aboutsummaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-06-14 16:20:52 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-06-14 16:41:53 +0200
commit067ba8cc5673896f7c09866d0b9587ff7fe4b9f7 (patch)
tree7f4cbbbea949d865ad45dad33476b851923e61c0
parentcb1f049e710cf8704ae29811b3058142c7f64840 (diff)
downloadffwd-067ba8cc5673896f7c09866d0b9587ff7fe4b9f7.tar.gz
ffwd-067ba8cc5673896f7c09866d0b9587ff7fe4b9f7.zip
server: add Qwen3-Embedding-8B model slot
-rw-r--r--README.md4
-rw-r--r--embed_server.c3
2 files changed, 5 insertions, 2 deletions
diff --git a/README.md b/README.md
index f056d40..8427674 100644
--- a/README.md
+++ b/README.md
@@ -83,8 +83,8 @@ Late-interaction reranking endpoint:
For Qwen3-Embedding models, `/v1/embeddings` accepts `dimensions` from 32 up to
the model's output size (1024 for `Qwen3-Embedding-0.6B`, 2560 for
-`Qwen3-Embedding-4B`). Truncated Matryoshka embeddings are re-normalized before
-encoding.
+`Qwen3-Embedding-4B`, 4096 for `Qwen3-Embedding-8B`). Truncated Matryoshka
+embeddings are re-normalized before encoding.
Example:
diff --git a/embed_server.c b/embed_server.c
index f22a99c..4c40afb 100644
--- a/embed_server.c
+++ b/embed_server.c
@@ -207,6 +207,7 @@ typedef enum {
MODEL_STD_4,
MODEL_QWEN3_06,
MODEL_QWEN3_4,
+ MODEL_QWEN3_8,
MODEL_CTX_06,
MODEL_CTX_4,
MODEL_LATE_06,
@@ -241,6 +242,8 @@ static const model_info k_models[] = {
EMBED_ATTENTION_CAUSAL, EMBED_POOL_LAST_TOKEN, 1},
{"Qwen3-Embedding-4B", MODEL_KIND_STANDARD, 2560, 32, 0, 0.0,
EMBED_ATTENTION_CAUSAL, EMBED_POOL_LAST_TOKEN, 1},
+ {"Qwen3-Embedding-8B", MODEL_KIND_STANDARD, 4096, 32, 0, 0.0,
+ EMBED_ATTENTION_CAUSAL, EMBED_POOL_LAST_TOKEN, 1},
{"pplx-embed-context-v1-0.6b", MODEL_KIND_CONTEXTUAL, 1024, 128, 0, 0.008,
EMBED_ATTENTION_BIDIRECTIONAL, EMBED_POOL_MEAN, 0},
{"pplx-embed-context-v1-4b", MODEL_KIND_CONTEXTUAL, 2560, 128, 0, 0.050,