jake/Lumotia
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(llm): bump model registry to Qwen3.5 + Qwen3.6 family (4 tiers)

Browse Source 
Replaces the three older Qwen3 variants with a four-tier ladder spanning
a wider hardware range:

- Qwen3_5_2B_Q4   (Minimal, 8 GB RAM, ~1.3 GB download)
- Qwen3_5_4B_Q4   (Standard, 16 GB RAM / 6 GB VRAM, ~2.7 GB) — DEFAULT
- Qwen3_5_9B_Q4   (High, 32 GB RAM / 12 GB VRAM, ~5.7 GB)
- Qwen3_6_27B_Q4  (Maximum, 64 GB RAM / 24 GB VRAM, ~17 GB)

All four GGUFs sourced from unsloth's HF org with pinned commit SHAs.
Sizes and SHA256 hashes verified against the live X-Linked-Etag /
X-Linked-Size headers on the LFS CDN. Q4_K_M quantisation throughout
(common sweet-spot for cleanup + task extraction).

recommend_tier rewritten to span four bands; default_tier moves from
the old 4B-Instruct-2507 to Qwen3.5 4B. The 27B Maximum tier honestly
needs 64 GB RAM to run without partial offload — surfaced in the
description string so the Settings UI can warn realistically.

In-tree smoke tests (smoke.rs, content_tags_smoke.rs) updated to
reference the new smallest tier so a developer's MAGNOTIA_LLM_TEST_MODEL
points at the cheapest GGUF to download. Crate description in
crates/llm/Cargo.toml refreshed to mention the new family.

NOTE (out of scope; not fixed): the size_bytes / sha256 / hf_url
methods could collapse into a single LlmModelMetadata table to remove
four parallel match arms. Layer 2 cleanup, separate session.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Jake
2026-05-01 09:57:21 +01:00
parent c42a144aad
commit 699cb7e08e
4 changed files with 81 additions and 61 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
crates/llm/Cargo.toml
View File

@@ -2,7 +2,7 @@
name = "magnotia-llm" name = "magnotia-llm"
version = "0.1.0" version = "0.1.0"
edition = "2021" edition = "2021"
description = "Local LLM engine for Magnotia (Qwen3 via llama-cpp-2): transcript cleanup, task extraction, micro-step decomposition" description = "Local LLM engine for Magnotia (Qwen3.5 / Qwen3.6 via llama-cpp-2): transcript cleanup, task extraction, micro-step decomposition"
[features] [features]
# Default desktop build keeps the existing openmp + vulkan acceleration. # Default desktop build keeps the existing openmp + vulkan acceleration.

136
crates/llm/src/model_manager.rs
View File

@@ -12,100 +12,119 @@ use tokio::io::{AsyncReadExt, AsyncWriteExt};
#[allow(non_camel_case_types)] #[allow(non_camel_case_types)]
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Serialize, Deserialize)] #[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Serialize, Deserialize)]
pub enum LlmModelId { pub enum LlmModelId {
#[serde(rename = "qwen3_1_7b")] #[serde(rename = "qwen3_5_2b")]
Qwen3_1_7B_Q4, Qwen3_5_2B_Q4,
#[serde(rename = "qwen3_4b_instruct_2507")] #[serde(rename = "qwen3_5_4b")]
Qwen3_4BInstruct2507Q4, Qwen3_5_4B_Q4,
#[serde(rename = "qwen3_14b")] #[serde(rename = "qwen3_5_9b")]
Qwen3_14BQ5, Qwen3_5_9B_Q4,
#[serde(rename = "qwen3_6_27b")]
Qwen3_6_27B_Q4,
} }
impl LlmModelId { impl LlmModelId {
pub fn default_tier() -> Self { pub fn default_tier() -> Self {
Self::Qwen3_4BInstruct2507Q4 Self::Qwen3_5_4B_Q4
} }
pub fn as_str(&self) -> &'static str { pub fn as_str(&self) -> &'static str {
match self { match self {
Self::Qwen3_1_7B_Q4 => "qwen3_1_7b", Self::Qwen3_5_2B_Q4 => "qwen3_5_2b",
Self::Qwen3_4BInstruct2507Q4 => "qwen3_4b_instruct_2507", Self::Qwen3_5_4B_Q4 => "qwen3_5_4b",
Self::Qwen3_14BQ5 => "qwen3_14b", Self::Qwen3_5_9B_Q4 => "qwen3_5_9b",
Self::Qwen3_6_27B_Q4 => "qwen3_6_27b",
} }
} }
pub fn display_name(&self) -> &'static str { pub fn display_name(&self) -> &'static str {
match self { match self {
Self::Qwen3_1_7B_Q4 => "Qwen3 1.7B", Self::Qwen3_5_2B_Q4 => "Qwen3.5 2B",
Self::Qwen3_4BInstruct2507Q4 => "Qwen3 4B Instruct 2507", Self::Qwen3_5_4B_Q4 => "Qwen3.5 4B",
Self::Qwen3_14BQ5 => "Qwen3 14B", Self::Qwen3_5_9B_Q4 => "Qwen3.5 9B",
Self::Qwen3_6_27B_Q4 => "Qwen3.6 27B",
} }
} }
pub fn file_name(&self) -> &'static str { pub fn file_name(&self) -> &'static str {
match self { match self {
Self::Qwen3_1_7B_Q4 => "Qwen3-1.7B-Q4_K_M.gguf", Self::Qwen3_5_2B_Q4 => "Qwen3.5-2B-Q4_K_M.gguf",
Self::Qwen3_4BInstruct2507Q4 => "Qwen3-4B-Instruct-2507-Q4_K_M.gguf", Self::Qwen3_5_4B_Q4 => "Qwen3.5-4B-Q4_K_M.gguf",
Self::Qwen3_14BQ5 => "Qwen3-14B-Q5_K_M.gguf", Self::Qwen3_5_9B_Q4 => "Qwen3.5-9B-Q4_K_M.gguf",
Self::Qwen3_6_27B_Q4 => "Qwen3.6-27B-Q4_K_M.gguf",
} }
} }
pub fn size_bytes(&self) -> u64 { pub fn size_bytes(&self) -> u64 {
match self { match self {
Self::Qwen3_1_7B_Q4 => 1_107_409_472, Self::Qwen3_5_2B_Q4 => 1_280_835_840,
Self::Qwen3_4BInstruct2507Q4 => 2_497_281_120, Self::Qwen3_5_4B_Q4 => 2_740_937_888,
Self::Qwen3_14BQ5 => 10_514_570_624, Self::Qwen3_5_9B_Q4 => 5_680_522_464,
Self::Qwen3_6_27B_Q4 => 16_817_244_384,
} }
} }
pub fn minimum_ram_bytes(&self) -> u64 { pub fn minimum_ram_bytes(&self) -> u64 {
match self { match self {
Self::Qwen3_1_7B_Q4 => 8 * 1024_u64.pow(3), Self::Qwen3_5_2B_Q4 => 8 * 1024_u64.pow(3),
Self::Qwen3_4BInstruct2507Q4 => 16 * 1024_u64.pow(3), Self::Qwen3_5_4B_Q4 => 16 * 1024_u64.pow(3),
Self::Qwen3_14BQ5 => 32 * 1024_u64.pow(3), Self::Qwen3_5_9B_Q4 => 32 * 1024_u64.pow(3),
Self::Qwen3_6_27B_Q4 => 64 * 1024_u64.pow(3),
} }
} }
pub fn recommended_vram_bytes(&self) -> Option<u64> { pub fn recommended_vram_bytes(&self) -> Option<u64> {
match self { match self {
Self::Qwen3_1_7B_Q4 => None, Self::Qwen3_5_2B_Q4 => None,
Self::Qwen3_4BInstruct2507Q4 => Some(8 * 1024_u64.pow(3)), Self::Qwen3_5_4B_Q4 => Some(6 * 1024_u64.pow(3)),
Self::Qwen3_14BQ5 => Some(16 * 1024_u64.pow(3)), Self::Qwen3_5_9B_Q4 => Some(12 * 1024_u64.pow(3)),
Self::Qwen3_6_27B_Q4 => Some(24 * 1024_u64.pow(3)),
} }
} }
pub fn description(&self) -> &'static str { pub fn description(&self) -> &'static str {
match self { match self {
Self::Qwen3_1_7B_Q4 => "Low tier for 8 GB RAM and CPU-heavy machines.", Self::Qwen3_5_2B_Q4 => "Minimal tier for 8 GB RAM and CPU-heavy machines.",
Self::Qwen3_4BInstruct2507Q4 => { Self::Qwen3_5_4B_Q4 => {
"Default tier for cleanup and task extraction on 16 GB systems." "Standard tier for cleanup and task extraction on 16 GB systems."
}
Self::Qwen3_5_9B_Q4 => "High tier for 32 GB RAM with a 12 GB+ GPU.",
Self::Qwen3_6_27B_Q4 => {
"Maximum tier for 64 GB RAM with a 24 GB GPU; partial CPU offload below that."
} }
Self::Qwen3_14BQ5 => "High tier for 32 GB+ RAM and larger GPUs.",
} }
} }
pub fn hf_url(&self) -> &'static str { pub fn hf_url(&self) -> &'static str {
match self { match self {
Self::Qwen3_1_7B_Q4 => { Self::Qwen3_5_2B_Q4 => {
"https://huggingface.co/unsloth/Qwen3-1.7B-GGUF/resolve/d7f544eead698dbd1f15126ef60b45a1e1933222/Qwen3-1.7B-Q4_K_M.gguf" "https://huggingface.co/unsloth/Qwen3.5-2B-GGUF/resolve/f6d5376be1edb4d416d56da11e5397a961aca8ae/Qwen3.5-2B-Q4_K_M.gguf"
} }
Self::Qwen3_4BInstruct2507Q4 => { Self::Qwen3_5_4B_Q4 => {
"https://huggingface.co/unsloth/Qwen3-4B-Instruct-2507-GGUF/resolve/a06e946bb6b655725eafa393f4a9745d460374c9/Qwen3-4B-Instruct-2507-Q4_K_M.gguf" "https://huggingface.co/unsloth/Qwen3.5-4B-GGUF/resolve/e87f176479d0855a907a41277aca2f8ee7a09523/Qwen3.5-4B-Q4_K_M.gguf"
} }
Self::Qwen3_14BQ5 => { Self::Qwen3_5_9B_Q4 => {
"https://huggingface.co/unsloth/Qwen3-14B-GGUF/resolve/a04a82c4739b3ef5fa6da7d10261db2c67dd1985/Qwen3-14B-Q5_K_M.gguf" "https://huggingface.co/unsloth/Qwen3.5-9B-GGUF/resolve/3885219b6810b007914f3a7950a8d1b469d598a5/Qwen3.5-9B-Q4_K_M.gguf"
}
Self::Qwen3_6_27B_Q4 => {
"https://huggingface.co/unsloth/Qwen3.6-27B-GGUF/resolve/82d411acf4a06cfb8d9b073a5211bf410bfc29bf/Qwen3.6-27B-Q4_K_M.gguf"
} }
} }
} }
pub fn sha256(&self) -> &'static str { pub fn sha256(&self) -> &'static str {
match self { match self {
Self::Qwen3_1_7B_Q4 => { Self::Qwen3_5_2B_Q4 => {
"de942b0819216caa3bfe487180dd1bb37398fa1c98cb42bb0bbac7ab7d6e8a12" "aaf42c8b7c3cab2bf3d69c355048d4a0ee9973d48f16c731c0520ee914699223"
} }
Self::Qwen3_4BInstruct2507Q4 => { Self::Qwen3_5_4B_Q4 => {
"bf52d44a54b81d44219833556849529ee96f09da673a38783dddc2e2eaf17881" "00fe7986ff5f6b463e62455821146049db6f9313603938a70800d1fb69ef11a4"
}
Self::Qwen3_5_9B_Q4 => {
"03b74727a860a56338e042c4420bb3f04b2fec5734175f4cb9fa853daf52b7e8"
}
Self::Qwen3_6_27B_Q4 => {
"5ed60d0af4650a854b1755bd392f9aef4872643dc25a254bc68043fa638392a0"
} }
Self::Qwen3_14BQ5 => "6f87abc471bd509ad46aca4284b3cfa926d8114bc491bb0a7a3a7f74c16ef95b",
} }
} }
} }
@@ -121,9 +140,10 @@ impl FromStr for LlmModelId {
fn from_str(value: &str) -> Result<Self, Self::Err> { fn from_str(value: &str) -> Result<Self, Self::Err> {
match value { match value {
"qwen3_1_7b" => Ok(Self::Qwen3_1_7B_Q4), "qwen3_5_2b" => Ok(Self::Qwen3_5_2B_Q4),
"qwen3_4b_instruct_2507" => Ok(Self::Qwen3_4BInstruct2507Q4), "qwen3_5_4b" => Ok(Self::Qwen3_5_4B_Q4),
"qwen3_14b" => Ok(Self::Qwen3_14BQ5), "qwen3_5_9b" => Ok(Self::Qwen3_5_9B_Q4),
"qwen3_6_27b" => Ok(Self::Qwen3_6_27B_Q4),
other => Err(format!("Unknown LLM model id: {other}")), other => Err(format!("Unknown LLM model id: {other}")),
} }
} }
@@ -154,9 +174,10 @@ pub enum DownloadError {
} }
const ALL_MODELS: &[LlmModelId] = &[ const ALL_MODELS: &[LlmModelId] = &[
LlmModelId::Qwen3_1_7B_Q4, LlmModelId::Qwen3_5_2B_Q4,
LlmModelId::Qwen3_4BInstruct2507Q4, LlmModelId::Qwen3_5_4B_Q4,
LlmModelId::Qwen3_14BQ5, LlmModelId::Qwen3_5_9B_Q4,
LlmModelId::Qwen3_6_27B_Q4,
]; ];
static ACTIVE_DOWNLOADS: LazyLock<Mutex<std::collections::HashSet<LlmModelId>>> = static ACTIVE_DOWNLOADS: LazyLock<Mutex<std::collections::HashSet<LlmModelId>>> =
@@ -206,16 +227,15 @@ pub fn model_info(id: LlmModelId) -> LlmModelInfo {
} }
pub fn recommend_tier(total_ram_bytes: u64, total_vram_bytes: Option<u64>) -> LlmModelId { pub fn recommend_tier(total_ram_bytes: u64, total_vram_bytes: Option<u64>) -> LlmModelId {
if total_vram_bytes.unwrap_or(0) >= 16 * 1024_u64.pow(3) let vram = total_vram_bytes.unwrap_or(0);
&& total_ram_bytes >= 32 * 1024_u64.pow(3) if vram >= 24 * 1024_u64.pow(3) && total_ram_bytes >= 64 * 1024_u64.pow(3) {
{ LlmModelId::Qwen3_6_27B_Q4
LlmModelId::Qwen3_14BQ5 } else if vram >= 12 * 1024_u64.pow(3) && total_ram_bytes >= 32 * 1024_u64.pow(3) {
} else if total_vram_bytes.unwrap_or(0) >= 8 * 1024_u64.pow(3) LlmModelId::Qwen3_5_9B_Q4
|| total_ram_bytes >= 16 * 1024_u64.pow(3) } else if vram >= 6 * 1024_u64.pow(3) || total_ram_bytes >= 16 * 1024_u64.pow(3) {
{ LlmModelId::Qwen3_5_4B_Q4
LlmModelId::Qwen3_4BInstruct2507Q4
} else { } else {
LlmModelId::Qwen3_1_7B_Q4 LlmModelId::Qwen3_5_2B_Q4
} }
} }
@@ -389,15 +409,15 @@ mod tests {
#[test] #[test]
fn model_path_contains_model_dir_and_filename() { fn model_path_contains_model_dir_and_filename() {
let path = model_path(LlmModelId::Qwen3_1_7B_Q4); let path = model_path(LlmModelId::Qwen3_5_2B_Q4);
assert!(path.to_string_lossy().ends_with("Qwen3-1.7B-Q4_K_M.gguf")); assert!(path.to_string_lossy().ends_with("Qwen3.5-2B-Q4_K_M.gguf"));
assert!(path.starts_with(model_dir())); assert!(path.starts_with(model_dir()));
} }
#[test] #[test]
fn recommend_tier_prefers_mid_by_default() { fn recommend_tier_prefers_mid_by_default() {
let tier = recommend_tier(16 * 1024_u64.pow(3), None); let tier = recommend_tier(16 * 1024_u64.pow(3), None);
assert_eq!(tier, LlmModelId::Qwen3_4BInstruct2507Q4); assert_eq!(tier, LlmModelId::Qwen3_5_4B_Q4);
} }
#[tokio::test] #[tokio::test]

2
crates/llm/tests/content_tags_smoke.rs
View File

@@ -24,7 +24,7 @@ fn extract_content_tags_returns_valid_pair() {
let engine = LlmEngine::new(); let engine = LlmEngine::new();
engine engine
.load_model(LlmModelId::Qwen3_1_7B_Q4, &model_path, true) .load_model(LlmModelId::Qwen3_5_2B_Q4, &model_path, true)
.expect("load model"); .expect("load model");
let transcript = "Tomorrow I need to run through the grant application one more time \ let transcript = "Tomorrow I need to run through the grant application one more time \

2
crates/llm/tests/smoke.rs
View File

@@ -26,7 +26,7 @@ fn llama_cpp_2_smoke_generates_and_wraps() {
let engine = LlmEngine::new(); let engine = LlmEngine::new();
engine engine
.load_model(LlmModelId::Qwen3_1_7B_Q4, &model_path, true) .load_model(LlmModelId::Qwen3_5_2B_Q4, &model_path, true)
.expect("load model"); .expect("load model");
let completion = engine let completion = engine