![Apex-35B Debuts: 35B Parameters, 3B Active [2026]](https://shattered.io/wp-content/uploads/2026/10/apex-intelligence-apex-35b-v1-0-35b-3b-active-automated-research-2026-gen.webp)
Apex Intelligence announced Apex-35B-V1.0 in Beijing on October 29, 2026, and the pitch is unusual. According to the company’s announcement carried by The Straits Times, it is a general-purpose model built with automated AI research, and it is the first model the company has shipped.
The model starts from Qwen3.6-35B-A3B, a sparse mixture-of-experts base with 35 billion total parameters and 3 billion active per token. Apex Intelligence then pushed it through mid-training and post-training on proprietary datasets and research trajectories. That is a familiar recipe for the open-weight crowd, which we mapped in our comparison of six open-weight model families. The twist is the system the company says did much of the work.
This analysis separates what Apex Intelligence has stated from what nobody has verified yet. The headline numbers on speed come from the company itself, and the announcement was distributed as a paid press release. Independent benchmarks, pricing and named spokespeople are absent from the records reviewed here. We say so wherever it matters.
Don't miss new tech stories on Google
Add Shattered once in the Google app and our stories show up in your news suggestions.
What Apex Intelligence Announced on October 29 #
The facts that hold up are short. Apex Intelligence, a company that announced from Beijing, released Apex-35B-V1.0 on October 29, 2026. The company describes it as a general-purpose model and as its first model. The same announcement, republished by PR Newswire APAC, frames the release around a process rather than a single benchmark score. The process is automated AI research.
The Straits Times page lists the item under paid press releases and states that the issuer is solely responsible for the content. That label matters for how readers should weigh the claims. A paid release gives the company full control over wording, timing and emphasis, and no reporter tested anything before it went live.
Still, the release carries concrete, checkable artifacts. The weights sit on Hugging Face, and the code, deployment guidance and usage examples sit on GitHub. Anyone with enough hardware can download the model and measure it. That openness is the best argument for taking the release seriously, because it turns marketing claims into testable ones within days.
Apex-35B-V1.0 Specs at a Glance #
The table below collects every specification confirmed in the announcement records. Where a detail is missing, it does not appear. The model card lives on the company’s GitHub repository, which points readers toward the weights and deployment notes.
| Attribute | Apex-35B-V1.0 |
|---|---|
| Developer | Apex Intelligence | | Announced | October 29, 2026, Beijing |
| Base model | Qwen3.6-35B-A3B |
| Architecture | Sparse mixture-of-experts (MoE) |
| Parameters | 35B total, 3B active per token | | Context length | 262,144 tokens (input plus output) | | Modalities | Text in, text out | | Thinking mode | Dynamic | | Weight precision | BF16 | | Approximate BF16 weight storage | 72 GB | | Weights | Hugging Face | | Code and guides | GitHub | | Price | Unconfirmed |
Two details deserve a second look. The 262,144-token window counts input and output together, so a long prompt eats into the room left for the answer. And the model is text only. Teams that need image or audio input will have to look elsewhere or bolt on another component.
Built on Qwen3.6-35B-A3B: The Lineage #
Apex-35B-V1.0 is not a from-scratch pretrain. Apex Intelligence says it took Qwen3.6-35B-A3B and continued training it. That is a legitimate and common path, since a strong base saves months of compute. It also means the model inherits the base’s architecture, tokenizer family and most of its knowledge. Whatever the new model does better, it does on top of someone else’s foundation.
The name tells you the shape. The 35B figure is total parameters, and A3B signals roughly 3 billion active per token. Readers following the Qwen, DeepSeek and Mistral families can see how this base compares in our DeepSeek vs Qwen vs Mistral breakdown. A 35B sparse model sits at the small, cheap end of that spectrum.
What Mid-Training and Post-Training Add
The announcement names two stages. Mid-training sits between the original pretraining and the final alignment, and teams use it to push in domain data at scale. Post-training covers the instruction tuning and reinforcement steps that shape behavior. Apex Intelligence says both stages drew on proprietary datasets and research trajectories.
Research trajectories are the interesting ingredient. In practice they are step-by-step records of how a task was worked through, such as an experiment being planned, run and analyzed. The announcement does not publish them, and the datasets stay proprietary. So the open part is the weights, while the recipe that made them different stays closed.
How a Sparse MoE Gets 3B Active From 35B #
A sparse mixture-of-experts model splits its feed-forward layers into many expert blocks. A small router picks a few experts for every token, and the rest sit idle. With 35 billion parameters in total and 3 billion active per token, only about 8.6% of the weights do work on any given step. The arithmetic is simple: 3 divided by 35.
That ratio explains the appeal. Compute per token tracks the active count, so decoding costs look like those of a 3B dense model. Memory tells a different story. Every expert must stay loaded, which is why the weights need about 72 GB in BF16. You pay for the big model in storage and the small model in arithmetic.
The dynamic thinking mode adds another lever. A model with dynamic thinking decides how much reasoning to spend on a given prompt. Easy questions get quick answers, and harder ones trigger longer chains. For agent loops that call a model hundreds of times, that choice can shift cost by a wide margin. Apex Intelligence has not published figures on how often the mode triggers.
Apex Research: The Automated AI Research System #
The most distinctive claim in the announcement is about how the model was made. Apex Intelligence says it built a system called Apex Research, which it describes as an automated research system designed to enable recursive self-improvement. The announcement lists the activities the system automates:
- Data construction
- Experiment design
- Code execution
- Training-result analysis
- Training optimization
Read that list as a loop. The system builds data, designs an experiment, runs the code, studies the result and tunes the next training run. A human team used to own each of those steps. Handing them to software compresses the cycle time, at least in principle, and it is the core of the automated AI research story.
Be careful with the phrase recursive self-improvement. It is the company’s description of what Apex Research is designed to enable. The announcement does not show a model rewriting its own training code unsupervised, and it does not publish evidence of compounding gains across generations. Design intent and demonstrated outcome are different things, and only the first has been stated.
The Performance Claims Under the Microscope #
Apex Intelligence reports gains in both training and inference, each measured under comparable workloads. The company supplied ranges, not single figures. The table translates them into plain multipliers so readers can judge the scale.
| Metric | Reported gain (per Apex Intelligence) | Plain-English reading |
|---|---|---|
| Training throughput | 45% to 60% higher | About 1.45x to 1.6x the tokens per unit of time |
| Decoding speed | 1.6x to 1.8x faster | A 100-token reply finishes in roughly 56% to 63% of the time |
| Time to first token | 30% to 35% lower | A 1-second wait drops to about 0.65 to 0.70 seconds |
The right-hand column is our arithmetic, not a company statement. The left-hand figures come only from Apex Intelligence, as relayed in the announcement coverage. No third party has reproduced them.
Why Comparable Workloads Is the Phrase to Watch
Every performance figure carries the qualifier comparable workloads. The announcement records do not say what the comparison point is. It could be the Qwen3.6-35B-A3B base, an earlier internal pipeline or a different serving stack. A 1.7x decoding gain over a poorly tuned baseline means little, while the same gain over an optimized one would be striking.
Hardware, batch size and sequence length also swing these numbers. Time to first token depends heavily on prompt length, and decoding speed depends on batching. Until Apex Intelligence or an outside lab publishes configurations, treat the ranges as directional. Run your own tests on your own prompts before planning capacity around them.
Hardware Footprint: What 72 GB of BF16 Weights Means #
BF16 stores each parameter in 2 bytes. Multiply 35 billion by 2 and you get 70 GB, which lines up with the approximate 72 GB figure in the announcement once you add small overheads. That number covers the weights alone. The KV cache for a long context, plus runtime buffers, comes on top.
In practice, a single accelerator with 80 GB of memory leaves very little headroom for a 262,144-token window. Teams will likely split the model across two cards or use quantized variants. The announcement records mention only BF16 weights, so any smaller build would come from the community or from later releases.
For local experiments, the choice of runtime matters as much as the card. We compared the main options in our guide to Ollama, LM Studio, llama.cpp and vLLM. Sparse MoE models often behave differently across those runtimes, so test more than one before drawing conclusions about speed.
Target Workloads: Coding, Agents, Question Answering, Deep Search #
The announcement points the model at four jobs: coding, agent workflows, question answering and deep search. The first two share a need for long, structured context. An agent that reads a repository, edits files and reruns tests burns thousands of tokens per step, so the 262,144-token window fits the pitch.
Deep search is the least standard entry. It usually means a model that issues many queries, reads results, follows leads and assembles an answer. That workload rewards cheap tokens and fast first responses, which is where a 3B-active model and the claimed 30% to 35% cut in first-token time would matter most.
No benchmark scores for any of these tasks appear in the cited records. So we cannot say whether Apex-35B-V1.0 beats its base on coding or agent tasks, or by how much. The pitch is coherent, but the proof is still missing.
Availability: Hugging Face, GitHub and the Missing Price #
Distribution follows the standard open-weight pattern. The weights are on Hugging Face, and the accompanying code, deployment guidance and usage examples are on GitHub. For developers, that means no waitlist, no gated API and no per-token meter on day one. You download, load and measure.
Two things remain unconfirmed. First, no price appears in the cited records, so there is no hosted offering we can cost out. Second, the records reviewed here do not spell out a license, and licensing decides whether a company can ship the model in a commercial product. Read the model card before building on it.
Competitive Context: Where a 35B Sparse Model Fits #
Apex-35B-V1.0 enters a crowded open-weight field with a modest footprint. At the top end, labs now ship models with hundreds of billions or even a trillion parameters. We covered one such release in our report on Mistral Large 4 and its 1T parameters. Against that scale, a 35B model targets a different buyer.
That buyer wants good-enough quality at low serving cost, and wants to run the model on hardware they control. The 3B-active design gives a cost profile closer to a small dense model, while the 35B total keeps knowledge capacity higher than a true small model. Our running tally of the field lives in the October roundup of open-source LLMs.
The honest comparison is against its own parent. Qwen3.6-35B-A3B is the model Apex Intelligence fine-tuned, so the fair test is whether the derivative beats the base on the tasks the company advertises. If it does not, buyers can simply use the base. If it does, the automated research pipeline has earned its headline. Nobody has run that head-to-head in public yet.
Historical Context: From Hand-Tuned Training to Automated Loops #
Machine learning teams have long automated pieces of the pipeline. Hyperparameter search, neural architecture search and automated data filtering all predate the current model wave. What changed in the last few years is that language models themselves became capable enough to write experiment code, read logs and propose the next run.
Apex Research sits at the end of that arc. It bundles data construction, experiment design, code execution, result analysis and training optimization into one system. Earlier tools handled a single stage. The claim here is end-to-end coverage, which is a bigger promise and a harder one to verify from outside the company.
The geopolitical frame also matters. Chinese labs have leaned on open weights as strategy, a theme we explored in our piece on China’s open-weight AI approach and the DHS DeepSeek order. A Beijing-announced release built on a Qwen base fits that pattern, and it may draw the same scrutiny from security teams that earlier releases did.
What Is Still Unconfirmed #
A tidy ledger helps here. The cited records give solid specifications and distribution details, and they leave everything about independent validation open. No named executive, researcher or outside expert appears with an on-record statement, so this article carries no quotes section. Inventing one would mislead readers.
| Question | Status | What the records say |
|---|---|---|
| Parameter counts and architecture | Confirmed | 35B total, 3B active, sparse MoE |
| Base model | Confirmed | Qwen3.6-35B-A3B |
| Weights and code availability | Confirmed | Hugging Face and GitHub |
| Training and inference gains | Company claim | 45% to 60% training, 1.6x to 1.8x decoding, 30% to 35% lower first-token time |
| Independent benchmark results | Unconfirmed | None in the cited records |
| Price or hosted API | Unconfirmed | No price appears |
| Named spokesperson quote | Unconfirmed | No verified on-record speaker |
| Recursive self-improvement in practice | Unconfirmed | Described as a design goal of Apex Research |
Market Impact for Developers and Buyers #
The immediate market effect is small, and that is worth saying plainly. A first model from a new company, announced through a paid release, does not move chip stocks or reorder the leaderboard. Its significance is as a signal about method, and the method is what other labs will watch.
For engineering teams, the practical impact comes in three parts. Open weights lower the cost of experimenting, since nobody needs a contract to try the model. A 3B-active design keeps serving bills in check once you clear the memory hurdle. And a 262,144-token window opens up repository-scale coding and long-document work. Teams already running agent stacks should add it to their evaluation queue.
For buyers of closed APIs, the pressure is indirect. Every competent open-weight model that serves agent and coding workloads at low cost strengthens the case for self-hosting. The more automated research shortens the path to such models, the faster that pressure builds. Our [AI and machine learning hub](https://shattered.io/ai-machine-learning/) tracks how those dynamics play out across vendors.
Five Predictions for Apex-35B-V1.0 and Automated Research #
These are our judgments, not reported facts. We frame each as a testable bet so readers can check us later.
- Independent benchmarks will land within weeks. Open weights invite quick testing, and community evaluators usually publish coding and agent results fast. Expect the first third-party numbers before the end of the year.
- The head-to-head with the Qwen3.6 base will decide its reputation. If the derivative wins clearly on coding and agents, the automated research pitch gains credibility. If the gap is small, the story shrinks to a fine-tune.
- Quantized builds will appear quickly. A 72 GB BF16 footprint is awkward for many labs, so community quantizations are likely to follow soon after release.
- Other labs will advertise their own automated research loops. Once one company markets the process as the product, rivals tend to describe similar internal systems.
- Claims of recursive self-improvement will face pressure for evidence. Without published results showing gains compounding across versions, skeptics will keep calling the phrase marketing.
Risks and Open Questions for Adopters #
Open weights shift responsibility onto the user. When a company hosts a model, it owns the monitoring, the patching and the abuse controls. When you download the weights, those jobs become yours. Teams in regulated sectors should ask what safety evaluation, if any, Apex Intelligence ran before release. The records reviewed here describe none.
Provenance is the second question. The training data is proprietary and the research trajectories are unpublished, so outsiders cannot audit what went in. That limits how far a compliance officer can vouch for the model. A release that publishes weights but hides the data is open in one sense and opaque in another.
The third risk is dependency. Apex-35B-V1.0 is a derivative of Qwen3.6-35B-A3B, so any weakness in the base carries over. If the base model changes its terms or gets superseded, the derivative inherits the fallout. Teams should pin exact versions of the weights and keep a fallback model ready.
Finally, version numbers carry a hint. The V1.0 tag marks a first release from a first-time model maker. Early versions often ship with rough edges in tool calling, long-context stability and multilingual behavior. None of that is confirmed for this model, but a cautious rollout makes sense. Start with a shadow deployment, compare outputs against your current model for a few weeks and promote it only when the data supports the move.
How to Evaluate Apex-35B-V1.0 Yourself #
Since the weights are public, a disciplined test plan beats waiting for headlines. Start by pulling the repository and reading the deployment guidance on GitHub. Then run three comparisons against Qwen3.6-35B-A3B using identical prompts and identical hardware.
- Measure time to first token across short, medium and very long prompts.
- Measure decoding speed at several batch sizes with the same serving stack for both models.
- Score a fixed set of your own coding and agent tasks, not a public benchmark the model may have seen.
Log memory use at the full 262,144-token window, since long contexts stress the KV cache more than the weights. If your numbers land inside the claimed 1.6x to 1.8x decoding range, the company’s figures hold for your setup. If they do not, you will know exactly where the claim breaks.
Frequently Asked Questions #
What is Apex-35B-V1.0?
Apex-35B-V1.0 is a general-purpose text model from Apex Intelligence, announced October 29, 2026, in Beijing. It is the company’s first model and uses a sparse mixture-of-experts design with 35 billion total parameters and 3 billion active per token.
What base model does Apex-35B-V1.0 use?
It builds on Qwen3.6-35B-A3B. Apex Intelligence says it applied mid-training and post-training using proprietary datasets and research trajectories.
What is Apex Research?
Apex Research is the company’s automated research system. It handles data construction, experiment design, code execution, training-result analysis and training optimization, and the company says it is designed to enable recursive self-improvement.
How much memory does the model need?
The BF16 weights take roughly 72 GB of storage. Running the full 262,144-token context needs additional memory for the KV cache, so plan for more than the weight size alone.
How much faster is it, according to Apex Intelligence?
The company reports 45% to 60% higher training throughput, 1.6x to 1.8x faster decoding and 30% to 35% lower time to first token under comparable workloads. No independent test has confirmed those figures yet.
Where can I download it?
The model weights are on Hugging Face. Code, deployment guidance and usage examples are on the company’s GitHub repository.
Does it have a price?
No price appears in the cited records. The weights are downloadable, so self-hosting costs are your own hardware and power, while any hosted pricing remains unannounced.