Qwen3 on the Clone: Checkpoint 3

Covers: Chapter 2 (Qwen3-ASR model installation and AI Tools) plus the llama-ui memory-file test saved 18/10/2026. ASR = Automatic Speech Recognition (speech to text). It is the original goal; it has not been started.

---

1. Problem & Resolution

Item	State
Python 3.15.0 built on the Clone, copied to Live, python3.15 wrapper on both	Done on Clone. Live test output never seen, so Live is unconfirmed.
Time zone fix (two glibc symlinks)	Clone only. Not on Live.
Clone disk	Grown to 95 GB, 31 GB free.
Qwen3-4B (Q4_K_M) served by llama-server, browser UI working	Done. Short chats work (14 to 55 token replies, 1.3 to 1.7 tokens/s).
8B model file	Still on disk, not in use.
Loading the memory file (about 21 KB) into the 4B	Works but slow: 5,243 tokens read in 28 min 14 s (3.09 tokens/s).
Summary of the memory file	Accurate against the file. It opened with a Claude: [2026-10-18--18:09:26] line copied from the file's preference text. The time was not measured by the model.
Reply speed after the file was loaded	Collapsed to 0.36 tokens/s (107 tokens in 4 min 55 s) against 1.3 to 1.9 before. Cause not tested.
Qwen3-ASR itself	Not started.

Open from the start of the chapter: does llama-mtmd-cli support the Qwen3-ASR model? Not yet checked.

2. Risk Analysis

• Context limit: the server runs with -c 8192. The memory file used 5,243 tokens, and the long summary was still generating when the PDF was saved. It was probably cut off at the limit.
• Unknown server flags: ps -o args= -C llama-server has not been run. We do not know whether -np 1, -t 2/-t 4 or --verbose were in force during any of these numbers. The earlier log showed slot id 3, which suggests -np 1 was not applied. Speed comparisons are not trustworthy until this is settled.
• Live is behind the Clone on the Python build verification and on the time zone symlinks.
• A first reply said "I don't have direct access to load files" even though the file was already in context. The 4B is unreliable on statements about itself.

3. Code Change

None this checkpoint. No scripts, packages or configuration were changed after Checkpoint 2.

Next steps (options, your call on timing)

1. Run ps -o args= -C llama-server on the Clone and note the real flags.
2. Retest the 4B with a trimmed memory file of about 1,000 tokens (estimated 5 minutes to read).
3. Check whether llama-mtmd-cli can load Qwen3-ASR, then decide whether the ASR work starts in a fresh conversation.
4. Confirm Python 3.15 on Live and, if wanted, the two time zone symlinks.

Release card

Rule	Status
Code delivered	Not needed (no code)
Version bump / history entry	Not needed
Zip with dotted version	Not needed
Legacy zipper used	Not needed
File name not reused	Done (_3; earlier checkpoint was _2)