leonsarmiento/Apodex-1.1-mini-6bit-XL-mlx
Apodex-1.1-mini by Apodex AI — a reasoning-first agent model for complex, long-horizon research tasks (Qwen/Qwen3.5-35B-A3B fine-tune), quantized for Apple Silicon using the BaseQuant_XL 6/8-bit recipe.
Only ~3B parameters active per token. Apodex works directly with papers, datasets, spreadsheets, images, and code — cleaning data, selecting methods, running analyses, recovering from errors, and turning raw inputs into verifiable deliverables within one continuous task.
This is a full multimodal build — the base model's vision tower is preserved (the fine-tune itself is text-trained; the tower carries over from Qwen3.5-35B-A3B).
About XL Quantization
BaseQuant_XL is a fully data-agnostic, static quantization. No calibration dataset, no sensitivity analysis, no importance matrix. Precision is allocated purely by architectural role — routing-critical layers get higher precision, bulk expert parameters get lower precision. The result is a transparent, faithful capture of the source model.
Data-dependent calibration quantizations (iMatrix, AWQ, GPTQ, oQ, oQ4e, etc.) use a calibration set to guide bit allocation. This can produce a skewed representation of the model: domains well-represented in the calibration data (English, popular topics, public or leaked benchmarks) are preserved better, while underrepresented domains (non-English languages, niche use cases, your own data) are preserved worse. XL avoids this trade-off entirely — it generalizes honestly because it is never fit to any particular data distribution.
Quickstart
pip install -U mlx-vlm
python -m mlx_vlm.generate --model leonsarmiento/Apodex-1.1-mini-6bit-XL-mlx --max-tokens 4096 --temperature 1.0 --top-p 0.95 --prompt "Research the current state of solid-state battery manufacturing and produce a short brief."
Apodex follows the Qwen3.5 chat template — tool calls are emitted as <tool_call><function=...><parameter=...></parameter></function></tool_call> and reasoning as <think>...</think>. Launch with the matching parsers so the server returns standard OpenAI-style tool_calls and reasoning_content fields.
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 |
visual (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 |
- Bits per weight: 6.808 · Total size: 29.87 GB (6 shards) · Group size: 64
Notes specific to this build:
- Source ships an MTP (multi-token prediction) layer (785 tensors) — dropped in this build; the MTP config declaration is stripped so engines don't flag a phantom MTP.
- Source stores merged per-layer expert tensors — converts via stock
mlx_vlmsanitize (no custom sanitize needed).
Recommended Inference Parameters
Per the source model card:
| Parameter | Value |
|---|---|
temperature |
1.0 |
top_p |
0.95 |
repetition_penalty |
1.05 |
max_context_length |
262,144 |
max_tokens |
32,768 |
tool_call_parser |
qwen3_coder |
reasoning_parser |
qwen3 |
Thinking is on by default (<think>...</think> before the answer); with a reasoning parser enabled the chain-of-thought is returned in a separate reasoning_content field.
Agentic Usage
Apodex is trained for native function calling — pass tool schemas via the tools= parameter of the chat-completions API (rendered into the prompt by the chat template) rather than inlining them in the system prompt, so the qwen3_coder parser can recover structured tool_calls. The system prompt only needs the role and the high-level objective. Source-recommended harness: ApodexAI/FrontierAgent.
Model Overview
| Property | Value |
|---|---|
| Architecture | Qwen3.5-family MoE (35B-A3B) + native vision encoder |
| Parameters | 35.9B total / ~3B active per token |
| Experts | 256 (8 routed + 1 shared) |
| Attention | Hybrid linear + full attention |
| Modalities | text, image, video → text |
| Context window | 262,144 tokens native |
| Thinking | <think>...</think> |
| Tool calling | <tool_call><function=...> (qwen3_coder parser) |
| License | Apache 2.0 |
Source Model Benchmarks (from Apodex AI)
Apodex-1.1-mini with the Agent Team harness remains competitive with frontier models — leading FrontierFinance at 50.2 and nearly matching the best result on APEX-Agents at 27.7. The Agent Team setup consistently outperforms a ReAct loop across all reported benchmarks.
The full Apodex-1.1 Agent Team reports 38.5 APEX-Agents, 78.8 GDPVal, 54.3 FrontierFinance, 63.3 FrontierScience-Research, 35.3 BioMysteryBench, 56.1 Humanity's Last Exam.