Known errors (chapter 1)

Start and connect

What you see
curl: (7) Failed to connect to localhost port 11434 after 0 ms: Couldn't connect to server
Cause
Nothing is running at port 11434.
Fix
Open the Ollama app, or run ollama serve.
Status
Seen on our Mac (curl 8.7.1).
What you see
Our Python scripts: Cannot reach Ollama at http://127.0.0.1:11434. Is it running? [Errno 61] Connection refused
Cause
Nothing is running at port 11434.
Fix
Open the Ollama app, or run ollama serve.
Status
The [Errno 61] Connection refused part: seen on our Mac (Python 3.14.6, URLError). The first part is our script's text (timing.py).
What you see
Node: TypeError: fetch failed, cause ECONNREFUSED
Cause
Nothing is running at port 11434.
Fix
Open the Ollama app, or run ollama serve.
Status
Seen on our Mac (Node v22.23.2).
What you see
Node: ReferenceError: fetch is not defined
Cause
This Node has no built-in fetch. Our .mjs scripts need Node 18 or newer.
Fix
Install Node 18 or newer. node -v shows yours.
Status
Seen on our Mac (Node v16.20.2, 2026-10-23).
What you see
Error: listen tcp 127.0.0.1:11434: bind: address already in use after ollama serve
Cause
The Ollama app already runs the server on port 11434, and the serve command stops at its listen step.
Fix
Do not run ollama serve while the app runs: the server is already there. Or quit the app first.
Status
Seen on our Mac (macOS, Ollama 0.35.0, 2026-10-23).
What you see
The Mac installer will not run
Cause
Ollama on macOS needs macOS 14 (Sonoma) or newer.
Fix
Update macOS to version 14 or newer.
Status
In Ollama's docs.

Model

What you see
HTTP 404 {"error":"model 'qwen3:1.7b' not found"}
Cause
The model is not pulled, or the name has a typo.
Fix
ollama pull qwen3:1.7b, then check ollama list.
Status
In Ollama's code. Seen on our Mac: curl, speed.mjs, speed.py and practice_table.py all printed this one before the pull (2026-10-23).
What you see
HTTP 404 model "qwen3:1.7b" not found, try pulling it first
Cause
Same, found at another step.
Fix
ollama pull qwen3:1.7b, then check ollama list.
Status
In Ollama's code. None of our four tries printed it.
What you see
practice_table.py stops after the 0.6b rows with HTTP 404, and no table file appears
Cause
The 1.7b model is not pulled. The script saves its two files only after the last run, so the 0.6b rows are lost too.
Fix
ollama pull qwen3:1.7b, then run the script again.
Status
Seen on our Mac (2026-10-23).
What you see
You used qwen3:4b and it thinks, even with "think": false
Cause
That tag is a thinking-only build: ollama show qwen3:4b lists thinking levels true only. Its thinking and a closing </think> mark come back inside message.content.
Fix
Use qwen3:1.7b for the bigger model.
Status
Seen on our Mac (Ollama 0.35.0, 2026-10-23).

Thinking

What you see
TTFT is many seconds and message.content stays empty for a long time
Cause
think was left out. v0.35.0 turns thinking on for qwen3.
Fix
Send "think": false.
Status
In Ollama's code. Seen on our Mac: lesson 3's two runs with no think field both came back with thinking (2026-10-23).
What you see
HTTP 400 "<model>" does not support thinking
Cause
think: true sent to a model that cannot think.
Fix
Use qwen3:0.6b, or drop think for that model.
Status
In Ollama's code.
What you see
Our scripts: No answer text came back (done_reason: length). Did thinking use up the token budget?
Cause
A token limit ended the reply before any answer text. With thinking on, thinking may use the whole budget.
Fix
Raise num_predict, or turn thinking off. think_on.py sets no limit.
Status
Our script's text (timing.py). Seen on our Mac with thinking on and num_predict 150, the Draft reply prompt (2026-10-23).
What you see
curl's first byte is fast, but the answer feels slow
Cause
With thinking on, the first byte is thinking.
Fix
Time the first non-empty message.content in code.
Status
In curl's docs.

Timing and counters

What you see
tokens/s like 0.00000006
Cause
Times are in nanoseconds.
Fix
Multiply by 10^9.
Status
In Ollama's docs.
What you see
tokens/s much lower than expected
Cause
You divided by total_duration. It includes loading, prompt reading and waiting.
Fix
Divide by eval_duration.
Status
In Ollama's docs and code.
What you see
Line count is not equal to eval_count
Cause
Ollama does not promise one line per token. When the model stops on its own, its last token only marks the end, so it sends no text line. With thinking on, Ollama can also hold text back until it can split it.
Fix
Count tokens with eval_count.
Status
In Ollama's code. Seen on our Mac: 26 text lines for eval_count 27, thinking off; 677 and 682, thinking on (2026-10-23).
What you see
budget_check.py: error: unrecognized arguments: for lesson 4:
Cause
You pasted the label for lesson 4: with the numbers.
Fix
Paste only the part after the colon.
Status
Seen on our Mac (Python 3.14.6).
What you see
TypeError [ERR_PARSE_ARGS_UNEXPECTED_POSITIONAL]: Unexpected argument 'for' from budget-check.mjs
Cause
You pasted the label for lesson 4: with the numbers.
Fix
Paste only the part after the colon.
Status
Seen on our Mac (Node v22.23.2).
What you see
The first run is much slower
Cause
Cold start: the model loaded on that call.
Fix
One warm-up call, or report cold and warm apart.
Status
In Ollama's docs.
What you see
A line {"error":"..."} in the stream, while HTTP was 200
Cause
The error came after the stream started.
Fix
Check every line for error (our scripts do).
Status
In Ollama's code.
What you see
Python httpx.ReadTimeout (your own code, if it uses httpx)
Cause
httpx waits 5 s by default.
Fix
timeout=None. All our scripts, lesson 2's speed.py included, use urllib with a long timeout (600 s in the practice and break it scripts and in speed.py).
Status
In httpx's docs.
What you see
Node: SyntaxError: Unterminated string in JSON at position 66 (line 1 column 67), or Expected ',' or '}' after property value in JSON
Cause
JSON.parse got half a line: a network chunk does not end at a line break.
Fix
Keep a buffer, split on \n, keep the last half line.
Status
Seen on our Mac, on a line we cut by hand (Node v22.23.2, 2026-10-23). The position changes with the cut.
What you see
A repeat prompt has a much lower TTFT, and prompt_eval_cached_count is high
Cause
The prompt cache was reused.
Fix
For fair tables, change the start of each prompt (our scripts add a random tag).
Status
In Ollama's docs. Seen on our Mac: prompt_eval_cached_count 22 of 23 when first-call.sh ran again (2026-10-23).
What you see
prompt_eval_cached_count is missing
Cause
Ollama leaves the field out when it has no value.
Fix
Nothing to fix. Our script records null.
Status
In Ollama's code.

Practice table

What you see
Long prompt tokens are not bigger than medium, or look cut
Cause
The prompt is longer than the context window.
Fix
Check the CONTEXT column in ollama ps. If the long prompt is cut, compare only the short and medium rows.
Status
In Ollama's code: the default context depends on GPU memory. Seen on our Mac: 4096, with 24 GB of memory and 17.8 GiB of GPU memory in the server log (2026-10-21).
What you see
A measured row has a big load ms
Cause
The model was unloaded mid-table (5 minutes idle, or memory pressure).
Fix
Rerun the script.
Status
In Ollama's docs.
What you see
tokens/s jumps a lot between the 3 runs
Cause
Other work on the machine.
Fix
Close other apps, plug in power, rerun.
Status
Our advice, not a measured fact.
What you see
ollama ps shows 100% CPU
Cause
The model is not on the GPU.
Fix
Read the server log. The top of this page says where it is.
Status
In Ollama's docs.

Break it 1 (five at once)

What you see
HTTP 503 server busy, please try again. maximum pending requests exceeded
Cause
More than OLLAMA_MAX_QUEUE (512) requests waiting for a model to load. Five requests will not do this.
Fix
Send fewer at once, or raise OLLAMA_MAX_QUEUE.
Status
In Ollama's code. The two spaces are in Ollama's text.
What you see
Log line model architecture does not currently support parallel requests
Cause
You used a model family that Ollama forces to 1 slot.
Fix
Use qwen3:0.6b for this lab.
Status
In Ollama's code.

Break it 2 (keep_alive: 0)

What you see
ollama ps lists the model, and UNTIL says Stopping...
Cause
The model's time is up. Ollama unloads it when its last request ends.
Fix
Wait a second and run ollama ps again.
Status
In Ollama's code. Seen on our Mac only with a fast poll, during a keep_alive: 0 call on a cold model: from the end of loading to the reply, about 0.1 s (2026-10-23). By hand you will most likely not see it.
What you see
The model reloads even without keep_alive: 0
Cause
More than 5 minutes passed since the last call.
Fix
Expected. That is the default.
Status
In Ollama's docs and code.
What you see
load_duration is small, but the first call after a long idle is still slow
Cause
A developer saw this on the MLX engine (Ollama 0.34.2, M4 Pro): the wait was in reading the prompt. Our qwen3:0.6b on v0.35.0 runs on llama.cpp, which the report says keeps the weights in place.
Fix
Measure TTFT on your side, from send to the first answer word. On the MLX engine, Ollama 0.40.0 and newer have a fix.
Status
Seen in a report. Not verified yet: our engine after a long idle.

The check page

What you see
"My answer is close but marked wrong"
Cause
The band is ±5%, or the unit is wrong (ms against s).
Fix
Read the unit in the question.
Status
What you see
"I used my own numbers"
Cause
The check uses the recorded log, not your run.
Fix
Use the numbers in the log on the page.
Status

Copying the scripts

What you see
first-call.sh: line 2: http://localhost:11434/api/chat: No such file or directory
Cause
The file lost its line breaks when it was copied. The curl -N line ended up inside the # comment above it, so the URL ran as a command.
Fix
Download the file from the lesson, or copy it again with its line breaks.
Status
Seen on our Mac (2026-10-21).
What you see
curl: option -o: requires parameter, then ttft.sh: line 4: /dev/null: Permission denied
Cause
The same: lost line breaks split the command in the wrong places.
Fix
Download the file from the lesson, or copy it again with its line breaks.
Status
Seen on our Mac (2026-10-21).
What you see
{"error":"invalid character '}' looking for beginning of object key string"} from first-call-once.sh
Cause
A comma after the last item, "stream": false,, makes the request bad JSON.
Fix
Remove the comma after "stream": false.
Status
Seen on our Mac (2026-10-23). curl still exits with status 0; the server log shows 400.