leonsarmiento/Salience-1.5-Pro-6bit-XL-mlx

Salience 1.5 Pro by Vection Labs — a 35B-A3B sparse MoE software-engineering agent with native vision, quantized for Apple Silicon using the BaseQuant_XL 6/8-bit recipe.

Only ~3B parameters active per token — the decode speed of a small model with the capacity of a 35B one. Built for hard engineering work: writing and debugging real code, repo-scale edits, robust backend systems, and driving tools through long-horizon tasks.

Quickstart

pip install -U mlx-vlm

python -m mlx_vlm.generate --model leonsarmiento/Salience-1.5-Pro-6bit-XL-mlx --max-tokens 256 --temperature 0.6 --top-p 0.95 --prompt "Implement an LRU cache in Python with O(1) get/put."

Works with LM Studio — vision mmproj included. Thinking mode is on by default (emits <think>...</think>). Pass enable_thinking=False for direct answers.

Quantization Strategy

BaseQuant_XL recipe — precision is allocated by layer importance, not applied uniformly:

Layers	Bits	Rationale
mlp.gate (router), shared_expert_gate, lm_head, shared_expert	bf16	Routing decisions and output projection — any quantization noise here causes expert misrouting or output degradation
embed_tokens, self_attn, linear_attn	8-bit	Every-token layers — near-lossless, attention quality preserved
vision_tower, switch_mlp (routed experts)	6-bit	Bulk parameters — 256 experts with only 8 active per token; redundancy absorbs quantization noise. 6-bit is the sweet spot for routed experts (higher bits can cause overthinking)

• Bits per weight: ~6.8 · Total size: ~28 GB · Group size: 64

Recommended Inference Parameters

Parameter	Value
temperature	0.6 – 0.85
top_p	0.95
top_k	20
presence_penalty	1.1
reasoning_parser	deepseek_r1

Lower temperature for precise code edits; higher for long agentic sessions.

Model Overview

Property	Value
Architecture	Qwen3.5 MoE (35B-A3B) + native vision encoder
Parameters	35B total / ~3B active per token
Experts	256 (8 routed + 1 shared)
Modalities	text, image, video → text
Context window	262,144 tokens native
Attention	Hybrid linear + full (full every 4th layer)
Thinking	On by default (<think>...</think>)
Tool calling	XML-style (<tool_call><function=...><parameter=...>)
License	Apache-2.0

MLX Benchmarks (n=50, instruct mode)

Benchmark	Salience-1.5-Pro 6bit-XL	Qwen3.6-35B 6bit-XL	Gemma-4 26B 6bit-XL
MMLU	74%	64%	76%
MMLU_PRO	78%	64%	82%
ARC_CHALLENGE	90%	90%	90%
HUMANEVAL	80%	78%	98%
MBPP	76%	78%	82%
Average	79.6%	74.8%	85.6%

Salience leads Qwen3.6 on every benchmark and is competitive with Gemma-4 — +14 on MMLU_PRO reflects strong structured reasoning from its software-engineering tuning. Gemma-4 remains the coding champion (HUMANEVAL 98%, MBPP 82%).

MATHQA Thinking Efficiency (n=10, seed=42, 8192-token budget)

Reasoning efficiency measures how concisely a model arrives at the correct answer — fewer reasoning chars per correct answer means less context window consumed and less energy per answer. All three models achieve 10/10 verified accuracy; the differentiator is how much thinking it takes to get there.

Model	Verified Acc	Reason chars/correct	Time/correct
Gemma-4 26B 6bit-XL	10/10 (100%)	2,420	39.8s
Qwen3.6-35B 6bit-XL	10/10 (100%)	3,562	32.1s
Salience-1.5-Pro 6bit-XL	10/10 (100%)	5,350	47.7s

Salience's SWE-focused tuning produces thorough, methodical reasoning chains — longer than Gemma-4's concise style, but every chain reaches the correct answer. Gemma-4 remains the reasoning-efficiency champion; Salience trades more thinking for broader engineering depth and multimodal capability.

Capabilities

• Code & SWE execution — runnable code, repo-scale edits, methodical debugging, robust backends
• Agentic tool-use — multi-step planning, tool orchestration, long-horizon task execution
• Deep reasoning — structured, inspectable chains of thought with one-token switch to disable
• Multimodal perception — diagrams, screenshots, documents, video as first-class inputs

Source Model

Made by Vection Labs. Base architecture: Qwen3.5 MoE (35B-A3B). MLX quantization by leonsarmiento.