Skip to content

Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.

fix: add DeepSeek V4/V3.2/R1 tokenizer mappings and context limits (#912)

yiihao0c7087539d92
diff --git a/headroom/providers/openai.py b/headroom/providers/openai.py
index 320fb59e6..3ebd0c549 100644
--- a/headroom/providers/openai.py
+++ b/headroom/providers/openai.py
@@ -99,9 +99,16 @@ _CONTEXT_LIMITS: dict[str, int] = {
# these are the manual fallback when LiteLLM doesn't know the model.
"deepseek-v4-flash": 1_000_000,
"deepseek-v4-pro": 1_000_000,
+ "deepseek-v3.2": 128_000,
+ "deepseek-v3-0324": 128_000,
+ "deepseek-v3": 128_000,
+ "deepseek-v2": 128_000,
"deepseek-chat": 131_072,
"deepseek-reasoner": 131_072,
+ "deepseek-r1": 131_072,
+ "deepseek-r1-0528": 131_072,
"deepseek-coder": 16384,
+ "deepseek-coder-v2": 128_000,
}
# Fallback pricing - LiteLLM is preferred source
diff --git a/headroom/providers/openai_compatible.py b/headroom/providers/openai_compatible.py
index 8a590d51c..32c88db70 100644
--- a/headroom/providers/openai_compatible.py
+++ b/headroom/providers/openai_compatible.py
@@ -86,9 +86,22 @@ _DEFAULT_CONTEXT_LIMITS: dict[str, int] = {
# DeepSeek
"deepseek": 1048576,
"deepseek-coder": 128000,
+ "deepseek-coder-v2": 128000,
+ "deepseek-coder-v2-lite": 128000,
"deepseek-v2": 128000,
+ "deepseek-v2-lite": 128000,
"deepseek-v3": 1048576,
+ "deepseek-v3-0324": 128000,
+ "deepseek-v3.2": 128000,
+ "deepseek-r1": 131072,
+ "deepseek-r1-0528": 131072,
+ "deepseek-reasoner": 131072,
"deepseek-v4": 1048576,
+ "deepseek-v4-pro": 1_000_000,
+ "deepseek-v4-flash": 1_000_000,
+ "deepseek-chat": 131072,
+ "deepseek-r1-distill-qwen": 131072,
+ "deepseek-r1-distill-llama": 131072,
# Yi
"yi": 32768,
"yi-34b": 32768,
diff --git a/headroom/tokenizers/huggingface.py b/headroom/tokenizers/huggingface.py
index c48c16474..750971f33 100644
--- a/headroom/tokenizers/huggingface.py
+++ b/headroom/tokenizers/huggingface.py
@@ -63,13 +63,30 @@ MODEL_TO_TOKENIZER: dict[str, str] = {
"qwen2.5": "Qwen/Qwen2.5-7B",
"qwen2.5-7b": "Qwen/Qwen2.5-7B",
"qwen2.5-72b": "Qwen/Qwen2.5-72B",
- # DeepSeek
+ # DeepSeek V1 / Coder (legacy, 2023-2024)
"deepseek": "deepseek-ai/deepseek-llm-7b-base",
"deepseek-7b": "deepseek-ai/deepseek-llm-7b-base",
"deepseek-67b": "deepseek-ai/deepseek-llm-67b-base",
"deepseek-coder": "deepseek-ai/deepseek-coder-6.7b-base",
+ "deepseek-coder-v2": "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct",
+ "deepseek-coder-v2-lite": "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct",
+ # DeepSeek V2/V3 family (2024)
"deepseek-v2": "deepseek-ai/DeepSeek-V2",
+ "deepseek-v2-lite": "deepseek-ai/DeepSeek-V2-Lite",
"deepseek-v3": "deepseek-ai/DeepSeek-V3",
+ "deepseek-v3-0324": "deepseek-ai/DeepSeek-V3-0324",
+ "deepseek-v3.2": "deepseek-ai/DeepSeek-V3.2",
+ # DeepSeek R1 reasoning family (2025)
+ "deepseek-r1": "deepseek-ai/DeepSeek-R1",
+ "deepseek-r1-0528": "deepseek-ai/DeepSeek-R1-0528",
+ "deepseek-reasoner": "deepseek-ai/DeepSeek-R1",
+ # DeepSeek V4 family (2025-2026)
+ "deepseek-v4-pro": "deepseek-ai/DeepSeek-V4-Pro",
+ "deepseek-v4-flash": "deepseek-ai/DeepSeek-V4-Flash",
+ # DeepSeek API aliases (routed through the proxy)
+ "deepseek-chat": "deepseek-ai/DeepSeek-V3",
+ "deepseek-r1-distill-qwen": "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B",
+ "deepseek-r1-distill-llama": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B",
# Yi family
"yi": "01-ai/Yi-6B",
"yi-6b": "01-ai/Yi-6B",
diff --git a/tests/test_huggingface_tokenizer_timeout.py b/tests/test_huggingface_tokenizer_timeout.py
index 8d7e8b888..22b9e98ac 100644
--- a/tests/test_huggingface_tokenizer_timeout.py
+++ b/tests/test_huggingface_tokenizer_timeout.py
@@ -113,6 +113,14 @@ def test_count_messages_fails_open_to_estimation(monkeypatch: pytest.MonkeyPatch
assert tokens > 0 # estimation fallback, no exception, no hang
+def test_deepseek_model_aliases_resolve_to_expected_tokenizers() -> None:
+ assert get_tokenizer_name("deepseek-v3.2") == "deepseek-ai/DeepSeek-V3.2"
+ assert get_tokenizer_name("deepseek-v4-pro") == "deepseek-ai/DeepSeek-V4-Pro"
+ assert get_tokenizer_name("deepseek-v4-flash") == "deepseek-ai/DeepSeek-V4-Flash"
+ assert get_tokenizer_name("deepseek-r1") == "deepseek-ai/DeepSeek-R1"
+ assert get_tokenizer_name("deepseek-r1-0528") == "deepseek-ai/DeepSeek-R1-0528"
+
+
def test_invalid_timeout_env_falls_back_to_default(monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.setenv("HEADROOM_HF_TOKENIZER_LOAD_TIMEOUT_SECS", "not-a-number")
assert hf_mod._load_timeout_secs() == hf_mod._LOAD_TIMEOUT_DEFAULT
diff --git a/tests/test_providers/test_universal.py b/tests/test_providers/test_universal.py
index 9f34e622e..26190ed46 100644
--- a/tests/test_providers/test_universal.py
+++ b/tests/test_providers/test_universal.py
@@ -88,6 +88,11 @@ class TestOpenAICompatibleProvider:
assert provider.get_context_limit("deepseek-v4") == 1048576
assert provider.get_context_limit("deepseek") == 1048576
assert provider.get_context_limit("deepseek-v2") == 128000
+ assert provider.get_context_limit("deepseek-v3.2") == 128000
+ assert provider.get_context_limit("deepseek-v4-pro") == 1_000_000
+ assert provider.get_context_limit("deepseek-v4-flash") == 1_000_000
+ assert provider.get_context_limit("deepseek-r1") == 131072
+ assert provider.get_context_limit("deepseek-coder-v2") == 128000
def test_get_context_limit_unknown_model(self):
"""Test context limit for unknown models (defaults to 128K)."""