whatmodel.sh / DeepSeek-V2.5
DeepSeek-V2.5
↗ share
+ contribute
DeepSeek-V2.5 is an open-weights language model from DeepSeek, released September 2024. Its published size is 236B parameters (21B active per token) and a 164k-token context window. Licence: deepseek.
Catalogue identity notes
Arena entry 'deepseek-v2.5' (text). LMArena's leaderboard table (arena-leaderboard space @6f2eeeb) labels 'deepseek-v2.5' as 'Deepseek-v2.5' (DeepSeek, DeepSeek), linking https://huggingface.co/deepseek-ai/DeepSeek-V2.5. Maker repo deepseek-ai/DeepSeek-V2.5 @c85b5ed: licence tag other (deepseek). Parameters: DeepSeek-V2 architecture: 236B total / 21B active (HF safetensors metadata 235.7B). Released 2024-09: month the maker's Hugging Face repo was created (createdAt 2024-09-05). September 2024 merge of DeepSeek-V2-Chat and DeepSeek-Coder-V2-Instruct; distinct from DeepSeek-V2.5-1210.
$ how it compares
The LLMs ranked either side of DeepSeek-V2.5, out of 828 with a score, then more from DeepSeek.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 345 |  Claude 3.5 Haiku | 54±1 | comparison only | ||
| 346 |  Llama 3.1 Tulu 3 70B | 54±1 | local | ||
| 347 | DeepSeek-V2.5 | 54±1 | local | ||
| 348 | DeepSeek-V2 Chat (0628) | 54±1 | local | ||
| 349 |  GLM-4 9B Chat 1M | 53±5 | local | ||
| more from DeepSeek | |||||
| 21 | DeepSeek V4.1 Flash | 78±1 | api $0.26/M +local | ||
| 31 | DeepSeek V4 Pro (0813) | 76±1 | api $1.63/M +local | ||
| 35 | DeepSeek V3.2 Speciale | 76±3 | comparison only | ||
$ what it's good at
coding▁▁▁▃▆██▇▆▄▃▁ #188 of 458 · 31% of its score · 4 tests
Beats Gemini 2.5 Flash, Qwen3.5 4B, Qwen3 30B-A3B Instruct 2507 here, though they rank higher overall. Trails QwQ-32B-Preview here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| CRUXEval (ZeroEval, 800 problems) | 70.0% ✎ | ▁▂▆▃▆█▆▃▂·▁▁ beats 29/47 | 14% |
| Aider code editing (133 Exercism Python exercises, two attempts) | 72.2% ✎ | ▁·▁▂▄▁▆▄█▆▂▂ beats 54/62 | 11% |
| WebApp1K-Duo-React pass@1 (onekq-ai) | 49.0% ✎ | ▁··▁·▁▁▃▆█▇█ beats 6/45 | 3.0% |
| WebApp1K-React pass@1 (onekq-ai) | 83.4% ✎ | ▃▁▁█▁▄▅▃▄▇▁▄ beats 22/31 | 2.8% |
chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #214 of 393 * · 8.9% of its score · 2 tests
Beats Granite 4.2 8B, Mistral Small 4 (119B-A6.5B), Gemma 2 9B IT SimPO here, though they rank higher overall. Trails LFM2.5 1.2B Instruct, Granite 4.0 H Micro 3B, Gemma 3n E4B here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| LMArena (text) | 1307.2 ✎ | ▁▁▁▂▃▃▃▆▆▇█▃ beats 137/379 | 5.9% |
| IFEval prompt-level strict | 77.6% ✎ | ▄▄··▄·▄··█·▄ beats 0/6 | 3.0% |
math▁▂▃▅▆▅▄█▄▁▁▁ #270 of 398 · 16% of its score · 3 tests
Beats Molmo 2 8B here, though they rank higher overall. Trails DeepSeek R1 Distill Qwen 1.5B, Mixtral 8x7B v0.1 (base), InternVL2-4B here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| MATH Level 5 (ZeroEval, 721 problems) | 44.7% ✎ | ▁▁▁▆█▂▇▇▄··▂ beats 22/38 | 8.9% |
| GSM8K (ZeroEval, zero-shot) | 92.5% ✎ | ▁▄·▁▄█▅▅▅███ beats 26/43 | 4.7% |
| MATH-500 lab | 74.8% ✎ | ▁▃▂▃▅▇▅▇█▅▂▁ beats 17/84 | 2.5% |
other tests4 not in a task · 44% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| MMLU-Redux (ZeroEval, 2,778 questions) | 80.3% ✎ | ▁·▁▁▄▄▄▅█▄▂▁ beats 29/53 | 22% |
| ZebraLogic (ZeroEval, 1,000 puzzles) | 22.1% ✎ | ▁▁▃▆██▁▁·▁▁▁ beats 34/57 | 12% |
| MMLU-Pro | 65.8% ✎ | ▃▂▅█▃·█▃▅··▂ beats 17/23 | 10% |
| DROP F1, 3-shot (simple-evals) | 81.4% ✎ | ▅▂▆█▆·▃·▂▅·▂ beats 9/23 | 0.3% |
$ where to run it
llama.cpp · recommended
version
- Install llama.cpp · Homebrew and winget installs name the commands llama-cli and llama-server instead.macOS / Linux
curl -LsSf https://llama.app/install.sh | shWindowsirm https://llama.app/install.ps1 | iex - Run DeepSeek-V2.5 · needs ~142.5 GB · from huggingface.co/lmstudio-communitychat in the terminal
llama cli -hf lmstudio-community/DeepSeek-V2.5-GGUF:Q4_K_M - Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf lmstudio-community/DeepSeek-V2.5-GGUF:Q4_K_M
LM Studio
version
- Install LM Studio · Open the app once before using the lms command.
- Run DeepSeek-V2.5 · needs ~142.5 GB · from huggingface.co/lmstudio-community
- Or download it in the terminal
lms get https://huggingface.co/lmstudio-community/DeepSeek-V2.5-GGUF@Q4_K_M -y - Then run lms ls to see its name, and lms chat followed by that name
Ollama
version
- Install Ollama
download it (macOS / Windows) ↗
Linuxcurl -fsSL https://ollama.com/install.sh | sh - Run DeepSeek-V2.5 · needs ~142.1 GB · from ollama.comchat in the terminal
ollama run deepseek-v2.5:236b-q4_0
Weights and documentation: huggingface.co/deepseek-ai ↗
$ how the score was worked out · WildEval ZeroEval published r… 62%, Aider code editing benchmark… 11%, MMLU-Pro author-released eval… 10% and 4 more
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
| source | results | share of the score |
|---|---|---|
 WildEval ZeroEval published results | 5 | 62% |
 Aider code editing benchmark runs (old edit leaderboard) | 1 | 11% |
| MMLU-Pro author-released evaluation runs | 1 | 10% |
 LMArena leaderboard dataset | 1 | 6% |
| WebApp1K models leaderboard (onekq-ai) | 2 | 6% |
| OpenCompass CompassAcademic leaderboard (full version) Space: committed predictions file | 2 | 3% |
| DeepSeek vendor-reported benchmark scores for its own models (Hugging Face model cards) · lab-reported | 1 | 2% |
| prior · pull toward the middle | 0% | |
| weighted average | = 53.6 |
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json