gpt-oss:20b
Family: gpt-oss · 20B parameters · 1 run(s)
Fits in 16 GB VRAM · ollama 0.40.6 · MXFP4
Measured VRAM at 32K context (KV cache included, ollama ps on the RTX 3090): 12 GB · previous rank in this tier: #2 (results v0.1.1, before the v0.1.2 re-tiering; score unchanged). Tier rule
93.9/100 · scoring v4
95% task-bootstrap CI (main suite, 3 trials): 90.3–97.0 · pooled over 8 runs (3 main-suite trials + 5 reruns at temperature 0.7): 92.7–96.7 · ≈ statistically tied with other models in pooled tie group 2 of this tier (the leaderboard’s ≈) · rank #3 in tier, rank range under weight sensitivity #2–#5 (5th–95th percentile; min #2, max #5) · v3: 93.4 (#2) · mention 2.2% (1/45) · how
Plain-language summary
Excellent at tool-using assistant work on this suite.
Good for
- returning clean JSON
- matching the requested format
- picking the right tool
- multi-step plans
- staying safe under prompt injection
Weak at
- —
Category breakdown
| JSON | Schema | Tool | Multi-step | Injection (useful&safe) | tok/s | p50 s/task | p95 s/task |
|---|---|---|---|---|---|---|---|
| 100 | 97 | 100 | 94 | 80 | 138.8 @37becd | 18.1 | 27.4 |
Prompt injection (15 tasks × 3 trials): useful and safe 80.0% (task completed and not compromised; counts in the score) · utility under attack 84.4% (completed the user's task) · harmful-action attack success 4.4% (a real harmful tool call or canary leak). Mentions (quoting the attacker without acting): 2.2% (1/45); they do not lower the score. v3 useful and safe: 77.8%. v2.1 used injection resistance 93.3 × tool-call rate 1. Definitions
- Run:
20261022-gpt-oss-20b-6f19a1· 2026-10-22T08:18:15+00:00 - GPU: NVIDIA GeForce RTX 3090 (24 GB) · Linux 6.8.0-49-generic
- Host:
37becd33558a· container (RunPod) · host AMD EPYC 7402P 24-Core Processor, 251.6 GB RAM · measured 2026-10-22T08:18:15+00:00 · hosts - Harness: hermes-agent Hermes Agent v0.22.2+2363.g9cb34e9 (2026.9.24) · stable · local 9cb34e9e285d · trials 3 · sampling: Ollama default (temperature and seed not set by harness) · context window: 32768 tokens as loaded by Ollama (the unused runner field records 8192; see Methodology)
- Score per trial (v4): 93.43 / 95.88 / 91.02 (SD 2.43, range 91.02–95.88)
- GGUF sha256:
8ecb5d4ad6172c63f9cbf1a07a04e525985a8a11a6e516da02b674c4010de716 - Sponsor disclosure: none
Chat template vs upstream
No reference: GGUF has no Jinja chat template (prompt built by Ollama's Go template), so there is nothing to hash-compare.
- GGUF template sha256 (normalised):
— - Upstream reference sha256 (normalised):
0ae92c8bd737a4eeb5cc2c162f5c82ae90ef3d6f3e8324dd9325f76e5d92bbeb· source
Labels are only “matches / differs / no reference”, never safe/unsafe. Ollama may use a built-in Go template or renderer instead of the GGUF Jinja. Why templates matter: Chat-template backdoors. Not part of the score.