DeepSeekLanguage model
DeepSeek Coder V2
DeepSeek Coder V2 has no overall index value yet because there are too few independent benchmark results.
32out of 100-1 vs. last month
based on search, usage and attention
Summary
DeepSeek Coder V2 stands out through its coding specialization and expanded language coverage. The Instruct variant has 236 billion total parameters, 21 billion active parameters, and a 128,000-token context window; compared with Coder-33B, coverage grew from 86 to 338 programming languages.
DeepSeek Coder V2 specifically means the Instruct variant here, not the Base version. The model card describes stronger code and mathematical abilities and a license that permits commercial use.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 31 published benchmark results
Page 1 of 3
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
Complexity-aware code comprehension (Show measurement, test conditions, and source)
- Source value
- 78.33
- Score
- 78.33
- Metric
- Overall automatic accuracy
- Unit
- %
- Category
- software-engineering
- Direction
- Higher is better
- Sample
- 300
- Source type
- independent-evaluator
- Evaluator
- Complexity-Aware Evaluation of LLM Comprehension, arXiv:2609.37405v1
- Status
- active
- Retrieved at
- 2026-10-20
- Methodology
- Independent evaluation. Exact model label: DeepSeek-Coder-V2. Zero-shot, temperature 0, one query per 300 Python-function tasks; no few-shot, chain-of-thought, or hints. Tasks: HumanEval, MBPP-sanitized, CRUXEval, LeetCode, BigCodeBench-Hard; 108 low, 120 medium, 72 high complexity. Automatic outputs are exact-matched after whitespace/quote/Boolean normalization. Immutable checkpoint/revision and serving route are not disclosed.
- Context
- The exact printed model label matches the assigned roster alias, but the paper does not identify an immutable revision. Do not transfer to Lite, Instruct, or another checkpoint.
BigCodeBench Full (Complete) (Show measurement, test conditions, and source)
- Source value
- 59.7
- Score
- 59.7
- Metric
- calibrated Pass@1 (greedy decoding)
- Unit
- %
- Category
- Full / Complete
- Direction
- Higher is better
- Sample
- 1140
- Source type
- official-leaderboard
- Evaluator
- BigCodeBench official leaderboard
- Status
- active
- Retrieved at
- 2026-10-28
- Methodology
- Official benchmark leaderboard. Exact leaderboard model row=DeepSeek-Coder-V2-Instruct; result split=Full; prompt variant=Complete; model link=https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Instruct; prompted=True; prefill=False; source model-date field=2024-06-17. Official leaderboard reports calibrated Pass@1 using greedy decoding. Evaluation run date and other per-model generation settings are not stated. Dataset tasks=1140.
- Context
- The leaderboard's JSON date field is retained as the model date, not an evaluation date. The site states calibrated Pass@1 using greedy decoding; it does not publish a run date or all per-model generation parameters. Complete and Instruct, and Full and Hard, are separate configurations.
BigCodeBench Full (Instruct) (Show measurement, test conditions, and source)
- Source value
- 48.2
- Score
- 48.2
- Metric
- calibrated Pass@1 (greedy decoding)
- Unit
- %
- Category
- Full / Instruct
- Direction
- Higher is better
- Sample
- 1140
- Source type
- official-leaderboard
- Evaluator
- BigCodeBench official leaderboard
- Status
- active
- Retrieved at
- 2026-10-28
- Methodology
- Official benchmark leaderboard. Exact leaderboard model row=DeepSeek-Coder-V2-Instruct; result split=Full; prompt variant=Instruct; model link=https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Instruct; prompted=True; prefill=False; source model-date field=2024-06-17. Official leaderboard reports calibrated Pass@1 using greedy decoding. Evaluation run date and other per-model generation settings are not stated. Dataset tasks=1140.
- Context
- The leaderboard's JSON date field is retained as the model date, not an evaluation date. The site states calibrated Pass@1 using greedy decoding; it does not publish a run date or all per-model generation parameters. Complete and Instruct, and Full and Hard, are separate configurations.
BigCodeBench Hard (Complete) (Show measurement, test conditions, and source)
- Source value
- 29.7
- Score
- 29.7
- Metric
- calibrated Pass@1 (greedy decoding)
- Unit
- %
- Category
- Hard / Complete
- Direction
- Higher is better
- Sample
- 148
- Source type
- official-leaderboard
- Evaluator
- BigCodeBench official leaderboard
- Status
- active
- Retrieved at
- 2026-10-28
- Methodology
- Official benchmark leaderboard. Exact leaderboard model row=DeepSeek-Coder-V2-Instruct; result split=Hard; prompt variant=Complete; model link=https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Instruct; prompted=True; prefill=False; source model-date field=2024-06-17. Official leaderboard reports calibrated Pass@1 using greedy decoding. Evaluation run date and other per-model generation settings are not stated. Dataset tasks=148.
- Context
- The leaderboard's JSON date field is retained as the model date, not an evaluation date. The site states calibrated Pass@1 using greedy decoding; it does not publish a run date or all per-model generation parameters. Complete and Instruct, and Full and Hard, are separate configurations.
BigCodeBench Hard (Instruct) (Show measurement, test conditions, and source)
- Source value
- 24.3
- Score
- 24.3
- Metric
- calibrated Pass@1 (greedy decoding)
- Unit
- %
- Category
- Hard / Instruct
- Direction
- Higher is better
- Sample
- 148
- Source type
- official-leaderboard
- Evaluator
- BigCodeBench official leaderboard
- Status
- active
- Retrieved at
- 2026-10-28
- Methodology
- Official benchmark leaderboard. Exact leaderboard model row=DeepSeek-Coder-V2-Instruct; result split=Hard; prompt variant=Instruct; model link=https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Instruct; prompted=True; prefill=False; source model-date field=2024-06-17. Official leaderboard reports calibrated Pass@1 using greedy decoding. Evaluation run date and other per-model generation settings are not stated. Dataset tasks=148.
- Context
- The leaderboard's JSON date field is retained as the model date, not an evaluation date. The site states calibrated Pass@1 using greedy decoding; it does not publish a run date or all per-model generation parameters. Complete and Instruct, and Full and Hard, are separate configurations.
HumanEval (Show measurement, test conditions, and source)
- Source value
- 90.2
- Score
- 90.2
- Metric
- score
- Unit
- No unit provided
- Benchmark version
- DeepSeek Coder V2 Instruct evaluation
- Category
- coding
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- DeepSeek Coder V2 official repository evaluation report
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact DeepSeek-Coder-V2-Instruct model row/widget in DeepSeek Coder V2 official repository evaluation report. Benchmark prompt, sampling/decoding settings, reasoning effort, quantization, and task run date are not stated in the cited result; none are inferred.
- Context
- The source presents a raw score without an explicit unit or scale; the published value is retained without percentage conversion.
MBPP+ (Show measurement, test conditions, and source)
- Source value
- 76.2
- Score
- 76.2
- Metric
- score
- Unit
- No unit provided
- Benchmark version
- DeepSeek Coder V2 Instruct evaluation
- Category
- coding
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- DeepSeek Coder V2 official repository evaluation report
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact DeepSeek-Coder-V2-Instruct model row/widget in DeepSeek Coder V2 official repository evaluation report. Benchmark prompt, sampling/decoding settings, reasoning effort, quantization, and task run date are not stated in the cited result; none are inferred.
- Context
- The source presents a raw score without an explicit unit or scale; the published value is retained without percentage conversion.
LiveCodeBench (Show measurement, test conditions, and source)
- Source value
- 43.4
- Score
- 43.4
- Metric
- score
- Unit
- No unit provided
- Benchmark version
- LiveCodeBench (report version not stated)
- Category
- coding
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- DeepSeek Coder V2 official repository evaluation report
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact DeepSeek-Coder-V2-Instruct model row/widget in DeepSeek Coder V2 official repository evaluation report. Benchmark prompt, sampling/decoding settings, reasoning effort, quantization, and task run date are not stated in the cited result; none are inferred.
- Context
- The source presents a raw score without an explicit unit or scale; the published value is retained without percentage conversion.
USACO (Show measurement, test conditions, and source)
- Source value
- 12.1
- Score
- 12.1
- Metric
- score
- Unit
- No unit provided
- Benchmark version
- DeepSeek Coder V2 Instruct evaluation
- Category
- coding
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- DeepSeek Coder V2 official repository evaluation report
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact DeepSeek-Coder-V2-Instruct model row/widget in DeepSeek Coder V2 official repository evaluation report. Benchmark prompt, sampling/decoding settings, reasoning effort, quantization, and task run date are not stated in the cited result; none are inferred.
- Context
- The source presents a raw score without an explicit unit or scale; the published value is retained without percentage conversion.
Defects4J (Show measurement, test conditions, and source)
- Source value
- 21
- Score
- 21
- Metric
- score
- Unit
- No unit provided
- Benchmark version
- DeepSeek Coder V2 Instruct evaluation
- Category
- code-fixing
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- DeepSeek Coder V2 official repository evaluation report
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact DeepSeek-Coder-V2-Instruct model row/widget in DeepSeek Coder V2 official repository evaluation report. Benchmark prompt, sampling/decoding settings, reasoning effort, quantization, and task run date are not stated in the cited result; none are inferred.
- Context
- The source presents a raw score without an explicit unit or scale; the published value is retained without percentage conversion.
SWE-Bench (Show measurement, test conditions, and source)
- Source value
- 12.7
- Score
- 12.7
- Metric
- score
- Unit
- No unit provided
- Benchmark version
- SWE-Bench (report version not stated)
- Category
- code-fixing
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- DeepSeek Coder V2 official repository evaluation report
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact DeepSeek-Coder-V2-Instruct model row/widget in DeepSeek Coder V2 official repository evaluation report. Benchmark prompt, sampling/decoding settings, reasoning effort, quantization, and task run date are not stated in the cited result; none are inferred.
- Context
- The source presents a raw score without an explicit unit or scale; the published value is retained without percentage conversion.
Aider (Show measurement, test conditions, and source)
- Source value
- 73.7
- Score
- 73.7
- Metric
- score
- Unit
- No unit provided
- Benchmark version
- DeepSeek Coder V2 Instruct evaluation
- Category
- code-fixing
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- DeepSeek Coder V2 official repository evaluation report
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact DeepSeek-Coder-V2-Instruct model row/widget in DeepSeek Coder V2 official repository evaluation report. Benchmark prompt, sampling/decoding settings, reasoning effort, quantization, and task run date are not stated in the cited result; none are inferred.
- Context
- The source presents a raw score without an explicit unit or scale; the published value is retained without percentage conversion.
Page 1 of 3
Popularity compared
How often the model is searched for, used and mentioned, compared with every model of its category.
Popularity index
The 100 most popular models of this category, 0 to 100, higher is more popular
20 of 101 models
As of November 2026 · Source: gradually.ai popularity index
Over time
Weekly values from October 18, 2026 to November 7, 2026
DeepSeek Coder V2Claude Haiku 4.5 32Holotron4-30B-A3B 32SOLAR 10.7B Instruct v1.0 31ZwZ2-4B 31
Method changes in the timeline, a jump at these points is no market event:
- 1November 1, 2026: Value only with search or usage data, 30-day attention
What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.
Similar models
More models of the same kind as DeepSeek Coder V2, chosen by model line, developer, size, price and release date.
- DeepSeek-V4-FlashDeepSeek · same developer
- DeepSeek-R1DeepSeek · same developer
- DeepSeek-V3DeepSeek · same developer
- DeepSeek V4 Flash 0731DeepSeek · same developer
- DeepSeek-V3 0324DeepSeek · same developer
- DeepSeek-R1-Distill-Llama-70BDeepSeek · same developer
- DeepSeek-R1 0528DeepSeek · same developer
- DeepSeek LLM 67BDeepSeek · same developer
All specifications
Architecture
Input and output
Watermarking and provenance
AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.
- AI watermarking
- Unknown
- no source
- Watermarking technique
- Not known
- no source
Licence and access
Frequently asked questions about DeepSeek Coder V2
How large is the context window of DeepSeek Coder V2?
131,072 tokens. That is how much text DeepSeek Coder V2 reads at once, about 200 A4 pages.
Page count as a rule of thumb from Artificial Analysis
Can I run DeepSeek Coder V2 locally?
Yes, the weights are open. With 236B parameters it needs a lot of memory, more a server than a laptop.
Stated by DeepSeek
- Open weights
- yes
- License
- MIT
- Parameters
- 236B
- active per token
- 21B
Sources and data date
Show all 21 sources
Benchmark sources at a glance
Which independent evaluators tested this model. This is coverage, not a quality ranking.
No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.
Without vendor claims and historical or disputed results; those stay labelled in the single values.
| Type | Evidence and data date |
|---|---|
| Input | Documented modalities. |
| Output | Documented modalities. |
| Research status | Research date October 15, 2026. 16 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | units-and-run-configuration Note The repository labels model parameter counts and benchmark names but does not specify score units, prompt/shot protocol, inference settings, quantization, or exact benchmark run dates; the AIME table explicitly reports 4 of 30. |
| Unresolved | MMLU-Pro measurement metadata Note The exact card shows the TIGER-Lab/MMLU-Pro label, adjacent text “Mmlu Pro”, and score 63.63. The adjacent text is not a metric definition; the captured source does not give dataset revision, task split, metric definition, sample size, confidence interval, unit/scale, direction, or evaluation settings. |
| Unresolved | task split, metric unit, denominator, run configuration, and confidence interval for Section 3.5 results Original wording The official repository identifies the 236B Instruct checkpoint and numeric table cells but does not supply the score scales, exact metrics, prompts, harness, sample counts, or confidence intervals. |
| Not found in the inspected sources | additional exact-checkpoint benchmark cells beyond the frozen official and 2026 paper rows Original wording The focused pass found no further exact-profile scores. Base, Lite, and other-size observations are not transferred. |
| Unresolved | Per-row metric scale and higher-or-lower-is-better direction for the ten closed-ended results Note The paper provides EM/Acc. labels but not row-level formulas or an explicit direction. Its opening summary marks MMLU as 79.2%; that wording does not establish a percent unit for the other nine rows. |
| Not published | Completed evaluation sample count and confidence intervals for Section 4.6 Note The report gives no evaluation N or confidence interval for these rows. The stated training-token total is not an evaluation denominator. |
| Unresolved | Exact inherited DeepSeek-V2 and MMLU simple-evals revision and run configuration Note The paper names both evaluation paths but does not provide a pinned commit, command, prompt manifest, or run settings for the reported rows. |
| Unresolved | Exact judge version and scoring setup for the four open-ended rows Note The report names GPT-4 judging for three rows but gives no snapshot or judge protocol. It does not identify the AlpacaEval score variant or a row-level run configuration. |
| Unresolved | Official MT-Bench evaluator documentation Note The attempted official FastChat judge-documentation path was unavailable to the fetcher; no additional MT-Bench protocol is attributed to this run. |
| Unresolved | conflicting-observations-benchmark-methodology Original wording Conflicting values were recorded for benchmark-methodology in the same source and retrieval date (https://arxiv.org/pdf/2406.11931). The separate observations remain visible; no value was silently chosen. |
| Additional source | DeepSeek Coder V2 official repository evaluation report (retrieved October 15, 2026) |
| Additional source | Complexity-Aware Evaluation of LLM Comprehension, arXiv:2609.37405v1 (retrieved October 20, 2026) |
| Additional source | DeepSeek-Coder-V2 official repository evaluation results, Section 3.5 (retrieved October 24, 2026; October 25, 2026) · DeepSeek Coder V2 repository · Editorial description reviewed October 2, 2026 |
| Additional source | BigCodeBench official leaderboard (retrieved October 28, 2026)12 |
| Additional source | Model metadata (retrieved October 25, 2026)12345 |
| Additional source | Model metadata (retrieved October 28, 2026)1234 |
| Additional source | OpenRouter: Popularity index (derived), CC BY 4.0 |
| Additional source | Vercel AI Gateway: Popularity index (derived), CC BY 4.0 |
| Additional source | Ollama: cumulative downloads for the popularity index |
| Additional source | Replicate: cumulative runs for the popularity index |
| Additional source | Google Trending Now: limited feed selection for Germany and the US |
| Additional source | Creative Commons Attribution 4.0 |