OpenGPTEvidence-based model decisionsStarting candidate 1
PreviewDeepSeek V4 Pro
deepseek-v4-proDeepSeekOpen-weight evidence · #2
09 · TASK RANKING
Private & local deployment
Open-weight candidates for controlled data paths and self-hosting.
- Official version
- 3
- Public evidence
- 1
- Data cutoff:
- Why these candidates
- Start with confirmed downloadable versions and open-weight public signals, then validate hardware, license, security, and serving cost.
- What remains uncertain
- Open weights do not automatically mean low cost, full commercial freedom, privacy, or easy operation.
- Public evidence
- Open-weight evidence
- Data cutoff
Starting candidate
3 / 3Three official versions to test first
Public data narrows the field. It does not replace a controlled test on your own work.
Starting candidate 2
StableLlama 4 Maverick 17B-128E
meta-llama/Llama-4-Maverick-17B-128E-InstructMetaNo comparable public result is currently mapped to this exact version.
Starting candidate 3
StableQwen3 235B-A22B
Qwen/Qwen3-235B-A22BAlibaba QwenNo comparable public result is currently mapped to this exact version.
Public evidenceThe models use different public evidence scales. Missing evidence is not treated as zero, and OpenGPT does not combine unlike sources into one score.⌄
DeepSeek V4 Pro
- Source
- LiveBench ↗
- Source rank
- #2
- Source value
- 71.6
- Mapping reviewed
Llama 4 Maverick 17B-128E
No comparable public result is currently mapped to this exact version.
Qwen3 235B-A22B
No comparable public result is currently mapped to this exact version.
A practical starting point, not a universal winnerSelect two or three candidates for a side-by-side public data comparison.
The models use different public evidence scales. Missing evidence is not treated as zero, and OpenGPT does not combine unlike sources into one score.
Open weights do not automatically mean low cost, full commercial freedom, privacy, or easy operation.
The evaluation opens with this task template and the first two selected models. You can edit every task before collecting answers.