Apodex 1.1 mini is the open-weight half of a release that also shipped an agent framework, and the framework may end up mattering more than the model. Apodex AI published both on 24 August 2026 alongside a technical report on arXiv, then followed with a company blog post on 15 September. FrontierAgent, the runtime and terminal interface, is on GitHub under Apache 2.0; the weights are on Hugging Face under the same licence.
The pitch is that this is not a research-and-summarise model. Apodex says the family opens files directly — a spreadsheet, a PDF, a dataset — chooses a method, writes and runs analysis code, and returns the tables, charts and scripts it produced with figures traceable back to source. The flagship is a 397-billion-parameter model. Apodex 1.1 mini is the sibling you can download and run yourself.
This piece works from the artefacts rather than the announcement: the Hugging Face configuration file, the model card, the GitHub repository, the arXiv abstract and the download counters. Where the launch coverage and the repositories disagree — and they do, in several specific places — we say which one to believe.
Table of contents
- What Apodex 1.1 mini Actually Is
- The FrontierAgent Framework Is the Other Half
- The Benchmark Numbers, and Who They Belong To
- What You Actually Get When You Download Apodex 1.1 mini
- What Apodex Says It Trained Differently
- Deep Discover, and What Apodex 1.1 mini Does Not Get
- Where the Downloads Are Going
- Running Apodex 1.1 mini Yourself
- Where the Apodex 1.1 mini Coverage Drifts From the Repositories
- Who Should Care About Apodex 1.1 mini
- References
What Apodex 1.1 mini Actually Is
Start with what is on the Apodex 1.1 mini repository page, because it settles several questions that the marketing leaves open.
A fine-tune, and the card says so
The model card lists Qwen/Qwen3.5-35B-A3B as the base model. Apodex 1.1 mini is a fine-tune, not a from-scratch pretrain, and the company is straightforward about it in the metadata even though the blog post is not. That is normal practice and it is a point in the release’s favour that the base is declared at all.
A mixture-of-experts, mostly asleep
The configuration file describes a qwen3_5_moe architecture with 40 layers, 256 experts and 8 experts activated per token. Hidden size is 2,048, the MoE intermediate size is 512, and there is a shared expert of the same width. The A3B in the base model name is the active parameter count: roughly three billion parameters do work on any given token.
The parameter count is 35.95 billion, not 35
Hugging Face’s own counter reads 35,951,822,704 parameters for the main repository. Everyone — Apodex, the arXiv abstract, the coverage — rounds that to 35B. It rounds more naturally to 36B, and the gap is about 0.95 billion parameters, or 2.7%. Not important for a buying decision; worth knowing if you are sizing memory.
Attention is mostly linear
The layer type list alternates three linear-attention layers to one full-attention layer, repeating ten times across the 40 layers. Only a quarter of the layers run full attention, which is how Apodex 1.1 mini reaches a 262,144-token position limit without the memory cost a fully quadratic stack would carry.
The FrontierAgent Framework Is the Other Half
FrontierAgent replaces AgentHarness, the single-agent evaluation harness Apodex shipped with version 1.0, and it is a considerably more ambitious object. It is also the piece you can use without running Apodex 1.1 mini at all.
Two workflows, one engine
| Mode | How it works | When it fits |
|---|---|---|
| ReAct | One stateful agent researches, reads files, writes deliverables and iterates in a task-scoped sandbox | Linear work with a single thread of reasoning |
| Agent Team | A coordinator keeps a task board, delegates to parallel sub-agents, collects structured reports and synthesises | Work that genuinely decomposes into independent branches |
The sandbox layout is the good part
Shell and file tools share one task-scoped filesystem with three directories: /inputs is read-only, /workspace holds working state, and /outputs holds persistent deliverables. Authorisation and sandbox failures are fail-closed. On macOS and Docker, /outputs maps to .apodex/runs/<session-id>/outputs on the host, next to the run’s checkpoint, trace, engine log and trajectories.
Asynchronous intervention
You can type while an agent is running. The instruction is queued and injected at the next safe turn boundary rather than interrupting the run. In Agent Team mode it steers the coordinator while sub-agents already in flight are allowed to finish. That is a genuinely well-considered design detail and it is the sort of thing a wrapper script cannot retrofit.
Approvals, traces and recovery
Mutating operations show a diff and require approval unless you pass --yes. Sessions are checkpointed, every action is traced locally, /revert restores session changes and --resume continues a saved run. The repository also bundles a benchmark runner with FrontierSearchBench and FrontierChallenge.
Adoption so far
The repository was created on 22 August 2026 and has 3,336 stars, 196 forks and 15 open issues, with commits still landing on 16 September. For a framework released three and a half weeks earlier that is real traction, though star counts measure attention rather than use.
The Benchmark Numbers, and Who They Belong To
This is where reading the model card rather than the press coverage changes the picture, so the attribution matters.
The flagship figures
Apodex reports the 397-billion-parameter flagship in its Agent Team configuration at 38.5 on APEX-Agents, 78.8 on GDPval, 54.3 on FrontierFinance, 63.3 on FrontierScience-Research, 35.3 on BioMysteryBench and 56.1 on Humanity’s Last Exam. Apodex says it blocks access to benchmark-hosting websites during evaluation to prevent answer retrieval, which is a sensible precaution to state.
The Apodex 1.1 mini figures are different numbers
The model card gives the mini its own paragraph: 50.2 on FrontierFinance, leading its comparison set, and 27.7 on APEX-Agent, described as nearly matching the best result. Those are the numbers that belong to Apodex 1.1 mini. The 38.5 and 63.3 figures do not.
Reading the two side by side
The generation-on-generation claim
Apodex reports 1.0 scoring 16.5 on APEX-Agents against 1.1’s 38.5, and 28.3 on FrontierScience-Research against 63.3. Those are multiples of 2.33 and 2.24 respectively, which supports the “more than doubles” framing. They are also the flagship’s numbers on both sides of the comparison, and they come from Apodex’s own harness.
What the comparison cannot tell you
Every figure here is vendor-run. The benchmarks FrontierFinance, FrontierScience-Research and APEX-Agents are not established third-party leaderboards with independent submission processes, and two of the three carry Apodex’s own “Frontier” naming. Treat the Apodex 1.1 mini scores as internal measurements published in good faith, not as external validation.
What You Actually Get When You Download Apodex 1.1 mini
The Hugging Face collection holds five official repositories, and the quantisations tell you something about the intended deployment.
The five official builds
| Repository | Reported parameters | Downloads | Likes |
|---|---|---|---|
| Apodex-1.1-mini | 35,951,822,704 | 9,863 | 134 |
| Apodex-1.1-mini-NVFP4 | 19,845,695,344 | 10,604 | 35 |
| Apodex-1.1-mini-FP8 | 35,953,788,784 | 6,687 | 15 |
| Apodex-1.1-mini-GPTQ-Int4 | Not reported | 6,457 | 20 |
| Apodex-1.1-mini-GGUF | Not reported | 832 | 18 |
Do not read the NVFP4 parameter count literally
That repository reports 19.85 billion parameters, roughly 55% of the main figure. The model has not shrunk. Hugging Face counts bytes in the safetensors index, and four-bit values are packed two to a byte, so a four-bit export of a 36-billion-parameter model reads as about half. The same trap catches every packed-FP4 release.
The quantised builds are outrunning the originals
The NVFP4 export has been downloaded 10,604 times against 9,863 for the bf16 original. Add FP8, GPTQ-Int4 and GGUF and the five official repositories total 34,443 downloads. A community GGUF conversion has been pulled 198,963 times on its own — roughly 5.8 times the entire official collection.
That download pattern is the real signal
People are not evaluating Apodex 1.1 mini in a research notebook. They are quantising it and running it locally, which is exactly the use case the release argues for: a model small enough to self-host that keeps the coordination behaviour of the flagship.
What Apodex Says It Trained Differently
The arXiv report names two training directions and one runtime, and they are the company’s actual claim to novelty.
Working capability, defined
The paper’s term for the target is “working capability”: sustained, verifiable progress toward a real-world objective. The argument is that reasoning and synthesis are not enough, because complex work also needs sustained interaction with files, information sources and executable code, plus state maintenance, failure recovery and verifiable delivery.
Environment Scaling
The first direction widens the diversity and verifiability of the executable file, search and code environments the model learns inside. In plain terms, more kinds of messy real inputs during training, each with a checkable outcome. This is the half that produces Apodex 1.1 mini’s file-handling behaviour rather than its coordination.
Agentic Coordination Scaling
The second trains agents to decompose long-horizon tasks, delegate parallel work, integrate asynchronous results and replan. Apodex is explicit that the asynchronous Agent Team is a trained capability inside the model rather than an orchestration script wrapped around it — which is the distinction that would separate Apodex 1.1 mini from a prompt-engineering layer over any open model.
AgentOS and Statement Review
AgentOS is the runtime holding tool calls, file state and task progress across a whole job, and it manages the creation, scheduling and lifecycle of sub-agents. Statement Review is the delivery check: it verifies key claims against their supporting sources, data and computations before output, flags conflicts and keeps the review inspectable.
Adaptive Recovery
The named answer to the failure mode Apodex is selling against — a long task that collapses at step forty on a missing value or a failed assumption, wasting every hour spent getting there. Instead of stopping or returning something vague, the system is meant to explain the adjustment it is making and carry on.
One published run, for scale
Apodex describes a single Deep Discover task on its web product running 395 steps across 28 sources and returning a 31 KB report, with the reviewer checking 51 claims before delivery. That is a useful shape to have in mind, though it is one vendor-selected example rather than a measured average.
Deep Discover, and What Apodex 1.1 mini Does Not Get
The headline capability of the release is a mode called Deep Discover, and its availability is the sharpest line in the announcement.
What the mode does
In Deep Discover the model assembles an Agent Team without being told how: it decides whether a task splits, how many sub-agents run and when they consolidate. The team is asynchronous, so each branch reports into a shared task state as it finishes rather than at the end, and a live task board carries plans, steps and exceptions.
Mid-run changes do not restart the job
Adding a file or changing a requirement mid-run keeps the intermediate results that still hold and replans only the affected parts. That is the behaviour most orchestration frameworks handle worst, and it is the clearest reason to care about the coordination being trained in.
The tier split
| Tier | Who can use it | Open weights |
|---|---|---|
| Deep Discover | Preview programme, Apodex Frontier | No |
| Deep Research | Open to everyone | No |
| Deep Solve | Open to everyone | No |
| Agent Team via FrontierAgent | Anyone, self-hosted | Yes, Apache 2.0 |
The consolation is the framework
Deep Discover as a product sits behind a preview programme you have to be admitted to. The coordination layer, though, is what FrontierAgent’s Agent Team mode exposes, and Apodex says Apodex 1.1 mini carries the same layer as the flagship. So the self-hosted route gets the mechanism without the hosted product wrapped around it.
Where the Downloads Are Going
The download distribution across the five official repositories is the only genuinely independent adoption signal in this release.
The split, by build
Seven in ten downloads are quantised
Adding NVFP4, FP8, GPTQ-Int4 and the official GGUF gives 24,580 of 34,443 downloads, which is 71.4%. Only 28.6% of people taking Apodex 1.1 mini want the full-precision weights. That is a local-inference audience, not a fine-tuning one.
The community conversion dwarfs all of it
One third-party GGUF repository reports 198,963 downloads against 34,443 for the five official builds combined — a factor of 5.8. Community conversions typically surface in desktop inference tools before official ones do, which is where the rest of the demand appears to be sitting. The same dynamic played out around distillation debates we covered in Anthropic detailing campaigns from Alibaba, Moonshot and DeepSeek.
What the numbers are not
Hugging Face download counts include automated pulls and repeated fetches from CI. They are directional, not a user count, and they say nothing about whether anyone kept using Apodex 1.1 mini after the first run.
Running Apodex 1.1 mini Yourself
The practical requirements for Apodex 1.1 mini are modest by frontier standards but not as light as the announcement suggests.
Serving the model
Apodex recommends SGLang or vLLM for an OpenAI-compatible endpoint. Both example commands use tensor parallelism of 8 and a context length of 262,144, with qwen3_coder as the tool-call parser and qwen3 as the reasoning parser. Eight-way tensor parallel at bf16 is a multi-GPU node, which is why the quantised exports matter so much.
Recommended sampling
The card specifies temperature 1.0, top-p 0.95, repetition penalty 1.05, a maximum context of 262,144 and a 32,768-token output cap. Apodex 1.1 mini follows the Qwen3.5 chat template, emitting tool calls as <tool_call><function=…> and reasoning inside <think> tags.
Pass tools properly or lose the format
The card is emphatic that tool schemas go through the tools= parameter of the chat-completions API rather than being inlined in the system prompt. Inline them and the chat template cannot emit the correct call format, and the server-side parser cannot recover structured tool_calls. This is the single most likely way to get poor results from Apodex 1.1 mini.
Installing FrontierAgent
The requirements are Git, Python 3.12, uv and an OpenAI-compatible endpoint, with Docker optional. The sequence is a clone, uv sync --python 3.12 --extra dev, copying .env.example to .env, adding your endpoint, then uv run frontier-agent --mode react or --mode agent_team.
The API route
Apodex is offering free access on its API platform, with list prices quoted at $0.30 per million input tokens for the flagship and $0.10 for Apodex 1.1 mini — a three-to-one ratio. The repository advertises the free window as two weeks, which given a 24 August release means it may already have closed.
Where the Apodex 1.1 mini Coverage Drifts From the Repositories
Several claims in circulation about Apodex 1.1 mini do not survive a look at the source, and one of the write-ups carrying them is marked sponsored.
“One command on macOS and Linux”
Both the company blog and the TestingCatalog write-up describe FrontierAgent as starting with one command and no pre-installation. The repository’s own quick start is a clone, a uv sync, a config file copy, an edit and then the run command, with Python 3.12 and uv as prerequisites. Five steps, not one.
The benchmark caption
TestingCatalog places the 38.5 and 63.3 figures under a heading reading “Apodex 1.1 mini benchmarks”. Those are the flagship’s Agent Team scores. The mini’s own numbers, 27.7 and 50.2, are lower and are published separately on the model card.
The context window
Coverage cites a 260K context window. The configuration file sets max_position_embeddings to 262,144, which is 256K in the conventional sense and 262K if you are rounding raw tokens. Neither reading gives 260K.
The release date
The blog post is dated 15 September and the coverage ran on 16 September, but the arXiv report was submitted on 24 August, the GitHub repository was created on 22 August and the weights were last modified on 26 August. This is a three-week-old release being covered as news. Nothing wrong with that — just do not read the timing as a signal of freshness.
Who Should Care About Apodex 1.1 mini
The honest answer is narrower than the announcement implies, and it depends on what you already run.
Teams that cannot send data to a frontier API
This is the strongest case. An Apache 2.0 model with a declared base, a 256K window and a coordination layer trained in rather than scripted around it is a credible on-premises option for regulated work. That argument holds whether or not the benchmark numbers hold.
Anyone building multi-agent systems
FrontierAgent is worth reading even if you never run Apodex 1.1 mini. The task board, the three-directory sandbox, the fail-closed authorisation and the queued asynchronous intervention are design patterns you can lift. Our coverage of BuildBetter’s outcome-priced agent shows a different commercial answer to the same coordination problem.
The open-weight argument
A 36-billion-parameter model reaching 92% of a 397-billion-parameter sibling on one benchmark is the distillation case in miniature, and it is the pattern Y Combinator’s Garry Tan has been urging on American labs — see our write-up of his call for US open-weight labs to distill frontier models.
Who should wait
If you are already served by a hosted frontier model and have no data-residency constraint, there is no urgency here. Apodex 1.1 mini is a competent open-weight agent model with vendor-run benchmarks and a three-week-old framework. Apodex says pretraining for 2.0 is underway, built around what it calls a Heavy-Duty Solver, so the ground will move again.
References
Apodex-1.1-mini model card (Hugging Face)
FrontierAgent runtime and terminal interface (GitHub)
Apodex 1.1: Scaling Agentic Intelligence for Complex Work (arXiv)
Apodex 1.1 technical blog post (Apodex AI)
Apodex launches Apodex v1.1 Deep Discover model (TestingCatalog, sponsored)
Qwen3.5-35B-A3B base model (Hugging Face)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.