explainx / blog
Apodex shipped 1.0 with open 35B Apodex-1.0-mini — #1 on FutureX at 59.17, ahead of Claude Sonnet 4.6, DeepSeek-V4-Pro, GPT-5.5. Verification-centric deep research, AgentHarness, vLLM setup.
Jun 30, 2026
InternScience's Agents-A1 claims trillion-class agent performance at 35B MoE scale — BrowseComp 75.5, GAIA 96.0, IFEval 94.8 — with open weights on Hugging Face and ModelScope. Here is what the paper claims, how to serve it, and the honest gaps on coding and Qwen lineage.
Jun 1, 2026
Hermes WebUI brings autonomous AI agents to your browser with persistent memory, self-improving skills, and 24/7 scheduled jobs. Unlike Claude Code or OpenClaw, Hermes learns your environment and gets more capable the longer it runs - all self-hosted on your infrastructure.
May 22, 2026
Qwen 3.7-Max is built for the agent era, excelling in SWE-bench, Terminal Bench, and sustained execution across thousands of tool calls. Discover how environment scaling and cross-harness generalization drive its frontier performance.
June 29, 2026 — 9:31 PM: Apodex posted on X that since shipping Apodex 1.0, its open 35B Apodex-1.0-mini has been outperforming models many times its size on FutureX — three #1 finishes in four weeks, with fresh scores of 59.17 (#1) and 58.42 (#2), ahead of Claude-Sonnet-4.6 (56.32), DeepSeek-V4-Pro (53.58), and GPT-5.5 (52.51).
Apodex calls itself a Self-Evolving Heavy-Duty Solver — built for questions whose answer isn't written down yet. That is future prediction, geopolitics, market moves, and open research problems where retrieval plus verification beats param count.
| Question | Answer |
|---|---|
| What model is it? | Apodex-1.0-mini — 35B-A3B MoE, Qwen3.5 base, Apache 2.0 |
| Where to try? | apodex.ai — Deep Research mode uses Apodex-1.0-mini |
| FutureX scores? | #1 @ 59.17, #2 @ 58.42 (June 29 leaderboard) |
| Open weights? | Hugging Face — apodex/Apodex-1.0-mini |
| Why 35B beating Sonnet is wild? | Harness + verification-centric training — not raw logits alone |
| FIFA 2026 winner? | Apodex replied "checking our notes 👀" — the benchmark vibe in one reply |
Apodex 1.0 is a verification-centric agent stack, not a single checkpoint drop:
| Component | Role |
|---|---|
| Apodex-1.0-mini | Open 35B-A3B weights — standard ReAct tool agent |
| Apodex-1.0-H | Heavy-duty mode — async agent team, shared evidence pool, global verifier |
| Apodex-1.0-0.8B / 2B / 4B SFT | Smaller open models trained on deep-research SFT data alone |
| AgentHarness | Open eval repo — reproduce BrowseComp, DeepSearchQA, HLE, etc. |
| AgentOS | Task-agnostic runtime for building and evaluating agent workflows |
| apodex.ai API | Product surface — Deep Research powered by mini |
Paper title: Apodex-1.0: A Verification-Centric Agent Team for Discoverative Intelligence (tech blog, Hugging Face model card).
Most deep-research agents conflate finding and trusting. Apodex splits the loop:
In heavy-duty deployment, Apodex claims coordination of up to 150 sub-agents over 15,000 steps on a single task — test-time scaling as architecture, not just more tokens.
That rhymes with Perplexity's agentic search thesis (search as code, evidence chains) and the agent memory systems debate — agents need data management, not just longer context.
FutureX evaluates future prediction — forecasting outcomes before ground truth exists. That is deep research under uncertainty: track sources, detect regime change, synthesize conflicting signals, and commit to a calibrated answer.
Apodex submitted four experimental prediction harnesses built only on Apodex-1.0-mini. On the June 29 board they occupied #1 and #2 (and previously held top four per LinkedIn launch copy).
| Rank (Jun 29) | Score | Model / harness |
|---|---|---|
| #1 | 59.17 | Apodex harness (Apodex-1.0-mini) |
| #2 | 58.42 | Apodex harness (Apodex-1.0-mini) |
| — | 56.32 | Claude-Sonnet-4.6 |
| — | 53.58 | DeepSeek-V4-Pro |
| — | 52.51 | GPT-5.5 |
explainx.ai read: FutureX rewards evidence discipline, not memorization — the same axis where BrowseComp and DeepSearchQA live. A 35B open model leading here is a procurement signal for teams blocked from Fable/Mythos who still need frontier-grade research loops.
From the Hugging Face model card — open checkpoints on the four-benchmark suite:
| Model | BrowseComp | BrowseComp-ZH | HLE-Text | DeepSearchQA |
|---|---|---|---|---|
| Apodex-1.0-mini | 71.5 | 80.6 | 46.8 | 82.2 |
| Apodex-1.0-4B-SFT | 48.8 | 63.5 | 32.9 | 69.9 |
| Apodex-1.0-2B-SFT | 27.9 | 35.0 | 18.2 | 49.9 |
| Apodex-1.0-0.8B-SFT | 13.9 | 10.7 | 11.2 | 25.8 |
Apodex-1.0-H (heavy agent team, not weights alone) pushes further — 90.3 BrowseComp, 94.4 DeepSearchQA, 60.8 HLE-text, 46.7 FrontierScience-Research, per the launch blog — edging GPT-5.5 on BrowseComp in Apodex's table and beating Claude Opus 4.8 and Kimi K2.6 on DeepSearchQA.
Data > params at small scale: Apodex-1.0-4B-SFT beats every open 30B-class model on BrowseComp and BrowseComp-ZH in their comparison — a direct challenge to "just scale parameters" thinking in the China free-models playbook.
Evaluations block benchmark-hosting websites during runs to reduce answer leakage from public repos — standard hygiene, worth noting when comparing vendor numbers.
Qwen/Qwen3.5-35B-A3B-Base → Qwen/Qwen3.5-35B-A3B → apodex/Apodex-1.0-mini
Post-training recipe (from model card):
Apodex claims additive gains — general knowledge (MMLU-Pro/Redux, C-Eval), math (AIME, HMMT), instruction-following (IFEval, IFBench), and long-context (LongBench v2) stay within ~1 point of matched Qwen3.5 bases. Coding is preserved too — Apodex-1.0-H reports 79.0 SWE-bench Verified and 58.4 Terminal-Bench v2 in the launch blog.
For local MoE coding trade-offs on the same Qwen family, see Qwen 3.6 27B dense vs 35B A3B — different fine-tune, same hardware class.
| Apodex-1.0-mini | Agents-A1 | |
|---|---|---|
| Ship date | Apodex 1.0 + FutureX post Jun 29 | ModelScope Jun 30 |
| Base | Qwen3.5-35B-A3B | qwen3_5_moe |
| Pitch | Verification-centric deep research | Heterogeneous long-horizon agent |
| BrowseComp | 71.5 (mini ReAct) | 75.51 (vendor table) |
| Standout | FutureX #1, DeepSearchQA 82.2 | GAIA 96, IFEval 94.8 SOTA |
| Open eval | AgentHarness | Agents-A1/evaluation |
| Product | apodex.ai Deep Research | ModelScope / HF weights |
Neither replaces Kimi K2.7-Code or LongCat-2.0 for repo-scale coding — pick by task axis.
python3 -m sglang.launch_server \
--model-path apodex/Apodex-1.0-35B-A3B \
--tp 8 \
--host 0.0.0.0 \
--port 1234 \
--context-length 262144 \
--tool-call-parser qwen3_coder \
--reasoning-parser qwen3
vllm serve apodex/Apodex-1.0-35B-A3B \
--tensor-parallel-size 8 \
--max-model-len 262144 \
--enable-auto-tool-choice \
--tool-call-parser qwen3_coder \
--reasoning-parser qwen3
| Parameter | Value |
|---|---|
temperature | 1.0 |
top_p | 0.95 |
repetition_penalty | 1.05 |
max_context_length | 262144 |
max_tokens | 32768 |
Apodex uses the Qwen3.5 chat template — tool calls as <function=...> blocks; pass schemas via OpenAI tools= parameter, not inlined in system prompt. Full Python agent loop is on the Hugging Face card.
git clone https://github.com/ApodexAI/AgentHarness
# Download benchmark pack from Hugging Face datasets/apodex/Deep-Research-Benchmarks
# See AgentHarness README for ReAct eval protocol
Supported suites include BrowseComp, BrowseComp-ZH, DeepSearchQA, HLE (text), FrontierScience, SuperChem, WideSearch — see AI benchmarks guide for how to read each.
Point OpenAI-compatible endpoints at your server — same pattern as OpenCode local stack or Codex OSS. For MCP-heavy research, pair with Claude Code MCP patterns even if the model is Apodex-hosted.
X confirmed: Apodex-1.0-mini powers Deep Research mode on apodex.ai.
Use cases Apodex targets:
That is a different buyer than "fast codegen" — closer to enterprise Fable alternatives research tiering: Apodex / DeepSeek / GLM API for volume + verify, closed frontier for edge cases if accessible.
| Concern | Notes |
|---|---|
| Harness vs model | FutureX scores reflect prediction harness + mini, not bare chat completion |
| Heavy vs mini gap | 90.3 BrowseComp is Apodex-1.0-H team mode — not downloadable as one file |
| Benchmark leakage | Apodex blocks hosts; other labs may not — compare under AgentHarness when possible |
| Sonnet 4.6 naming | Verify exact API snapshot and harness parity on futurex.live |
| Self-evolving claim | Marketing term — read tech report for what actually updates (data, verifiers, policies) |
Apodex-1.0-mini is the strongest public signal yet that open 35B + verification architecture can beat closed Sonnet-, DeepSeek-, and GPT-class entries on future-facing research evals — not just static knowledge tests.
If you build research agents in 2026, run three checks:
The 35B class is having a week: Apodex on FutureX (Jun 29), Agents-A1 on GAIA (Jun 30), Qwen 3.6 27B dense for local coding (Jun 29–30). Pick the axis you actually ship.
Official: apodex.ai · Hugging Face — Apodex-1.0-mini · FutureX leaderboard · AgentHarness GitHub · Discord
FutureX scores and benchmark tables reflect Apodex's June 29, 2026 public posts and Hugging Face model card — leaderboard positions change daily. Verify live at futurex.live before citing ranks. Last updated: June 29, 2026.