Model Profile · Round 01
Qwen 3.5 Flash
Alibaba · open weights · local-friendly · 128K ctx
⚠ FloorFloor on agency and instruction drift. Caveat emptor.
Composite Score
7.1
/100 · canonical
Arena ELO (R1)
1487
±47 · n=401
Multi-Turn ELO (R2)
1412
±47 · n=216
Reliability Rank
#19
failure-mode rubric
▌ Section 02 · The Lede
What this model is for.
Qwen 3.5 Flash sits at #8 community ELO and #9 reliability, with two structural floors: agency respect bottoms at 2.5/5 (lowest in the pool) and instruction drift bottoms at 2.5/5 — both the lowest single-session scores on those axes in our pool. F12 instruction drift mean is also the lowest in our 11 at 3.17. The 42% NSFW win rate is mid-pack but the floor scores tell the story. It's the field's caveat-emptor model.
▌ Section 03 · At a Glance
Cross-test position
Qwen 3.5 Flash sits at #20 on Composite, the caveat to watch.
Composite
20
Arena ELO
8
Multi-Turn
20
Rubric
20
Adversarial
19
Cost · Latency
6
▌ Section 04 · Strength & Weakness
Where it shines. Where it stumbles.
▲ Strength
No standout strength on the dimensions we tested. Best read as a budget option for non-critical scenes where structural failure is acceptable.
▼ Weakness
Catastrophic floors on agency respect (lowest session 2.5) AND instruction drift (lowest session 2.5) — two structural Tier-1 failures. F12 mean of 3.17 is also the lowest in our pool. Llama is the only other model with floors on both axes, but Qwen's are deeper on each. Reach for almost anything else if the system prompt matters.
▌ Section 05 · Failure Modes
Per-axis breakdown.
Six adversarial probes per session, twenty sessions per model, judged by Sonnet 4 against a fixed rubric. Further right = the model handled the failure mode better. Each axis is drawn as a band on the rubric’s 1 to 5 scale, not a number: judges disagree by about 0.3 at the model level, so overlapping bands are a tie. The right column is the rank within the rp-bench pool.
F1 · Agency
Doesn't write your character's actions
#20
F2 · POV / Tense
Holds 2nd-person, present-tense narration
#20
F3 · Lore
Doesn't break worldbuilding
#9
F8 · Momentum
Pushes scene forward when user goes passive
#9
F12 · Instruction Drift
Keeps to the system prompt
#21
F13 · Context Attention
Holds character cards 50+ turns deep
#17
“Two structural floors — agency and instruction drift, both bottoming at the worst scores in the pool.”
Round 01 verdict · Caveat emptor
▌ Section 06 · Subjective Dimensions
Engagement · Voice · Collaboration.
All three dimensions scored 1 to 5 by the Sonnet 4 LLM judge across twenty 12-turn multi-turn sessions. The same battery feeds the failure-mode rubric above; these are the subjective half of that judgment, drawn as bands with no number for the same reason.
Engagement
Tone Consistency
Collaboration
▌ Section 07 · Behavioral Metrics
How it writes.
Quantitative signals from the same 20 multi-turn sessions, compared against the population mean across all 11 models.
Avg words / turn
229↓
pop avg 265 · -13%
Unique-word ratio
0.634↓
pop avg 0.657 · -4%
Repetition score
0.069↑
pop avg 0.048 · +44%
▌ Section 08 · Flaw Hunter
Adversarial probe score.
Score of 100 minus deductions across 22 fail-mode flag types on adversarial 12-turn sessions. Further right = fewer flaws caught. Drawn as a band, not a number: about ten points either way is rater noise, so bands that overlap are tied.
▌ Craft band
From the Round 01/02 flaw-hunter pool. The current card carries the newer single-rater band.
Fatal/sess 0.50
Major/sess 6.50
▌ Top flaws caught
recycled_descriptionpurple_prosenarrating_emotions
▌ Section 09 · Sample Responses
Highest- and lowest-rated turns.
▌ Pending Round 02
Best- and worst-rated sample responses ship with the raw-vote endpoint in Round 02. When that lands, this section will surface the model’s highest- and lowest-scoring blind-arena turns side by side, scored on the same rubric the leaderboard uses.
▌ Round 01 verdict
Qwen 3.5 Flash is the model the failure-rank board flags hardest. There's no axis where it stands out, and two — agency respect and instruction drift — where its floor is the worst in the field. Avoid for any scene where character control matters. The 42% NSFW win rate isn't a saving grace; the structural failures show up regardless of mode.
▌ Section 10 · Compare & Drill
Stack it against another model.
Profile · Qwen 3.5 Flash · Round 01