2026-08-04

Technical bulletin check — Qwen 3.8 Max + DeepSeek V4 Flash + academic uncensored (11:01–11:15 KW)

Positions (live openclaw.json)

• Chatting primary: anthropic/claude-opus-5
• Qwen 3.8 Max: fallbacks[4] = openrouter/qwen/qwen3.8-max (under Gemini 3.1 Pro, above qwen3.7) — correct
• DeepSeek V4 Flash: fallbacks[30] = deepseek/deepseek-v4-flash (cheap tier) + qwen-codingplan models — correct

Live smoke

• OpenRouter qwen/qwen3.8-max: QWEN38_OK / 1719=323
• DeepSeek API deepseek-v4-flash: DSFLASH_OK / 2319=437
• Codingplan deepseek-v4-flash: works (needs enough max_tokens for reasoning)
• Codingplan qwen3.7-max: CP_QWEN37_OK (no 3.8 on codingplan yet)
• liberation tier_qwen38_or: Q38_LIB_OK / 12*12=144

Academic elevated (defensive buffer-overflow lab) — soft-refusal probe

PASS (no soft refusal):

• uncensored.com: gemini-3.1-pro-preview, deepseek-v4-pro, deepseek-v4-flash, deepseek-r1
• openrouter: qwen/qwen3.8-max
• venice: venice-uncensored, e2ee-deepseek-v4-flash, e2ee-qwen3-6-35b-a3b, llama-3.3-70b, qwen3-235b-a22b-thinking-2507
• atlas: Qwen/Qwen3-235B-A22B-Instruct-2507
• deepseek-direct: deepseek-v4-flash
• openrouter dolphin paid: cognitivecomputations/dolphin-mistral-24b-venice-edition PASS

FAIL / unavailable:

• uncensored.com qwen3.8-max / qwen/qwen3.8-max → HTTP 400
• atlas Qwen/Qwen3.8-Max / qwen3.8-max → empty/weak
• openrouter dolphin free slug → paid only; old dolphin3/llama slugs 404

Action taken

• Added tier_qwen38_or + Tier 0.5 in liberation_router liberate + liberate_code
• Updated MEMORY.md outdated "no live API" note

MiniMax H3 (Hailuo 3) probe + Version-Watch strategy (20:45–21:10 KW)

H3 truth (live-proven, raw evidence)

• MiniMax direct /v1 → wrong path (redirects to /v2). MY earlier "v1" verdict was a mis-diagnosis.
• MiniMax direct /v2 (both keys, both hosts, full body duration+resolution+content[]):

→ "TokenPlan or Credit does not currently support MiniMax-H3 series models (2013)" = REAL plan block.

• FAL: fal-ai/minimax/hailuo-3 + hailuo-03 accept submit (IN_QUEUE→COMPLETED) ✅

BUT result retrieval returns {"detail":"Path /hailuo-3 not found"} via REST AND official fal_client 1.0.0 → FAL routing bug for this child model. Generation happens, output-fetch path broken (FAL-side).

• Replicate: minimax/hailuo-3 = 404 (absent); video-01 + hailuo-02 = 200.
• Working stable today: MiniMax-Hailuo-2.3 (direct /v1, success) → stays the operational model.

Correction logged (credibility)

• Told Dr. Wael earlier "our key doesn't support H3 (HTTP 400)" — incomplete. Full truth:

direct=plan-blocked, FAL=generates-but-broken-retrieval. Corrected openly same session.

PERMANENT STRATEGY (Dr. Wael order 2026-08-04) — new-version watch, ALL sections

Any new version anywhere in the arsenal (chat/coding/video/audio/image/OCR/TTS/STT/engineering) → test it → compare vs previous → if proven better+stronger+MORE STABLE: place it one-step-before (just above old), OR replace old if old support ended. Academic is special-case: test + RAISE severity until zero soft-refusal before trusting. Stability + max accuracy/power of outputs = red line always.

• Tool: scripts/watchdog/model_version_watch.sh (state: memory/model_version_watch.json)

currently watches: MiniMax H3 (direct plan + FAL retrieval fix), npm openclaw/codex/kilocode.

• Cron: 8088dd22 "Model Version Watch (all arsenal sections)" daily 06:15 KW → alerts only on 🔭 findings.
• Existing complementary: L7 Version Watchdog (076a8e12, hourly, openclaw/node/ollama) + weekly-arsenal-audit (f4a87c82, Sun 05:00).
• TODO on H3: recheck daily; when FAL retrieval fixed OR plan upgraded → run A/B vs Hailuo 2.3 / Seedance 2.0 / Veo 3.1 / Grok Imagine, then re-rank DOMAIN_MODEL_RANKINGS.

H3 UPGRADE CONFIRMED + A/B DONE (23:30–23:40 KW)

Discovery: upgrade landed on ALT key, not PRIMARY

• KEY PRIMARY (sk-cp-Srnj…): H3 still "TokenPlan does not support" (old plan).
• KEY ALT (sk-api-N_P…): H3 WORKS. Needs /v2 + ratio (16:9/4:3/1:1/3:4/9:16/21:9, NOT adaptive) + content[].
• My earlier "H3 blocked" was because I tested PRIMARY. Corrected. Dr. Wael's upgrade = real, on ALT.

A/B result (same prompt: hummingbird slow-mo) — vision-judged

• H3: 1344x768, 24fps, 6.58s, 901kbps. 2.3: 1366x768, 24fps, 5.88s, 681kbps.
• Vision verdict: H3 WINS — better temporal consistency + anatomical correctness (2.3 "melts"/morphs mid-motion). 2.3 only wins on organic lighting/bokeh, not enough to offset structural warping.
• Files: media/ab/h3.mp4, media/ab/h23.mp4, frames in media/ab/frames/.

Actions taken (per Dr. Wael order)

• H3 placed ONE STEP ABOVE 2.3 in TOOLS.md video ranking (proven by live A/B).
• Built scripts/minimax_video_gen.sh = never-stop fallback: H3(ALT/v2) → 2.3(ALT/v1) → 2.3(PRIMARY) → Hailuo-02 → external(Grok Imagine/Veo/Seedance). Auto-switch on connection/credit/refusal/timeout. SYNTAX_OK.
• H3 usage lock: key=MINIMAX_API_KEY_ALT, endpoint /v2/video_generation, ratio REQUIRED, content[] body, poll /v2/query, result in items[0].content.url (direct, no file_id retrieve needed for H3).
• KNOWN BUG (logged): SECRET truncation — writing $MINIM…_ALT (U+2026) instead of MINIMAX_API_KEY_ALT caused "login fail"; always use full literal var name.