Apex Intelligence, a Beijing lab founded in June 2026 by Yongchao Chen, an assistant professor at Tsinghua University's College of AI, has released its first model — Apex-35B-V1.0. The 35B MoE model, built on Qwen3.6-35B-A3B, beats its own base on 20 of 21 public benchmarks and on several research tests outscores systems more than ten times its size. Weights are published under Apache 2.0.
Key takeaways
- 35B total parameters, about 3B active per token, a 262,144-token context window shared by input and output
- 20 wins out of 21 public benchmarks against the base Qwen3.6-35B-A3B — the single loss is LiveCodeBench v6 (82.3 vs 84.5)
- BFCL V4: 84.0 points, the highest score in the company's whole comparison, ahead of the 753B GLM-5.3 (75.8)
- BF16 weights of roughly 72 GB plus GGUF variants, all under Apache 2.0
- Angel?Angel: the earliest startup funding round, preceding Series A and angel-plus rounds: nearly US$50 million, co-led by IDG Capital and XtalPi, among others
Picture the following situation
An internal evaluation shows the model failing on fluid dynamics tasks: it picks the wrong numerical method and cannot fix its solution after seeing the simulation output.
In a conventional process a research team steps in: someone analyses the failures, builds new examples and reruns training.
Apex Intelligence says software took that role. Apex Research identifies the reasoning gap itself, generates and checks new scientific examples, then feeds them into the next round. On the internal CFD-Bench this produced a 32.1-point gain.
But a problem appears: how does anyone know that model-generated data teaches something rather than cementing the model's own mistakes?
The answer is what the company's whole construction rests on.
What exactly was released?
Apex-35B-V1.0 is a Mixture of Experts model: 35B parameters in total, only about 3B active per token. At inference it costs about as much as a 3B model while keeping a far larger one's capacity. The base is Qwen3.6-35B-A3B, an open MoE model from Qwen, trained further on proprietary datasets and research trajectories.
The context window is 262,144 tokens, shared between input and output, so a long reasoning chain eats the same budget as a long prompt. The BF16 weights take about 72 GB. The company also claims a dynamic thinking mode: the model decides each turn whether to reason and for how long.
Supported runtimes are SGLang, vLLM, TokenSpeed, KTransformers, Transformers and llama.cpp, the first three with ready Multi-Token Prediction commands. Alongside the BF16 weights there are GGUF builds from Q4_K_M up.
How does the self-improvement loop work?
According to the company, Apex Research works in three areas, each with a concrete number.
The first is data: the system hunts for the model's weak spots and generates training material aimed at them, as in the fluid dynamics example above.
The second is experimentation. Agent trajectories showed that overly long reasoning contexts and imbalanced data mixtures degrade agent behaviour, so the system tested other sequence-length distributions and sampling ratios.
The third is infrastructure. Optimising GPU kernels?GPU kernels: low-level compute routines executed directly on the graphics processor, workload balancing and KV-cache?KV-cache: a store of keys and values from earlier tokens so the model does not recompute them for every new token management delivered — on the company's figures — 45–60% higher training throughput, 1.6–1.8× faster decoding and 30–35% lower time to first token.
A simplified view of the described mechanism, not an actual training run.
What do the benchmarks say?
Against the base the lead is almost complete: 20 wins out of 21 benchmarks, the single loss being LiveCodeBench v6 (82.3 against 84.5).
Tools and research tasks
The strongest result is BFCL V4 — a test of whether a model correctly invokes tools, that is function calling. Apex-35B-V1.0 scores 84.0, best in the company's comparison, ahead of the 753B GLM-5.3 (75.8).
The research tests look similar. On FrontierScience (60 scientific problems graded against official rubrics?rubrics: fixed lists of scoring criteria used to grade an answer) it scored 40.0 against 30.7 for GLM-5.3, and on PaperBench, which measures reproducing a paper's results, 60.9 against 42.5.
General tasks
On general tasks the largest models still lead: Terminal-Bench 2.1 gives Apex 65.2 against 89.9, SWE-bench Verified 73.4 against 79.8. On MMLU-Pro, though, it beats even DeepSeek-V4.1-Flash-1017 (87.2 against 68.3).
Robotics
Robotics is a separate thread. On the internal RoboRSI, driving a robot in the RobotTwin and RoboCasa simulators, it scores 44.5 against 35.0 for the base, and on Real2Sim (a physics simulation rebuilt from one photograph) 21.6 against 2.1.
| Benchmark | Apex-35B-V1.0 | Comparison |
|---|---|---|
| BFCL V4 | 84.0 | GLM-5.3 — 75.8 |
| FrontierScience | 40.0 | GLM-5.3 — 30.7 |
| MMLU-Pro | 87.2 | DeepSeek-V4.1-Flash-1017 — 68.3 |
| LiveCodeBench v6 | 82.3 | Qwen3.6-35B-A3B — 84.5 |
| RoboRSI | 44.5 | Qwen3.6-35B-A3B — 35.0 |
| Real2Sim | 21.6 | Qwen3.6-35B-A3B — 2.1 |
Table: scores as reported by Apex Intelligence, one run per task (pass@1).
What do these names mean?
- MoE and active parameters — a model split into specialised subnetworks, only a few of which fire per token. Here 3B of 35B weights compute, and that share sets the running cost.
- RSI (recursive self-improvement) — an arrangement where the output of the model's research returns as an improvement to that same model. A feedback loop, not a one-off training run.
- Harness — the layer giving a model access to a terminal, files and tools. The compared models were evaluated inside Claude Code, except on BFCL V4.
Together they are the company's thesis: a cheap-to-run model can reach research level if its training data and configuration come from a closed measurement loop.
Why is ordinary fine-tuning not enough?
Standard further training: a team assembles a dataset, runs the training, measures the result and repeats. The bottleneck is not compute but human time.
Apex Intelligence moves that stage into software, but only where a hard metric exists: a fluid dynamics task can be simulated and its error computed, and a robot run in RobotTwin either succeeds or fails.
The second problem is subtler: a model generating its own data can amplify its own errors, producing examples consistent with what it already wrongly believes. Hence "validated" in the company's description — only results checked outside the model, by simulation, test or measurement, enter the next round.
That narrows the range: the loop works where a cheap, trustworthy judge exists. Without one, classic fine-tuning with people in the loop remains.
Is it working already?
Yes — the weights are publicly available. Apex-35B-V1.0 was published on the Hugging Face Hub with GGUF variants, under Apache 2.0, and the announcement on the company's site is dated 28 October 2026.
Not released: Apex Research. The system is described in company materials but published neither as code nor as a service, so verification rests on its own statements.
The benchmark results are the company's own, one run per task (pass@1), with no independent replication yet.
Why does this matter?
If the claims hold, the interesting part is not the model's size but the source of its edge: not more parameters, but better-chosen training data and configuration, selected automatically. That shifts the centre of competition from GPU access towards the quality of the measurement loop. For teams without billions to spend, it is one of very few realistic paths — for now a thesis, not an externally verified result.
What next?
- A further, larger self-evolving model has been announced, extending into chip design and quantum computing — with no date
- The August 2026 XtalPi partnership is meant to connect the model to automated lab infrastructure — the first test of the RSI loop outside simulation
- Independent replication is the credibility test, above all on BFCL V4 and PaperBench
Sources
- Apex Intelligence — Introducing Apex-35B-V1.0
- Hugging Face — Apexintelligence-AI/Apex-35B-V1.0
- Apex Intelligence — Apex Intelligence Raises Nearly US$50 Million in Angel Funding to Build Self-Evolving Foundation Models
- Apex Intelligence — Apex Intelligence Establishes a Strategic Partnership with XtalPi to Build an Open Ecosystem for AI for Science
- 机器之心 — 清华27岁助理教授杀入AI自进化赛道！首个35B模型刚刚亮相
