OsaurusAI/LFM3-Audio-2B-JANG_6M

JANG_6M build of LiquidAI/LFM3-Audio-2B for Apple Silicon. LFM3-Audio 2B: the LFM3 trunk with a 17-layer LFM2 speech encoder (16 kHz).

8-bit token mixers, 6-bit FFN, and every quantized matrix calibrated with GPTQ on the exact storage grid plus activation-aware scaling, using calibration data weighted toward tool calls, terminal agents, reasoning and code.

Bundle

Modalities	text, speech in
On-disk size	2.01 GiB (one shard), float16
Attention q/k/v/o, short-conv in/out projections	8-bit affine, group 32, GPTQ
FFN w1 / w3 / w2	6-bit affine, group 32, AWQ + GPTQ
Tied embedding / lm_head	8-bit affine, group 32 (it is the output head)
Kept at full precision	attention output gate, depthwise conv kernels, all norms, speech encoder and projector, speech-output decoder (unused for text)
GPTQ	142/142 tensors: fixed min-max grid (byte-identical layout to mx.quantize), act-order, Hessians accumulated over every calibration token, w2 Hessian rebuilt through the already-quantized w1/w3, best-of-RTN guard; mean calibration reconstruction error -53.5% vs round-to-nearest
AWQ	channel scales (alpha 0.25, clip 0.5-2) folded into ffn_norm and w3 rows; function-preserving, no runtime cost

Calibration (529,024 tokens)

Every sample is rendered through the model's own chat template: structured tools, assistant tool calls in LFM3's <tool_call><function=...> form, tool results as tool turns, and agent thoughts inside <think>. Part of the set is the model's own answers (card sampling) to reasoning, tool, terminal and coding prompts, so its serving distribution is in calibration.

domain	tokens
tool calling (Hermes, native XML calls)	154,969
terminal / SWE agent trajectories	174,363
reasoning (long traces)	105,223
coding and agentic text	46,072
general, science, systems, security, multilingual, long context	37,598
speech (ASR, spoken QA, spoken instructions)	10,799

Measured against the bf16 source

64 held-out sequences never seen in calibration (13,958 scored positions: tool calls, terminal and SWE agents, reasoning, code, general text, speech). KL divergence in nats per position (lower is better); top-1 is argmax agreement with bf16.

build	size	KL median	KL mean	KL p95	top-1
this bundle	2.01 GiB	0.00004	0.00119	0.0049	98.95%
same layout, round-to-nearest	2.01 GiB	0.00006	0.00190	0.0062	98.76%
uniform 6-bit gs64 (mlx-vlm default recipe)	1.77 GiB	0.00016	0.00424	0.0171	97.97%

Use

• Runtimes with LFM3 support: vMLX / Osaurus (vmlx-swift with LFM3, Osaurus with LFM3 media support), and mlx-vlm with the LFM3 port.
• Thinking: no thinking mode. Tool calls: XML function format. Sampling defaults: temperature 0.0, top_k 1, repetition_penalty 1.0.
• License: LFM Open License v1.0 (see LICENSE), same as the base model.