deepseek-coder-6.7B-base: benchmark results
DeepSeek's deepseek-coder-6.7b-base, a 6.7B code model using Multi-Head Attention, pretrained from scratch on 2 trillion tokens (87% code, 13% English/Chinese text) with a 16K context window. Provider: DeepSeek. Released 2023-11-02. Access: Open.
Aggregate Rating 1300 ± 25, rank #1484 of 1656 rated models, from 20 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| ShaderMatch | 17.51 | Clone Match Rate (%) | 95.2 |
| SAFIM | 63.4 | Average | 92.9 |
| CanItEdit (Lazy) | 27.76 | pass@1 (%; share of the 105 hand-written Python editing prob | 62.5 |
| EffiBench - NMU | 1.91 | Normalized Memory Usage | 58.5 |
| EffiBench - NET | 2.98 | Normalized Execution Time | 51.2 |
| EvalPlus (HumanEval+ & MBPP+) | 49.2 | Pass@1 avg (%) | 42.7 |
| CanItEdit (Descriptive) | 32.62 | pass@1 (%; share of the 105 hand-written Python editing prob | 41.7 |
| C3-Bench (Controllable Code Completion) - Implementation Control | 27.1 | Instruction-Following Rate (%; 1,286 ICC tasks: passes the u | 40 |
| Open LLM Leaderboard v1 - GSM8K | 17.97 | Accuracy (%) (5-shot) | 39.4 |
| C3-Bench (Controllable Code Completion) - Implementation Control - Pass@1 | 41.8 | Pass@1 (%; unit tests on the 1,286 instructed ICC tasks) | 32.5 |
| C3-Bench (Controllable Code Completion) - Scale Control | 4.9 | Instruction-Following Rate (%; 909 SCC tasks: AST node-type | 25 |
| GSM8K | 21.3 | Accuracy (%) | 22.6 |
Interactive version: theaggregate.ai/model?slug=deepseek-coder-6-7b-base · How It Works · Data refreshed daily, snapshot 2026-10-20.