Add Qwen 2.5 Coder 7B (GGUF Q4_K_M) model support
- Adds model constants and metadata for Qwen 2.5 Coder 7B Instruct - Implements GgufModelConfig::qwen25_coder_7b() constructor - Registers model in SUPPORTED_MODELS and SUPPORTED_MODEL_INFO - Enables tokenizer mapping for Android
keypair34 committed
Apr 26, 2026 at 07:11 UTC
d23c8cb7d4f7f3aaf2ca820db8f4e9c21b822cb3
2 files changed
+41
src/inference/engine.rs
+19
@@ -1956,6 +1956,25 @@ impl GgufModelConfig {
1956
}
1957
}
1958
1959
+ /// Qwen 2.5 Coder 7B Instruct (GGUF Q4_K_M) — ~4.4 GB.
1960
+ ///
1961
+ /// Strong coding model with tool/function calling support via the Qwen2.5
1962
+ /// chat template. Uses the `qwen2` architecture. Requires 8+ GB RAM.
1963
+ pub fn qwen25_coder_7b() -> Self {
1964
+ Self {
1965
+ model_id: super::models::BARTOWSKI_QWEN25_CODER_7B_INSTRUCT_GGUF.into(),
1966
+ files: vec![super::models::QWEN25_CODER_7B_GGUF_FILE.into()],
1967
+ tok_model_id: if cfg!(target_os = "android") {
1968
+ Some(super::models::QWEN25_CODER_7B_TOK_MODEL_ID.into())
1969
+ } else {
1970
+ None
1971
+ },
1972
+ display_name: "Qwen 2.5 Coder 7B (Q4_K_M)".into(),
1973
+ approx_memory: "~4.4 GB".into(),
1974
+ chat_template: None,
1975
+ }
1976
+ }
1977
+
1978
/// Qwen 3 4B Instruct (GGUF Q4_K_M) — ~2.7 GB.
1979
///
1980
/// Full OpenAI-compatible tool calling with extended thinking mode.
src/inference/models.rs
+22
@@ -52,6 +52,18 @@ pub const QWEN25_CODER_3B_GGUF_FILE: &str = "Qwen2.5-Coder-3B-Instruct-Q4_K_M.gg
52
/// Base model repo used for the HF tokenizer on Android.
53
pub const QWEN25_CODER_3B_TOK_MODEL_ID: &str = "Qwen/Qwen2.5-Coder-3B-Instruct";
54
55
+/// Pre-quantized Qwen 2.5 Coder 7B Instruct (GGUF Q4_K_M) — strong coding model with tool calling (~4.4 GB).
56
+///
57
+/// Uses the `qwen2` GGUF architecture. Trained on 5.5T tokens of code/math data with
58
+/// fill-in-the-middle and repo-level code understanding. Supports function/tool calling
59
+/// via the Qwen2.5 chat template. Requires ~8 GB RAM.
60
+pub const BARTOWSKI_QWEN25_CODER_7B_INSTRUCT_GGUF: &str =
61
+ "bartowski/Qwen2.5-Coder-7B-Instruct-GGUF";
62
+/// The specific GGUF filename to download from the bartowski Coder 7B repo.
63
+pub const QWEN25_CODER_7B_GGUF_FILE: &str = "Qwen2.5-Coder-7B-Instruct-Q4_K_M.gguf";
64
+/// Base model repo used for the HF tokenizer on Android.
65
+pub const QWEN25_CODER_7B_TOK_MODEL_ID: &str = "Qwen/Qwen2.5-Coder-7B-Instruct";
66
+
67
/// Pre-quantized Qwen 2.5 3B Instruct (GGUF Q4_K_M) — balanced option (~1.93 GB).
68
/// Ideal for macOS desktops and Android devices with sufficient RAM.
69
/// Not recommended as default on iOS: the 3B variant caused OOM on iPhone 16e (8 GB RAM)
@@ -110,6 +122,7 @@ pub const SUPPORTED_MODELS: &[&str] = &[
122
BARTOWSKI_QWEN3_4B_GGUF,
123
BARTOWSKI_QWEN3_8B_GGUF,
124
BARTOWSKI_QWEN3_1_7B_GGUF,
125
+ BARTOWSKI_QWEN25_CODER_7B_INSTRUCT_GGUF,
126
THEBLOKE_DEEPSEEK_CODER_6_7B_INSTRUCT_GGUF,
127
];
128
@@ -184,6 +197,14 @@ pub const SUPPORTED_MODEL_INFO: &[SupportedModelInfo] = &[
197
Smallest Qwen 3 variant with tool calling support.",
198
expected_size_bytes: 1_282_439_584,
199
},
200
+ SupportedModelInfo {
201
+ id: BARTOWSKI_QWEN25_CODER_7B_INSTRUCT_GGUF,
202
+ name: "Qwen 2.5 Coder 7B (GGUF)",
203
+ org: "Qwen / Alibaba",
204
+ description: "Strong coding model with tool calling support (~4.4 GB). \
205
+ Trained on 5.5T code tokens. Requires 8+ GB RAM.",
206
+ expected_size_bytes: 4_683_074_336,
207
+ },
208
SupportedModelInfo {
209
id: THEBLOKE_DEEPSEEK_CODER_6_7B_INSTRUCT_GGUF,
210
name: "DeepSeek Coder 6.7B (GGUF)",
@@ -206,6 +227,7 @@ pub fn tok_model_id_for_repo(hf_repo_id: &str) -> Option<&'static str> {
227
BARTOWSKI_QWEN25_3B_INSTRUCT_GGUF => Some(QWEN25_3B_TOK_MODEL_ID),
228
BARTOWSKI_QWEN25_CODER_1_5B_INSTRUCT_GGUF => Some(QWEN25_CODER_1_5B_TOK_MODEL_ID),
229
BARTOWSKI_QWEN25_CODER_3B_INSTRUCT_GGUF => Some(QWEN25_CODER_3B_TOK_MODEL_ID),
230
+ BARTOWSKI_QWEN25_CODER_7B_INSTRUCT_GGUF => Some(QWEN25_CODER_7B_TOK_MODEL_ID),
231
THEBLOKE_DEEPSEEK_CODER_6_7B_INSTRUCT_GGUF => Some(DEEPSEEK_CODER_6_7B_TOK_MODEL_ID),
232
_ => None,
233
}