio / Ollama / llama.cpp port setup, exact 5-field config, and error fixes.
On this page

Want to use Alibaba's Qwen 3.8 27B inside DeepSeek Harness (DSH) for coding and image reading? Before downloading 15+ gigabytes of files, make sure you pick the right path for your computer:
- Path A — Cloud API (no heavy GPU needed, 2-minute setup): Using an everyday laptop without a high-memory graphics card? Skip local installation completely and connect via Alibaba's cloud API using our Qwen / DashScope cloud setup guide.
- Path B — Run 100% locally offline for free (requires a strong GPU or high-memory Mac): Have an RTX 3090/4090 or a Mac with 32GB+ unified memory and want zero API bills with full local privacy? Follow the beginner walkthrough below.
One core concept makes everything click: DSH does not load multi-gigabyte model files by itself. Instead, you run a free local model app on your computer (such as beginner-friendly LM Studio or Ollama, or llama.cpp), and DSH talks to that app over a local port on your machine.
Step 0: Check whether your computer can run 27B locally
27B means 27 billion parameters. Based on the official Qwen 3.8 27B model card, check your hardware before downloading:
| Your Computer Hardware | Can It Run Qwen 3.8 27B Locally? | Plain Beginner Advice |
|---|---|---|
| Everyday laptop or 6GB-12GB GPU (RTX 3060 / 4060 8GB) | No (will freeze or run out of memory) | Pick a smaller 7B/8B local model, or use the cloud API link above |
| 16GB-20GB GPU (RTX 4060 Ti 16GB / RTX 4080) | Tight fit with 4-bit compression (Q4_K_M) | Keep context length modest (8192 to 16384) so VRAM does not overflow |
| 24GB+ GPU (RTX 3090 / 4090) or 32GB+ Apple Silicon Mac | Runs smoothly at Q4_K_M or Q5_K_M | Great match for offline coding, file reading, and local image checks |
4-step beginner walkthrough: connect local Qwen to DSH
Follow these four steps in order from zero to a working chat.
Step 1: Start your local model server app
Pick whichever local app you prefer to load Qwen 3.8 27B:
- Easiest visual app — LM Studio (no terminal commands): Search and download a
Qwen3.8-27BQ4_K_MGGUF inside LM Studio, open theDevelopertab on the left, load the model, and toggle Start Server on (default local URL ishttp://127.0.0.1:1234/v1; see our LM Studio visual guide). - Ollama: Pull and start the model in Ollama, which serves on
http://127.0.0.1:11434/v1by default (see our Ollama local guide). llama.cppserver: Launchllama-serverfollowing the official llama.cpp server docs (include--mmprojif you want image reading), which defaults tohttp://127.0.0.1:8080/v1.
Step 2: Copy the exact model ID in 10 seconds (avoids 90% of beginner errors)
Beginners often guess a model name like qwen3.8 and get an immediate Model not found error because the local app registered a longer filename.
With your local server running from Step 1, open this address right in your web browser (change 8080 to 1234 for LM Studio, or 11434 for Ollama):
http://127.0.0.1:8080/v1/models(Or run curl http://127.0.0.1:8080/v1/models in your terminal.)
Look for "id": "..." in the short text response and copy the exact string inside the quotes.
Step 3: Fill in the 5 fields inside DSH
Open DSH on your computer and click Settings -> Models -> Add a custom provider (full field reference in the official DSH provider docs). Enter these values:
| Field in DSH | What to Type | Why |
|---|---|---|
Provider ID | qwen-local | Short lowercase ID for your local server |
Display name | Local Qwen 3.8 27B | Friendly label shown in your model dropdown |
Base URL | http://127.0.0.1:8080/v1 | Use :1234/v1 for LM Studio or :11434/v1 for Ollama; keep /v1 at the end |
API (protocol) | openai-completions | Local servers expose an OpenAI-compatible Chat Completions endpoint |
Models | Paste the exact "id" from Step 2 | Must match /v1/models character for character |
API Key | local-key | Local apps do not charge money, yet the DSH adapter requires a non-empty placeholder string (never paste a paid cloud key here) |
Click Save.
Step 4: Test in 3 small stages (text -> folder read -> image)
Testing one capability at a time makes troubleshooting painless:
- Plain text first: Create an empty
dsh-testfolder on your Desktop, select it viaChoose workspace, pickLocal Qwen 3.8 27B, and send"Reply in one short sentence: what is 2 + 2?" - Read-only folder tool: Send
"List the files in this workspace without creating or editing any files"(more first-run tips in our Web UI beginner guide). Confirm you see an actual tool-call card run, not just a guessed text answer. - Image check (optional): Attach a simple screenshot. If text works fine yet image replies fail or hallucinate, check two things: your local engine must load the vision projector file (
mmproj), and the custom model entry in DSH must declare image input support (documented in GitHub Discussion #2329).
Quick troubleshooting table
| Error You See | Root Cause | Quick Fix |
|---|---|---|
ECONNREFUSED (connection refused) | Local server app is stopped or port number is wrong | Confirm LM Studio / Ollama / llama.cpp server is running and check port 1234, 11434, or 8080 |
404 or Model not found | Missing /v1 at the end of Base URL or mistyped model ID | End Base URL with /v1 and copy the exact "id" from /v1/models |
| Computer freezes or runs out of VRAM mid-chat | Context window set too high (such as 128K) for your GPU | Lower context length in LM Studio / Ollama / llama.cpp to 8192 or 16384 and restart the server |
Common questions
Why does DSH ask for an API key when the model runs on my own computer?
DSH uses one shared form for cloud and local OpenAI-compatible servers and checks that the credential field is not blank. Typing any placeholder word like local-key or ollama satisfies the check without sending data anywhere outside 127.0.0.1.
Should I set the context window to 128K like some online screenshots show?
Not on consumer GPUs. A model supporting 128K on paper still needs extra gigabytes of VRAM as context grows. Start at 8192 or 16384 first so responses stay fast and stable.
How should I test coding quality before using it on a real repo?
Run two small bug-fix tasks on a copy of a project using our coding benchmark template and verify the test suite passes.
Model and workflow guides
Gemini 4 Neon is here. When is it worth switching models?
Compare Gemini 4 Neon in two coding setups, work through a $0.45 API example, and find Neon in Gemini, Google AI Studio, and Antigravity.
Read →Models and costsGPT-6.2 Astra & Sol Guide: Specs, Pricing & Choice
Beginner guide to OpenAI's GPT-6.2 Astra and Sol: Claude vs DeepSeek price table, plain-English benchmarks, 272K cliff, and step-by-step DSH & Python setup.
Read →Models and costsGPT-6.1 Terra Guide: Uses, Pricing and Sol Comparison
Beginner guide to OpenAI's GPT-6.1 Terra: half-price visual, computer-use, and writing model compared with Sol and DeepSeek, plus DSH and Python setup steps.
Read →Model setup and costsCodex Costs: API Billing vs Subscription Usage
Beginner guide to OpenAI Codex costs: ChatGPT Plus login vs API-key billing table, real per-prompt dollar math, and 3 steps to lock a monthly budget cap.
Read →Model setup and costsConnect Grok API to DSH: Setup & Tools
Step-by-step beginner guide to connecting xAI's Grok API (grok-4.7) to DSH: Grok vs Groq comparison, exact 5-field custom provider setup, and 401/404 fixes.
Read →Model setup and costsChoose a Gemini Model by Task and Cost
Beginner guide to Google Gemini model selection: compare 3.8 Flash, 4 Argon, and DeepSeek prices, watch 2027 rate changes, and set up DSH in 4 simple steps.
Read →Model benchmarks and engineeringGemini 4 Argon Coding Benchmarks & 1M Window
Beginner guide to Google's Gemini 4 Argon coding scores vs Claude and GPT, how its shared 1M window and 65K output cap work, plus 4-step DSH setup.
Read →API costs and enterprise ROIGemini 4 Argon API Pricing, Cache & ROI Guide
Beginner guide to Gemini 4 Argon API pricing: Dec 2026 2x rate jump, how 95% auto-caching and thinking tokens bill, DeepSeek cost comparison, and calculator.
Read →Model setup and costsClaude Sonnet vs Opus: Which Should You Use?
Beginner guide to Claude Sonnet 5.5 vs Opus 5.5: side-by-side API prices, per-prompt cost math, clear rules on when to upgrade, and 4-step DSH setup.
Read →Models and costsOpus 5.5 vs Fable 5.2: Which Model Is Right for You?
Beginner guide to Claude Opus 5.5 vs Fable 5.2: side-by-side API prices, coding benchmarks, per-task cost math, 30-day retention rules, and 4-step DSH setup.
Read →Content creationHow to draft a video script with Media Desk for DeepSeek Harness
Turn a phone note into a video script, cover, and storyboard with Media Desk for DeepSeek Harness. See how to install it, set up images, and export a draft.
Read →Mobile and remote accessUse DeepSeek Harness on Your Phone: Beginner Guide
Zero-jargon beginner guide to using DeepSeek Harness on your phone: find your computer IP, connect over home Wi-Fi in 2 minutes, or use Tailscale on 4G/5G.
Read →Model interfaces and workflowsMiniMax H3 vs M-Series in DSH: Which Works?
Beginner guide to MiniMax H3 vs M3 in DSH: why H3 is for video while M3 handles coding, how to avoid region 401 errors, and 4-step DSH setup instructions.
Read →API credentials and troubleshootingKimi API Key Errors: Region, Model and Permissions
Beginner fix guide for Kimi / Moonshot API key errors: match .ai vs .cn URLs, avoid retired model names and K3 Max typos, decode 401/404/429, and set up DSH.
Read →Image editing and toolsAI Image Editing: Prompts, Masks and Review
Beginner guide to AI image editing without ruining your subject: copy our 4-line prompt template, avoid PNG mask traps, inspect edges, and use DSH safely.
Read →