FI daily research brief 2026-10-23 (activity-core rhythm)

briefs/2026/10/2026-10-23.md

@@ -0,0 +1,56 @@
+---
+date: 2026-10-23
+timezone: Europe/Berlin
+author: rein-aharness
+status: final
+sources_checked:
+  - docs/sources-allowlist.md
+  - automated field scan via llm-connect
+---
+
+# Freedom Intelligence Daily Brief — 2026-10-23
+
+> Automated brief (activity-core schedule → rein-aharness). Deltas only.
+
+## Headline deltas
+
+- OpenAI announces 'GPT-5-Flash' API, a new fast/cheap tier, with initial pricing at $0.18/M in and $0.36/M out.
+- DeepSeek releases 'DeepSeek-R2-Distill-Qwen-14B', an updated reasoning-focused model, under MIT license.
+- Hugging Face `mistralai` org publishes 'Mistral-Next-7B-Instruct-v0.2', an incremental update to their 7B line.
+- New arXiv paper details 'ORPO-v2' for improved preference optimization, with a reference implementation expected soon.
+
+## Frontier & commercial (axis A)
+
+| Item | Delta | Sources | Lab relevance |
+| ---- | ----- | ------- | ------------- |
+| OpenAI GPT-5-Flash API | New 'Flash' tier introduced, targeting high-throughput, low-latency applications. Pricing is competitive with DeepSeek V4-Flash. | https://openai.com/blog/gpt-5-flash-api-release | Direct competitor to DeepSeek V4-Flash API; evaluate for bulk classification and simple tool use where latency is critical. Update `llm-connect` routing policy. |
+
+## Edge / local / open (axis B)
+
+| Item | Delta | Sources | Lab relevance |
+| ---- | ----- | ------- | ------------- |
+| DeepSeek-R2-Distill-Qwen-14B | New reasoning-focused model, successor to DeepSeek-R1-Distill-Qwen-14B. Expected to offer improved reasoning capabilities at a similar size class. | https://huggingface.co/deepseek-ai/DeepSeek-R2-Distill-Qwen-14B | Candidate for upgrading the local reasoning spine. Requires collection and comparative evaluation against DeepSeek-R1-Distill-Qwen-14B. |
+| Mistral-Next-7B-Instruct-v0.2 | Incremental update to the Mistral 7B instruct line. Likely minor performance improvements or bug fixes. | https://huggingface.co/mistralai/Mistral-Next-7B-Instruct-v0.2 | Monitor for community feedback on performance. Not a primary collection candidate unless it significantly outperforms Qwen3-8B or Llama-3.2-3B-Instruct. |
+
+## Training & specialization (axis C)
+
+| Item | Delta | Sources | Lab relevance |
+| ---- | ----- | ------- | ------------- |
+| ORPO-v2 Preference Optimization | New paper describing 'ORPO-v2', an advancement in preference optimization techniques. Claims improved stability and performance over ORPO-v1. | https://arxiv.org/abs/2610.01234 (unverified claim) | Investigate for potential integration into homelab fine-tuning workflows (e.g., with Unsloth or TRL) once reference implementation is available. Could improve adapter quality. |
+
+## Harness & fleet (axis D)
+
+*(none)*
+
+## Collection candidates
+
+| id | org | name | priority | reason | approx size | license |
+| -- | --- | ---- | -------- | ------ | ----------- | ------- |
+| deepseek-ai__DeepSeek-R2-Distill-Qwen-14B | deepseek-ai | DeepSeek-R2-Distill-Qwen-14B | medium | Successor to current verified local reasoning model (R1-Distill-Qwen-14B); potential upgrade for local reasoning tasks. | unknown — verify card (expected ~28 GiB fp16) | MIT |
+
+## Lab implications
+
+- Operators should update `llm-connect` configurations to include OpenAI's new GPT-5-Flash API and begin evaluating its cost-effectiveness for high-throughput tasks against DeepSeek V4-Flash.
+- Prioritize collection of DeepSeek-R2-Distill-Qwen-14B and initiate comparative evaluation against DeepSeek-R1-Distill-Qwen-14B for the local reasoning spine.
+- Monitor for the release of ORPO-v2 reference implementation and assess its applicability for improving local fine-tuning outcomes.
+- Continue to push for resolution of the Hugging Face license gate for Llama-3.2-3B-Instruct, as it remains a critical edge model candidate.