nebius-build-week/tryouts-eval

Tryouts

Evaluation sets and full run logs for one question: which open model should a web research agent use? The same agent (same prompts, same Tavily search) ran across the open models on Nebius Token Factory and Claude Opus 5 as the baseline. Code to rerun everything on your own questions: https://github.com/Malek1414/tryouts. Leaderboard: https://huggingface.co/spaces/nebius-build-week/tryouts.

Built at the Nebius Build Week hackathon with NVIDIA, Berlin, 18 October 2026, by team Starke Ameisen.

Sets

Config	Items	What it tests
single_facts	50	Verified single facts in German and English. Every model answers from the same Tavily results.
hard_multihop	40	Questions from FRAMES (numerical, temporal, multiple constraints) that chain 2 to 5 facts. Each model plans its own searches, up to 3 rounds.
own_logs	16	Written from research our own agents ran in the last month (Nightcrawler and Searchify). Each answer is backed by a verbatim quote checked against the page, and some questions rest on a false premise, so the right answer is "unknown".
own_logs_deep (results only)	16	The own-logs questions in deep mode: full pages, Searchify's reranker (cross-encoder/mmarco-mMiniLMv2-L12-H384-v1), one query per fact and a quote gate.
tool_use	14	The model drives real MCP servers through tool calls: Tavily search, extract, map and crawl; semble code search on GitHub; a headless Chrome for pages that only render with JavaScript. Gold answers were checked against the source on 18 October 2026.

Fields: question, answer (gold), aliases, answer_type (exact, number, unanswerable), source, and for tool_use the intended tool. The results_* configs hold every run: answer, verdict, search queries or tool calls, sources, tokens, cost in USD and seconds.

The exact Tavily results every model read are in search_cache/tavily_cache.jsonl: drop it into the repo's results/ folder and a rerun replays the same searches without spending credits.

Scoring

Verdicts follow SimpleQA: correct, incorrect (a wrong answer given with confidence) and not_attempted (the model said it could not find it). On unanswerable items only "unknown" is correct. Numbers allow 1% tolerance. Anything a rule cannot settle goes to a fixed judge model (Kimi K2.6) that is not a contestant. Accuracy intervals are bootstrap 95% intervals (10,000 resamples). The pick is the cheapest model whose accuracy is not credibly more than 10 points below the best one.

Results (18 October 2026)

Model	Single facts	Hard multi-hop	Our own logs	Tool use (MCP)	$ / question (hard)
Claude Opus 5	100% (100%–100%), 0% wrong	88% (78%–98%), 12% wrong	69% (44%–88%), 19% wrong	100% (100%–100%), 0% wrong	$0.0619
Reflection Beam	100% (100%–100%), 0% wrong	80% (68%–92%), 12% wrong	69% (44%–88%), 19% wrong	86% (64%–100%), 7% wrong	$0.0047
Nemotron 3 Super	100% (100%–100%), 0% wrong	77% (64%–90%), 13% wrong	56% (31%–81%), 31% wrong	86% (64%–100%), 7% wrong	$0.0054
Nemotron 3.5 Super VL	–	77% (64%–90%), 8% wrong	75% (50%–94%), 19% wrong	100% (100%–100%), 0% wrong	$0.0076
DeepSeek V4 Flash	100% (100%–100%), 0% wrong	70% (55%–82%), 8% wrong	75% (50%–94%), 19% wrong	100% (100%–100%), 0% wrong	$0.0021
Nemotron 3.5 Lightning	96% (90%–100%), 4% wrong	65% (50%–80%), 20% wrong	69% (44%–88%), 25% wrong	64% (36%–86%), 36% wrong	$0.0028
Nemotron 3 Ultra	100% (100%–100%), 0% wrong	57% (42%–72%), 15% wrong	62% (38%–88%), 38% wrong	0% (0%–0%), 100% wrong	$0.0244
gpt-oss 120B	96% (90%–100%), 4% wrong	55% (40%–70%), 28% wrong	56% (31%–81%), 44% wrong	57% (29%–79%), 14% wrong	$0.0018
Qwen3 30B A3B	98% (94%–100%), 2% wrong	50% (35%–65%), 35% wrong	62% (38%–88%), 31% wrong	79% (57%–100%), 21% wrong	$0.0009
Nemotron 3.5 Lightning (no thinking)	94% (86%–100%), 6% wrong	38% (22%–52%), 52% wrong	56% (31%–81%), 44% wrong	–	$0.0005
Nemotron 3 Nano	96% (90%–100%), 4% wrong	–	–	86% (64%–100%), 0% wrong	–

Bold: the pick for that set (cheapest model not credibly more than 10 points below the best). Ranges: bootstrap 95% intervals. "Wrong" counts confident wrong answers, not "unknown".

Per-question replay: see the Space.

Run-to-run variance (hard set, two passes)

Model	Run 1	Run 2	Questions that flipped
Reflection Beam	82%	79%	5 of 39
Claude Opus 5	88%	90%	3 of 40
Nemotron 3 Super	77%	62%	8 of 39
Qwen3 30B A3B	49%	41%	9 of 39

Caveats

• Small sets: intervals are wide, read them before the point estimates.
• An early bug cached Tavily rate-limit replies as empty searches; affected runs were removed and rerun. See REPRODUCE.md in the repo.
• Nemotron 3 Ultra's tool calls came back from Token Factory as raw <tool_call> text or under invented tool names when 8 tools were offered (with one tool it called it correctly). This looks like a serving or chat-template issue, not the model's ability.
• Prices are Token Factory list prices on 18 October 2026.
• Facts with dates ("as of October 2026") can go stale.

Licence

Our questions and run logs: CC BY 4.0. hard_multihop items come from FRAMES (Apache 2.0, Google).