sevren-ai/Embla1-4bit

Embla 1 (sevren.ai) is a mixture-of-experts language model designed for phones. This repository contains the MLX 4-bit checkpoint (affine, group size 64, bf16 scales and biases): the exact weights used by the phone build.

The 4-bit format is the intended precision. Embla 1 was post-trained with quantization-aware training (QAT) specifically for this format, so these are the weights the training targeted. They are produced by quantizing the bf16 checkpoint sevren-ai/Embla1-bf16 with round-to-nearest.

Model details

• Parameters: 10B total, including 3B engram parameters; 1B active per token.
• Context length: 256K tokens (262,144).
• License: MIT.

Runtime

embla-metal is our optimized runtime for iPhones and Apple silicon Macs. This Metal decode engine is built for the checkpoint: it memory-maps the 4-bit weights directly into Metal and reads the engram table from storage.

Layout

• model.safetensors: the quantized layout used by mlx-lm. A quantized module holds <module>.weight (uint32, with eight 4-bit codes packed per word), <module>.scales, and <module>.biases (bf16, one per 64 inputs). Reconstruct a weight with mx.dequantize(weight, scales, biases, group_size=64, bits=4). Routers, expert biases, norm gains, AttnRes queries, and the engram conv remain in bf16.
• The 96 routed experts in each layer are stored as two stacks: mlp.experts_fc [E, f, d] and mlp.experts_proj [E, d, f].
• engram_table.safetensors: the engram's 23,488,192 x 128 rows, quantized the same way but with fp16 scales and biases (the phone's format). To recover the trained row exactly, cast a row's scale and bias to fp32, dequantize, and round to bf16. Using mx.dequantize in fp16 is one bf16 step off for a few percent of values.
• config.json: the architecture description (model_type: embla, identical to the transformers release) and mlx-lm's quantization block, with the routers listed as unquantized.

mlx-lm does not currently support embla, so this checkpoint does not load through mlx_lm.load. The transformers release (bf16, with reference model code) is the authoritative definition of the math.