# Apodex 1.1 Mini ![Organization](https://img.shields.io/badge/Org-Apodex-1f6feb) ![Params](https://img.shields.io/badge/Params-36B%20MoE-8957e5) ![Context](https://img.shields.io/badge/Context-262K%20tokens-1f6feb) ![License](https://img.shields.io/badge/License-Apache%202.0-22c55e) ![Modality](https://img.shields.io/badge/Modality-Text%20%2B%20Vision-8957e5) ![Released](https://img.shields.io/badge/Released-Aug%202026-555) > **Quick answer:** Apodex 1.1 Mini is the open-weight, locally-deployable member of Apodex's 1.1 agent-model family — a 36B-parameter Mixture-of-Experts model fine-tuned from Qwen3.5-35B-A3B, released under Apache 2.0 with a 262,144-token context window. Using Apodex's own Agent Team multi-agent harness, it scores 50.2% on FrontierFinance, 51.7% on FrontierScience-Research, and 27.7% on APEX-Agents — retaining most of the working capability of Apodex's much larger closed flagship model. ## At a Glance **Where Apodex 1.1 Mini leads** - Open weights under Apache 2.0 — no restrictions on self-hosting, fine-tuning, or commercial use, unlike the closed 397B flagship - Leads FrontierFinance among models near its size class at 50.2% and comes close to the flagship's APEX-Agents result (27.7% vs 38.5%) despite being roughly 11x smaller - Long native context (262,144 tokens) and multimodal (vision-capable) input support - Built-in reasoning-effort switch and Qwen3.5-style native tool calling (`` syntax), compatible with SGLang/vLLM deployment out of the box **Where it lags** - Meaningfully behind the flagship Apodex 1.1 on every shared benchmark (e.g. FrontierScience-Research 51.7% vs 63.3%, APEX-Agents 27.7% vs 38.5%) - As a Mixture-of-Experts model with a relatively small ~3B active-parameter budget per token, single-turn raw reasoning is a step below dense frontier models of similar total size - Apodex's own paper only reports Mini scores on three benchmarks (FrontierFinance, FrontierScience-Research, APEX-Agents) — no published Mini-scale numbers yet for coding, math, or general-knowledge benchmarks like SWE-bench Verified or HLE **Best for:** Teams that want Apodex's agentic "working capability" approach (sustained multi-step task completion via file/search/code tool use) in a self-hostable, fine-tunable form rather than the API-only flagship. ## What Apodex 1.1 Mini Is Apodex 1.1 Mini is the openly-released, locally-deployable counterpart to Apodex's flagship 1.1 model, described in Apodex's technical report "Apodex 1.1: Scaling Agentic Intelligence for Complex Work" (arXiv:2608.23283). It is fine-tuned from Alibaba's Qwen3.5-35B-A3B Mixture-of-Experts base model, inheriting that architecture's sparse expert-routing design (36B total parameters, a small fraction active per token) rather than being trained from scratch. Like the flagship, Mini is trained around Apodex's "working capability" framing — sustained, verifiable progress on real-world tasks via file, search, and code tool interaction — using the same underlying **AgentOS** execution substrate (a three-namespace filesystem separating read-only inputs, scratch workspace, and single-writer-lease outputs) and the same unified SFT + PIVOT-RL training recipe described in the paper. Apodex reports Mini's results under both a minimal single-agent **ReAct** scaffold and the fuller multi-agent **Agent Team** coordination harness (explicit task board, asymmetric verification, adaptive effort allocation); the Agent Team setting consistently produces the higher, and more representative, scores. ## Specifications | Field | Value | |---|---| | Organization | Apodex | | Parameters | 36B total (Mixture-of-Experts; ~3B active per token) | | Base model | Qwen/Qwen3.5-35B-A3B | | Context window | 262,144 tokens | | Architecture | Qwen3.5 MoE (`qwen3_5_moe`), native tool calling, thinking/reasoning-effort switch | | License | Apache 2.0 | | Release date | August 2026 | | Modality | Multimodal (text + vision; English and Chinese) | ## Pricing Open weights available on [Hugging Face](https://huggingface.co/apodex/Apodex-1.1-mini) (`apodex/Apodex-1.1-mini`) — free to self-host under Apache 2.0. FP8, NVFP4, and GPTQ-Int4 quantized variants are also published for lower-resource deployment. Recommended inference settings: temperature 1.0, top_p 0.95, repetition_penalty 1.05, max_tokens 32,768; deploy via SGLang (`--context-length 262144`) or vLLM (`--max-model-len 262144`). ## Context Window Apodex 1.1 Mini natively supports a **262,144-token (256K) context window** — enough for large multi-file codebases, long agent trajectories, or extended multi-document research tasks in a single request. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [FrontierFinance](/benchmarks/samaya-research/frontier-finance) | 50.2% | [Apodex 1.1 technical report](https://arxiv.org/abs/2608.23283) | 2026-08 | | [FrontierScience-Research](/benchmarks/openai/frontierscience-research) | 51.7% | [Apodex 1.1 technical report](https://arxiv.org/abs/2608.23283) | 2026-08 | | [APEX-Agents](/benchmarks/mercor/apex-agents) | 27.7% | [Apodex 1.1 technical report](https://arxiv.org/abs/2608.23283) | 2026-08 | *Scores are self-reported by Apodex under its Agent Team multi-agent coordination harness (the higher of the two settings Apodex reports; a minimal ReAct scaffold produces lower scores of 40.0% / 45.0% / 24.2% respectively on the same three benchmarks). Not Benchgen measurements — figures depend heavily on harness and coordination-strategy choices; see the source paper for methodology.* ## Apodex 1.1 Mini vs Alternatives | Model | Params | Context | FrontierFinance | APEX-Agents | License | |---|---|---|---|---|---| | **Apodex 1.1 Mini** | 36B (MoE) | 262K | 50.2% | 27.7% | Apache 2.0 | | [Apodex 1.1](/models/apodex/apodex-1-1) | 397B | — | 54.3% | 38.5% | Proprietary | | [Qwen3.5-35B-A3B](/models/alibaba/qwen3-5-35b-a3b) | 35B (MoE) | — | — | — | — | Apodex 1.1 Mini trades roughly 4 points of FrontierFinance and 11 points of APEX-Agents versus its own 397B flagship — a meaningfully smaller gap than the ~11x parameter difference would suggest, making it the more practical choice for teams that need self-hosted deployment or fine-tuning rights rather than the absolute ceiling score. ## How Apodex 1.1 Mini Performs on Real Agent Tasks Apodex's own framing is that Mini exists to prove its "working capability" approach — Environment Scaling (executable file/search/code worlds) and Agentic Coordination Scaling (the Agent Team multi-agent protocol) — transfers to a much smaller, locally-deployable model rather than requiring flagship-scale compute. The reported results on FrontierFinance (real investment-analyst workflows) and FrontierScience-Research (open-ended PhD-level science sub-problems) support that: Mini retains a large majority of the flagship's capability on both despite the parameter gap. However, Apodex has not yet published Mini-scale results on coding or terminal-agent benchmarks (SWE-bench Verified, Terminal-Bench 2.1) or general-knowledge/search benchmarks (HLE, DeepSearchQA) that it reports for the flagship — so its agentic coding and long-horizon search reliability specifically remain unverified by any published, independent number. ## Use Apodex 1.1 Mini via API ```python from openai import OpenAI # Apodex 1.1 Mini uses an OpenAI-compatible chat API when self-hosted via vLLM/SGLang client = OpenAI(base_url="http://localhost:8000/v1", api_key="not-needed") response = client.chat.completions.create( model="apodex/Apodex-1.1-mini", messages=[{"role": "user", "content": "Summarize this quarterly earnings call transcript."}], temperature=1.0, top_p=0.95, max_tokens=32768, ) print(response.choices[0].message.content) ``` ## Related Models & Benchmarks - Compare: [Apodex 1.1](/models/apodex/apodex-1-1) - Benchmarks: [FrontierFinance](/benchmarks/samaya-research/frontier-finance) · [FrontierScience-Research](/benchmarks/openai/frontierscience-research) · [APEX-Agents](/benchmarks/mercor/apex-agents) ## Frequently Asked Questions

What is Apodex 1.1 Mini?

Apodex 1.1 Mini is the open-weight, 36B-parameter Mixture-of-Experts member of Apodex's 1.1 agent-model family, fine-tuned from Qwen3.5-35B-A3B and released under Apache 2.0.

Is Apodex 1.1 Mini open source?

Yes — it's released under the Apache 2.0 license with open weights on Hugging Face, including quantized (FP8, NVFP4, GPTQ-Int4) variants.

How does Apodex 1.1 Mini compare to the full Apodex 1.1?

Mini scores meaningfully lower on every benchmark both models report (e.g. 27.7% vs 38.5% on APEX-Agents), but retains a large share of the flagship's capability at a fraction of the size and without requiring API access to a closed model.

--- *Specs from the Apodex 1.1 technical report (arXiv:2608.23283) and the official Hugging Face model card (huggingface.co/apodex/Apodex-1.1-mini). Last updated 2026-08-31.*

Apodex 1.1 Mini — Evaluations

Evaluations