Apodex 1.1 Mini: A Free, Open Research Agent Model Examined

Most small open models are pitched as cheap general assistants. Apodex 1.1 Mini is pitched at something narrower and harder: long research tasks where the model works through files, data, code and tools, and the output is supposed to be a deliverable you can check rather than a confident paragraph.

It appeared on OpenRouter as a free endpoint at the start of October 2026, which is how many people will first meet it. Before building on it, it helps to know what the model actually is, how much of the pitch depends on the software around it, and which numbers are claims rather than measurements.

What the model actually is

The Hugging Face model card is clear about lineage. Apodex-1.1-mini is a fine-tune of Qwen's Qwen3.5-35B-A3B, itself trained from the Qwen3.5-35B-A3B base model. The card lists about 36B total parameters in BF16. It does not state an active parameter count, but the base model's name indicates roughly 3B active per token.

That ratio is the practical headline. A 35B mixture-of-experts model with around 3B active is cheap to run per token and fits on hardware that a dense 35B model would strain. Active vs total parameters covers why the memory bill follows the total and the compute bill follows the active count.

Other facts from the card:

• Licence — Apache 2.0, which permits commercial use, modification and redistribution.
• Context — 262,144 tokens.
• Recommended output cap — 32,768 tokens, with temperature 1.0, top_p 0.95 and a repetition penalty of 1.05.
• Format — the Qwen3.5 chat template, with reasoning in a think block and tool calls parsed into standard OpenAI-style tool_calls by vLLM or SGLang.

The accompanying paper, arXiv 2608.23283, is dated 24 August 2026. The larger Apodex 1.1 model it also describes is available through Apodex's online service; Mini is the one this guide covers.

The harness is half the product

The benchmark numbers Apodex publishes for Mini come from what it calls an Agent Team setup. The model splits a task into parts, runs subagents in parallel, and writes their results into a shared task state so it can replan without starting over. The card links the harness that does this, a project called FrontierAgent published on GitHub, and reports that Agent Team beats a plain ReAct loop.

The published figures for Mini in that setup are 50.2 on FrontierFinance and 27.7 on APEX-Agents. For comparison, the card gives the full Apodex 1.1 54.3 and 38.5 on the same two benchmarks. The card says benchmark-hosting sites were blocked during evaluation to avoid answer leakage, which is a sensible precaution for a model with web access.

The implication is easy to miss. If you call Mini through a chat-completions endpoint with your own simple loop, you are not running the system those numbers describe. You are running a 3B-active model in a harness it was not tuned to report on. Expect weaker results until you either adopt FrontierAgent or build something with the same plan, delegate and replan shape.

About the forecasting claim

OpenRouter's listing describes Mini as built for research and forecasting tasks. No forecasting accuracy figure appears on the model card. That does not mean the claim is false, only that it is untested in public, and forecasting is unusually easy to evaluate yourself.

The method:

1. Collect thirty questions that have already resolved since the model's training data ends, with dates and outcomes recorded.
2. Give the model the information that was available before each resolution date, and ask for a probability, not a yes or no.
3. Score with the Brier score: the mean squared error between probability and outcome. Lower is better, and always answering 0.5 scores 0.25.
4. Run a frontier model and a plain base model on the same set. Mini only earns its place if it beats the cheap baseline clearly.

The same approach works for research deliverables: pick tasks where you already know the right answer and check whether the cited sources actually say what the model claims they say. A traceable chain of evidence is only as good as the evidence at the end of it. How to benchmark LLMs on your own work covers building the task set.

Free endpoint or self-hosted

The :free variant on OpenRouter is listed at zero cost per token. Free endpoints are rate-limited and can change terms or disappear, so treat it as a way to evaluate the model, not as production capacity. Free tier strategies covers how to get useful answers out of limits like these.

Ignore the maximum output figure on some aggregator listings, which is almost the entire context window. Use the card's 32,768-token cap as your working limit, and budget a long research run as many turns rather than one enormous answer.

Self-hosting is where the Apache 2.0 licence pays off. The card gives vLLM and SGLang commands with tensor parallelism across eight GPUs and the full 262K context. With roughly 3B active parameters a smaller deployment is plausible, but the KV cache for long contexts is what will set your memory floor, so test at the context length you actually need.

Who should try it

• Try it if you want an openly licensed research agent you can run on your own infrastructure, and you are prepared to adopt or reproduce its multi-agent harness.
• Try it as a cheap worker inside a larger system, where a stronger model plans and Mini does the file and data legwork.
• Skip it if you need a general chat model, image input, or results you can trust without building an evaluation first.