⚡ All JANG models are meant to be run in vMLX
JANGQ-AI/Laguna-S-2.1-JANG_4M
JANG_4M JANG affine bundle of poolside/Laguna-S-2.1 — 118B-parameter / ~8B-active MoE for agentic coding and long-horizon work, quantized for Apple Silicon with AWQ-protected routed experts.
Everything quality-critical stays high precision: attention (including the softplus output gate), the shared expert path, router, and norms. The 256-routed-expert bulk (~97% of parameters) carries the low bits, protected by activation-aware (AWQ) scales selected by measured quantization error on real expert weights and real calibration activations — the search always includes a no-AWQ baseline, so scales ship only because they beat it (2.9% lower reconstruction error at this bit width).
Bundle
| Field | Value |
|---|---|
| Source | poolside/Laguna-S-2.1 @ a50e85e (BF16, 235 GB) |
| Architecture | laguna / LagunaForCausalLM — 48 layers (12 global + 36 SWA w512), GQA 8 KV heads, 256 routed experts top-10 + 1 shared, dense layer 0, 1M context |
| On-disk size | 68.1 GB (15 shards) |
| Routed experts | 4-bit gate/up/down affine, group 64, AWQ input scales folded |
| Attention q/k/v/o + g_proj | 8-bit affine |
| Shared expert / dense FFN | 8-bit affine |
| Embeddings / lm_head | 6-bit / 8-bit affine |
| Router, e_score bias, norms | fp16 passthrough (routing bit-identical to unfolded math) |
| Modality | text-only (verified from tensor index — no vision/audio/video weights) |
Measured performance (M5 Max, 128 GB)
| Metric | Value |
|---|---|
| Decode | 33-44 tok/s greedy (cold-warm) |
| Load time | 5.3 s |
| Long-context cache parity | teacher-forced top-1 agreement 0.996 / 1.000 across a 2,913-token pass (pre / post the 512 sliding window) |
| BF16 vs quantized | greedy smoke + chat-with-thinking verified coherent |
Memory note: budget bundle size + ~10 GB. The reference Python runtime auto-sets the Metal wired limit (min(bundle x 1.2 + 8 GB, 118 GB)) — without a wired limit, decode throughput roughly halves on 64 GB+ working sets.
Chat protocol
- GLM-style think tags. Thinking is ON by default (vendor
default_chat_template_kwargs.enable_thinking=true); passenable_thinking=Falseto disable. The generation prompt ends<assistant><think>(thinking) or<assistant></think>(off). - Stop tokens:
eos_token_id = [2, 24]— id 24 is end-of-turn and MUST be in the stop set. - The template emits its own leading
〈|EOS|〉(= bos id 2): do not prepend another BOS. - Tool calls:
<tool_call>name<arg_key>k</arg_key><arg_value>v</arg_value></tool_call>(GLM-4.7-compatible parsing). - Sampling defaults (vendor): temperature 1.0, top_p 1.0, top_k 20.
Run it
Osaurus loads this bundle natively. Python reference runtime (mixed-precision affine aware):
pip install mlx mlx-lm transformers
python -m jang_tools.laguna.runtime --src ./Laguna-S-2.1-JANG_4M \
--prompt 'def fibonacci(n):' --max-new 64
config.json[quantization] carries per-module {bits, group_size, mode} overrides — loaders must honor per-module bits (a single top-level width mis-dequantizes the low-bit experts).
Source model quality (poolside, BF16 base)
Per the upstream model card: Terminal-Bench 2.1 70.2%, SWE-bench Multilingual 78.5%, SWE-Bench Pro 59.4% — competitive with much larger frontier MoEs. These figures are for the BF16 source model.
Quantized and verified by Jinho Jang (eric@jangq.ai) with the JANG toolchain. License: OpenMDW-1.1 (inherited from the source model).