DeepSeek-Coder 1.3B
The tiny coder that started DeepSeek's open-source run: 1.3B params, 16K context, fill-in-the-middle, under 3 GB VRAM.
About
DeepSeek-Coder-1.3B is the smallest member of the code model family that made DeepSeek's name before its chat models did. Trained from scratch on 2 trillion tokens, 87% of it code in English and Chinese, it was built for project-level work: a 16K context window and a fill-in-the-middle training objective make it genuinely useful for code completion and infilling inside an IDE, not just snippet generation. At 1.3B parameters it fits in under 3 GB of VRAM.
The base model scores 32.1 on HumanEval Python in third-party indexes, which was state-of-the-art territory for open 1B-class code models at its late-2023 release, and the instruct variant adds conversational coding help. The code is MIT and the weights carry DeepSeek's model license with commercial use allowed. Newer tiny coders like Qwen2.5-Coder-1.5B have since overtaken it on raw scores, but it remains a dependable, well-tooled baseline for on-device coding assistance.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- HumanEval (Python): 32.13 on the base model (third-party index, llm-explorer)
- Provider claimed SOTA among open-source code models at release across HumanEval, MultiPL-E, MBPP, DS-1000 (Jan 2024 paper)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: 2T-token training, 16K context, fill-in-the-middle objective and DeepSeek Model License (commercial use permitted) confirmed on the official deepseek-ai/deepseek-coder-1.3b-instruct model card and the Jan 2024 paper (2401.14196); HumanEval 32.13 from a third-party model index. Strengths: proven on-device coding baseline, permissive-enough licensing, broad quantization support. Docked for: (1) superseded on raw scores by newer tiny coders like Qwen2.5-Coder-1.5B; (2) provider's SOTA claims date to release and were never independently re-verified at this size. License: MIT (code); DeepSeek Model License (weights, commercial use permitted).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 7.0/10
- Manual review · sanity read + red-flag check
- 6.0/10
- · DeepSeek; released 2023-11
- · Paper 2401.14196 (Jan 2024)
- · HumanEval 32.13 base via third-party index; provider SOTA claim at release