Qwen3.6-27B-AWQ

Wallstreet Intern Test — Qwen3.6-27B-AWQ

Subject: GitLab Inc. (GTLB) Recommendation: BUY (12-month price target in the memo) Audit date: 2026-04-26 Auditor: Qwen3.6-27B AWQ-INT4 (Cyankiwi quantization), dense, thinking-mode vLLM config: --max-model-len 262144, --temperature 0.0, --reasoning-parser qwen3 --tool-call-parser qwen3_xml Wall-clock: 27 min, 56 iterations, 32K completion tokens Run name: 27b_invest_memo_v2 (the cherry-picked successful run of three; see "Variance" below)

Read this first — what this entry is and isn't

This entry is a successfully completed memo from a model that succeeded on this task only 1 of 3 attempts. The deliverable is real: a memo with explicit BUY recommendation, a 17 KB three-statement XLSX model, raw filings, extracted data, ADRs, sources with SHAs, and dated research notes. Tagged a release at end-of-run. But the next two attempts at the same task with the same flags produced very different outcomes — see Variance.

Read in this order

If you have 5 minutes: read memo/'s primary memo file. It's the PM-facing document.

If you have 20 minutes: read the memo, then decisions/ for the ADR-style records of non-obvious choices (which company, what discount rate, which competitors), then analysis/ for the sell-side-miss test.

If you want the full audit chain: every claim in the memo should trace through extracted/ → model/gitlab_three_statement_model.xlsx → memo. Quotes from management trace to line-numbered transcript text in extracted/transcripts/. External citations resolve via sources.md (URL + SHA-256 + timestamp).

Variance — read this before treating this entry as "27B's answer"

The model was run three times on the same task with the same flags:

run	wall	iters	finish	shipped?	notes
27b_invest_memo_v2	27 min	56	done_signal ⭐	yes ← this entry	2K-word memo, 6-sheet XLSX, 14-alternative ADR, full audit trail
27b_invest_memo_v3	11 min	40	model_stopped (parser fault)	partial	substantive work but cut off mid-emission
27b_invest_memo_v4	68 min	45	api_error: timed out	no	got stuck on a single ~1-hour inference call

So 1 of 3 runs shipped a complete memo. v2 (this entry) is the published result. The other two are kept in the source bench repo as failure-mode evidence.

The relevant comparison: cloud LLMs (Opus-4.7, GPT-5.5) shipped on every attempt. Local 30B-class quantized models on this task do not.

Repository layout

memo/                  PM-facing memo (markdown source + the thesis prose)
model/                 Three-statement model (XLSX)
raw/                   Original primary sources (filings, transcripts, press)
extracted/             Parsed/cleaned data from raw/, with extraction scripts
analysis/              Sell-side-miss test, competitive analysis, scenarios
research/              Working notes, dated, plus questions.md and dead-ends.md
decisions/             ADRs for non-obvious choices (company pick, discount rate, etc.)
sources.md             Every URL fetched with timestamp + SHA-256
tool-log.md            Tool calls in order, with one-line justifications

Caveats specific to a thinking-mode local model

Compared to the cloud entries on this benchmark:

• No board-of-advisors presentation follow-on. The wallstreet-intern-test benchmark prompt itself is just the memo task; the board-of-advisors-presentation/ folder under GPT-5.5/ was a separate follow-up task (see agent-pilot/task_board_presentation.md in the source bench repo). 27B-AWQ has separately produced board-deck artifacts in the source bench (run 27b_board_pres_v1 etc.) but those are not packaged as part of this entry.
• PDF rendering is the markdown source, not a typeset PDF. The cloud entries include a styled PDF; 27B's run produced markdown only. The PDF generation step in the spec was not exercised.
• Numbers are checkable but not all are checked. The XLSX model has 6 sheets with formulas, but I have not run a comprehensive consistency-check script across them. The cloud entries include model/key-outputs.ndjson, model/checks-inspect.ndjson, and model/formula-error-scan.ndjson machine-readable verification artifacts; this entry does not have those files because the agent didn't generate them. The XLSX itself opens cleanly and the formulas are sane on inspection.

Reproducibility

The exact harness invocation, vLLM container args, GPU snapshot, and task SHA are recorded in the run's receipt.json in the source bench repo (private to the maintainer).

To replay (rough):

python3 agent-pilot/harness.py replay_27b_invest_memo_v2 \
  agent-pilot/task_investment_memo.md \
  --model qwen3.6-27b-awq --port 8000 \
  --temperature 0.0

Note that v2 ran at temperature 0.0, before the deterministic-loop-trap discovery that informed our shift to temperature 0.3 for the later PR-audit task family. vLLM bf16 paths aren't bitwise-deterministic regardless; expect divergence.