⚠️ Re-download if you pulled this bundle before 2026-07-28

Earlier uploads of this bundle were missing top_k: 20 in generation_config.json and jang_config.json. Poolside's own Laguna-XS-2.1/generation_config.json omits top_k even though their model card specifies temperature=1.0, top_k=20, top_p=1 for every XS-2.1 evaluation (the S-2.1 config does ship it), and this bundle passed those values through verbatim.

Impact: with top_k unset, a runtime samples the full 100,352-token vocabulary unfiltered at temperature=1.0 / top_p=1.0 instead of the top 20 candidates — noticeably lower output quality, worst on the lower-bit profiles.

Fix: re-download generation_config.json and jang_config.json (a few KB — no need to re-pull the weights), or pass top_k=20 explicitly in your requests. Both files now carry top_k: 20 and are checked against each other at build time.

OsaurusAI/Laguna-XS-2.1-JANG_4M

JANG_4M JANG affine bundle of poolside/Laguna-XS-2.1 — 33B-parameter / ~3B-active MoE for agentic coding, quantized for Apple Silicon.

Attention (incl. the softplus output gate), shared expert, router, and norms stay high precision; the 256-routed-expert bulk carries 4-bit gate/up/down, protected by activation-aware (AWQ) scales chosen by measured quantization error against a no-AWQ baseline (0.7% lower reconstruction error at this bit width).

Bundle

Field	Value
Source	`poolside/Laguna-XS-2.1` @ `c405648`
Architecture	`laguna` — 40 layers (10 global + 30 SWA w512), 33B-A3B, 256 experts top-8 + shared, 262K ctx
On-disk size	19.8 GB (5 shards)
Routed experts	4-bit gate/up/down affine, group 64, AWQ 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
Modality	text-only (verified from tensor index)

Measured (M5 Max, 128 GB, greedy, wired)

Metric	Value
Decode	80.5 tok/s
Long-context cache parity	teacher-forced top-1 agreement 0.998 / 1.000 (pre/post the 512 sliding window, 2,913-token pass)

Chat / reasoning

GLM-style think tags; enable_thinking toggles reasoning (vendor serving default is ON (stamped in jang_config.chat); the template's own fallback is off, so pass enable_thinking=True explicitly if your runtime ignores bundle defaults)
The FULL chat template is inlined in tokenizer_config.json (upstream ships only an {% include %} stub that most runtimes cannot resolve — inlining is what makes the reasoning toggle actually work)
Stop tokens eos_token_id = [2, 24] — id 24 is end-of-turn and must be in the stop set
Template emits its own leading 〈|EOS|〉 (bos 2): do not prepend another
Tool calls: <tool_call>name<arg_key>k</arg_key><arg_value>v</arg_value></tool_call>
Vendor sampling: temperature 1.0, top_p 1.0, top_k 20

Loaders must honor the per-module {bits, group_size, mode} overrides in config.json[quantization].

Quantized and verified by Jinho Jang (eric@osaurus.ai). License: OpenMDW-1.1 (inherited).