Apex Intelligence's first model: continued training on Qwen3.6-35B-A3 using proprietary data and research trajectories, targeting coding, tool invocation, deep search, and research tasks such as fluid dynamics and robot simulation that require "writing code, running it, checking results, and iterating"; it autonomously decides whether to think and for how long in each turn, without manual mode switching. Officially, the development used their proprietary automated research system, Apex Research, to generate training data, test training recipes, and optimize kernels. Released on 10 28, not included in the last 7 periods.
Rising score 53 · Likes 58 · Downloads 1219
Reasons
- Officially, the largest improvements over the base Qwen3.6-35B-A3 are in agent and research metrics: BrowseComp 69.4 (base 42.1), AutomationBench 49.3 (10.0), SciCode main problems 62.6 (37.5), FrontierScience 40.0 (21.7), SWE-bench Pro v2 92.4 (72.4), Terminal-Bench 2.1 65.2 (46.1); BFCL V4 84.0 is the highest among all models in the table.
- Deployment instructions are detailed: startup commands for SGLang, vLLM, TokenSpeed, KTransformers, Transformers, and llama.cpp are provided, including pitfalls like driver versions and differing parser names across frameworks; officially recommended to use Claude Code (v2.1.280) as the agent framework.
- Quantization is comprehensive: there is also a GGUF repository, with BF16 plus 28 quantization types, Q4_K_M at 21.7 GB, and the smallest IQ1_S at 9.0 GB.
Usage
- vLLM ≥ 0.32.0: vllm serve Apexintelligence-AI/Apex-35B-V1.0 --served-model-name Apex-35B-V1.0 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder. BF16 is approximately 72 GB; a single card with 141 GB or more can run full context, H100 NVL requires 2 cards, and A100 40GB requires 4 cards.
- Local: download Q4_K_M from Apexintelligence-AI/Apex-35B-V1.0-GGUF, llama-server -m Apex-35B-V1.0-Q4_K_M.gguf --jinja -ngl 99 -c 32768; you must add --jinja to use the built-in chat template, do not switch to the Qwen3.6 template.
- Sampling temperature 1.0, top_p 0.95, top_k 20, max_tokens 65536 (including thinking); for multi-turn and tool calls, pass back the complete assistant message as-is, and set preserve_thinking=True in chat_template_kwargs.
Risks
- All scores are official self-tests, among which RoboRSI, Real2Sim, and CFD-Bench are internal benchmarks; "recursive self-improvement" is an official claim, with no method details provided in the card, and training data is not public.
- In the same table, much larger models like GLM-5.3, GLM-5.3-Flash, and DeepSeek-V4.1-Flash-1017 still score significantly higher on most metrics, for example Terminal-Bench 2.1 is 89.9 / 84.3 / 90.6; LiveCodeBench v6 82.3 is still slightly lower than the base's 84.5.
- Supports only text input and output; this is the first model from a new company, and future maintenance status is unknown.