**17 October 2026 (Europe/Berlin) — Open-weights efficiency wave meets harness-sensitive ARC-AGI-3 saturation; closed frontier still leads AA Intelligence Index v4.3.2.**

### Frontier & markets
DeepSeek released V4.1-Flash-1017 (MIT, open weights) today. It is a post-training update on the same Causal Encoder-Decoder MoE backbone (552B backbone + Engram; ~8B prefill / 16B decode active). Company-reported gains include Terminal-Bench 4.0 42.3 (from 31.2), DeepSWE v1.1 77.0 (from 74.2), HLE (no tools) 42.6, and shorter outputs / faster execution. API remains under the deepseek-flash name at roughly prior pricing (~$0.09/$0.55 per million). **Confirmed** (Hugging Face model card + DeepSeek API changelog).

Reflection AI shipped Beam 1.0 open weights (Apache 2.0) on 16 October: sparse MoE, 501B total / 23B active, 1M effective context, text-only, BF16/FP8/NVFP4 + DFlash2 drafter. Aimed at coding/agentic work; ecosystem includes AWS SageMaker, vLLM, SGLang, Mirror CLI. AA Intelligence Index placement reported around the low-30s among US open weights. **Confirmed** (reflection.ai blog + Hugging Face).

Meta released Muse Spark Open (15 October): natively multimodal MoE, 816B total / ~42B active, 1M context, open weights (Muse Model License). Pitched as strongest US open-weight; runs on paired M5 Ultra Mac Studios via MLX. Hosted at various providers. Muse Wonder / Wonder Full (12 October) are Meta’s first claimed frontier agent models (Active Memory + Active Focus); Wonder is API-available (~$4/$20), Wonder Full is Muse Code-only. AA Index v4.3.2 shows Wonder Full max near 55.8, Spark Open lower (~39 range in earlier reads). **Confirmed** (Meta AI Research blogs).

NVIDIA shipped Nemotron 3.5 Super VL (120B/12B multimodal) and Bolt (74B-A9B) open variants (15 October). Google expanded Gemini 4 Argon access (introductory $2/$10). Anthropic’s Claude Haiku 5.5 (7 October) is the cheapest/fastest small Claude yet ($0.10/$0.50 up to 100k prompts; effort control; ~75% cheaper than Haiku 4.5 blended). **Confirmed**.

On Artificial Analysis Intelligence Index **v4.3.2** (do not mix versions), Claude Opus 5.5 (max / default fallback) leads at 57.6, followed by Claude Sonnet 5.5 max (~56), Muse Wonder Full max (~55.8), then GPT-6 Astra / Gemini 4 Argon / others in the low-50s. Open-weight leaders sit lower (MiMo-V2.6-Pro ~46, etc.). Saturated exams are not used for ranking. **Confirmed** (AA-derived leaderboards as of 16–17 Oct).

### Science of intelligence
ARC Prize verified Gemini 4 Argon on ARC-AGI: 98.5% (v1), 91.3% (v2), and 97.5% (v3, provider-adapter harness that preserves opaque reasoning + auto-compaction, ~$4.8k) versus ~49.8% on the standard harness. Second lab above 95% on v3 after GPT-6 Astra (99.9%). Action efficiency matched Astra’s 62% reduction vs human baseline on public games. Chollet noted similar performance profile to Astra. Harness dependence remains a core measurement issue. **Confirmed** (ARC Prize posts + results page).

DeepSeek’s update explicitly advances the performance–cost Pareto on agent harnesses (Terminal-Bench, DeepSWE, etc.) while cutting output length—evidence that post-training and architecture (CED + KV compression) still move the efficient frontier without new scale. No new primary theory papers from Amodei, Hassabis, Tenenbaum, or Ganguli surfaced today.

### Engineering
Open-weight releases emphasize agent runtimes (Mirror CLI, MLX desktop, SageMaker, vLLM/SGLang day-0 support), speculative decoding, and NVFP4/FP8 quantization. DeepSeek and Beam both target long-horizon agentic coding with reduced active parameters and KV footprint. Meta demonstrated on-device Muse Spark Open workflows (iOS app changes, video editing in DaVinci). NVIDIA’s Nemotron variants and Indic pretraining data pipeline (with AI4Bharat) continue the open data/model stack. Energy/inference economics remain first-class: efficiency gains (shorter outputs, lower active params) are the near-term lever.

### Political / legal / social
Open-weights momentum (Beam, Muse Spark Open, DeepSeek, Nemotron, Jais 3) strengthens the Western/Chinese open alternative narrative and lowers barriers for sovereign or on-prem deployment. Meta’s Muse Model License and Reflection’s Apache 2.0 + partner ecosystem illustrate different openness regimes. No major new regulator texts today; availability remains gated on some frontier models (Argon free/Plus still limited). Rumor-level items (OpenAI internal events) are ignored without primary confirmation.

### Watchlist hits
- DeepSeek V4.1-Flash-1017 model card & benchmarks: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017
- Reflection Beam announcement: https://reflection.ai/blog/beam-is-available
- Meta Muse Spark Open: https://research.meta.ai/blog/muse-spark-open
- ARC Prize Argon results: https://arcprize.org/results/google-gemini-4-argon and @arcprize posts
- AA Index v4.3.2 snapshots (secondary aggregators): benchleader.com / artificialanalysis.ai

### Ledger actions
- **Frontier**: Opus 5.5 (max) still tops AA Index v4.3.2; Argon joins the ARC-AGI-3 >95% club on provider-adapter harness only. Wonder Full is the closest Meta challenger. No single model is best without task/harness.
- **Efficient**: DeepSeek V4.1-Flash-1017 and Beam (23B active) move the cost/performance Pareto for agentic coding. Haiku 5.5 remains the volume small-model reference.
- **Budget**: Open-weight MoEs (Spark Open, Beam, Nemotron Bolt, DeepSeek Flash) plus Haiku 5.5 / Gemini Flash tiers dominate low-cost agent routes.
- Five-benchmark atlas impact: Terminal-Bench 4.0, DeepSWE, ARC-AGI-3 (specify harness), AA Index components, and agent coding suites all shifted by today’s open updates. Rebase any composite that includes Terminal-Bench or ARC-AGI-3 before ranking.

**Sources**: Primary lab posts and model cards linked above; ARC Prize; Artificial Analysis-derived public leaderboards (v4.3.2 only).

Ask Intelligence Radar a task for a Task Router ticket (5 benchmarks → Frontier / Efficient / Budget + optional decomposition).
