FI daily research brief 2026-10-30 (activity-core rhythm)

briefs/2026/10/2026-10-30.md

@@ -0,0 +1,53 @@
+---
+date: 2026-10-30
+timezone: Europe/Berlin
+author: rein-aharness
+status: final
+sources_checked:
+  - docs/sources-allowlist.md
+  - automated field scan via llm-connect
+---
+
+# Freedom Intelligence Daily Brief — 2026-10-30
+
+> Automated brief (activity-core schedule → rein-aharness). Deltas only.
+
+## Headline deltas
+
+- Google announces `Gemini-3.5-Pro-128k` API, a new flagship tier with a 128k context window, priced above its Flash offerings.
+- The new `Gemini-3.5-Pro-128k` API is positioned as a direct competitor to OpenAI's `GPT-5-Flash` and `GPT-5.x` mid-tiers, offering enhanced reasoning.
+- Hugging Face `Qwen` org releases `Qwen3.5-7B-Instruct`, an updated 7B model with improved instruction following and a more permissive license.
+- The `Qwen3.5-7B-Instruct` is a ~7B dense model, providing a new strong open-weight option for single-GPU local inference, potentially replacing `Qwen3-8B`.
+
+## Frontier & commercial (axis A)
+
+| Item | Delta | Sources | Lab relevance |
+| ---- | ----- | ------- | ------------- |
+| Google Gemini-3.5-Pro-128k API | New API release. Offers 128k context window. Pricing is $0.25/Mtok input and $0.50/Mtok output, positioning it as a mid-tier flagship offering with enhanced reasoning capabilities compared to Flash mo | https://ai.google.dev/pricing | Provides another competitive option for high-context, reasoning-heavy API routing. Compare performance and latency against `GPT-5-Flash` and `Claude Sonnet-class` for specific tasks, especially for co |
+
+## Edge / local / open (axis B)
+
+| Item | Delta | Sources | Lab relevance |
+| ---- | ----- | ------- | ------------- |
+| Qwen `Qwen3.5-7B-Instruct` | New open-weight release on Hugging Face. This is a ~7B dense model with improved instruction following. License is Apache-2.0, replacing the previous Qwen3-8B license terms. | https://huggingface.co/Qwen/Qwen3.5-7B-Instruct | A significant new open-weight model for single-GPU local inference. Evaluate for potential inclusion in the strategic reserve as a default local instruct model, potentially replacing `Qwen3-8B` due to |
+
+## Training & specialization (axis C)
+
+*(none)*
+
+## Harness & fleet (axis D)
+
+*(none)*
+
+## Collection candidates
+
+| id | org | name | priority | reason | approx size | license |
+| -- | --- | ---- | -------- | ------ | ----------- | ------- |
+| Qwen__Qwen3.5-7B-Instruct | Qwen | Qwen3.5-7B-Instruct | high | New, highly capable open-weight 7B model with Apache-2.0 license, suitable for strategic reserve as a default local instruct. Offers improved instruction following over `Qwen3-8B`. | ~7B dense (unknown — verify card for exact GiB) | Apache-2.0 |
+
+## Lab implications
+
+- Update `llm-connect` routing policy to include `Gemini-3.5-Pro-128k` for complex, high-context reasoning tasks, and perform A/B testing against `GPT-5-Flash` and `Claude Sonnet-class` to determine optimal routing.
+- Initiate a collection request for `Qwen/Qwen3.5-7B-Instruct` to assess its performance and suitability as a strategic open-weight model for single-GPU inference, updating `inventory/catalog/` and potentially deprecating `Qwen3-8B`.
+- Review the `Qwen3.5-7B-Instruct` model card for exact parameter counts and recommended quantization strategies to inform collection and deployment planning, especially regarding its memory footprint compared to `Qwen3-8B`.
+- Continue to monitor for further updates from Google regarding their Pro API tiers, especially concerning latency and specific use-case performance in agentic workflows.