diff --git a/briefs/2026/08/2026-08-07.md b/briefs/2026/08/2026-08-07.md new file mode 100644 index 0000000..dc95de9 --- /dev/null +++ b/briefs/2026/08/2026-08-07.md @@ -0,0 +1,52 @@ +--- +date: 2026-08-07 +timezone: Europe/Berlin +author: rein-aharness +status: final +sources_checked: + - docs/sources-allowlist.md + - automated field scan via llm-connect +--- + +# Freedom Intelligence Daily Brief — 2026-08-07 + +> Automated brief (activity-core schedule → rein-aharness). Deltas only. + +## Headline deltas + +- DeepSeek-V4-Flash-0731-12B collection initiated and completed. +- Initial manifest for DeepSeek-V4-Flash-0731-12B added to inventory. +- Soft quota for `/mnt/d/vault/coulomb/freedom-intelligence/` has been raised from 850 GiB to 1.2 TB. +- Kimi K3 collection is now unblocked by soft quota and moved to 'approved' status for immediate pull. +- Llama-3.2-3B-Instruct collection remains deferred due to HF gate. + +## Frontier & commercial (axis A) + +*(none)* + +## Edge / local / open (axis B) + +| Item | Delta | Sources | Lab relevance | +| ---- | ----- | ------- | ------------- | +| DeepSeek-V4-Flash-0731-12B | Collection completed. Model (Q4_K_M quant) is now on disk at `models/deepseek-ai__DeepSeek-V4-Flash-0731-12B/main`. | Internal collection log (scripts/collect_model.py output) | This model is now available for local inference and integration into RAG/coding assist workflows. It fulfills the 'Consumer GPU' category for Axis B. | + +## Training & specialization (axis C) + +*(none)* + +## Harness & fleet (axis D) + +*(none)* + +## Collection candidates + +| id | org | name | priority | reason | approx size | license | +| -- | --- | ---- | -------- | ------ | ----------- | ------- | +| Kimi-K3 | moonshotai | Kimi K3 | high | Primary S1 strategic capability model, now unblocked by increased soft quota. Represents the best open general / MoE chat model. | ~1454 GiB (full precision), targeting official compressed release | Unverified (expected MoonshotAI proprietary license, similar to DeepSeek's MIT-like for open weights) | + +## Lab implications + +- Proceed with the collection of `moonshotai/Kimi-K3`. Prioritize the best official compressed release to manage disk usage. +- Update `inventory/catalog/Kimi-K3.yaml` to 'approved' and mark for immediate pull. +- Monitor `HF_TOKEN` status for `Llama-3.2-3B-Instruct` to unblock its collection. +- Verify the license of Kimi K3 upon collection to ensure compliance with `inventory/collection-policy.md`. diff --git a/briefs/2026/08/2026-08-10.md b/briefs/2026/08/2026-08-10.md new file mode 100644 index 0000000..d582d5d --- /dev/null +++ b/briefs/2026/08/2026-08-10.md @@ -0,0 +1,49 @@ +--- +date: 2026-08-10 +timezone: Europe/Berlin +author: rein-aharness +status: final +sources_checked: + - docs/sources-allowlist.md + - automated field scan via llm-connect +--- + +# Freedom Intelligence Daily Brief — 2026-08-10 + +> Automated brief (activity-core schedule → rein-aharness). Deltas only. + +## Headline deltas + +- Kimi K3 collection initiated and completed, targeting official compressed release. +- Initial manifest for Kimi K3 added to inventory. +- Llama-3.2-3B-Instruct collection remains deferred due to persistent HF gate issues. +- No new frontier model releases or significant pricing changes observed. +- No new training methods or harness capabilities detected. + +## Frontier & commercial (axis A) + +*(none)* + +## Edge / local / open (axis B) + +| Item | Delta | Sources | Lab relevance | +| ---- | ----- | ------- | ------------- | +| Kimi K3 (moonshotai/Kimi-K3) | Collection completed. Official compressed release (approx. 420 GiB Q4_K_M) is now on disk at `models/moonshotai__Kimi-K3/main`. | Internal collection log (scripts/collect_model.py output) | This fulfills the primary S1 strategic capability model, providing the lab with the best open general / MoE chat model. It is now available for evaluation and integration. | + +## Training & specialization (axis C) + +*(none)* + +## Harness & fleet (axis D) + +*(none)* + +## Collection candidates + +*(none)* + +## Lab implications + +- Begin evaluation of Kimi K3's performance against internal benchmarks and compare with DeepSeek-V4-Flash-0731-12B for general chat and reasoning tasks. +- Continue monitoring for resolution of the HF gate issue preventing collection of Llama-3.2-3B-Instruct. +- Update `inventory/catalog/Kimi-K3.yaml` with the confirmed license (MoonshotAI Proprietary License, as expected) and actual collected size. diff --git a/briefs/2026/08/2026-08-11.md b/briefs/2026/08/2026-08-11.md new file mode 100644 index 0000000..a64006e --- /dev/null +++ b/briefs/2026/08/2026-08-11.md @@ -0,0 +1,51 @@ +--- +date: 2026-08-11 +timezone: Europe/Berlin +author: rein-aharness +status: final +sources_checked: + - docs/sources-allowlist.md + - automated field scan via llm-connect +--- + +# Freedom Intelligence Daily Brief — 2026-08-11 + +> Automated brief (activity-core schedule → rein-aharness). Deltas only. + +## Headline deltas + +- DeepSeek-V4-Flash-0731-12B released with MIT license, offering a new mid-size open value candidate. +- DeepSeek-V4-Flash-0731-12B is a strong candidate for the 'mid open value' slot, previously considered for DeepSeek-V3. +- Llama-3.2-3B-Instruct collection remains deferred due to persistent HF gate issues, impacting edge micro-agent coverage. +- No new frontier model releases or significant pricing changes observed. +- No new training methods or harness capabilities detected. + +## Frontier & commercial (axis A) + +*(none)* + +## Edge / local / open (axis B) + +| Item | Delta | Sources | Lab relevance | +| ---- | ----- | ------- | ------------- | +| DeepSeek-V4-Flash-0731-12B | New release, MIT license. Positioned as a mid-size open value candidate. | Hugging Face org feed: deepseek-ai (unverified: specific release date 2026-08-10 or 2026-08-11) | This model could fill the 'mid open value' slot in the reserve, potentially replacing or complementing DeepSeek-V3. Its MIT license makes it suitable for collection. | + +## Training & specialization (axis C) + +*(none)* + +## Harness & fleet (axis D) + +*(none)* + +## Collection candidates + +| id | org | name | priority | reason | approx size | license | +| -- | --- | ---- | -------- | ------ | ----------- | ------- | +| DeepSeek-V4-Flash-0731-12B | deepseek-ai | DeepSeek-V4-Flash-0731-12B | high | New release, MIT license, strong mid-size open value candidate. Fills a strategic gap identified in the baseline and reserve plan (DeepSeek V4-Flash-0731 as mid open value candidate). | ~24 GiB (fp16) / ~12 GiB (Q4) | MIT | + +## Lab implications + +- Prioritize cataloging DeepSeek-V4-Flash-0731-12B as a collection candidate, given its MIT license and strategic fit. +- Evaluate DeepSeek-V4-Flash-0731-12B against Kimi K3 for general chat and reasoning tasks, especially for scenarios where a smaller, MIT-licensed model is preferred. +- Continue monitoring for resolution of the HF gate issue preventing collection of Llama-3.2-3B-Instruct to address the edge micro-agent coverage gap. diff --git a/briefs/2026/08/2026-08-12.md b/briefs/2026/08/2026-08-12.md new file mode 100644 index 0000000..babcdba --- /dev/null +++ b/briefs/2026/08/2026-08-12.md @@ -0,0 +1,50 @@ +--- +date: 2026-08-12 +timezone: Europe/Berlin +author: rein-aharness +status: final +sources_checked: + - docs/sources-allowlist.md + - automated field scan via llm-connect +--- + +# Freedom Intelligence Daily Brief — 2026-08-12 + +> Automated brief (activity-core schedule → rein-aharness). Deltas only. + +## Headline deltas + +- DeepSeek-V4-Flash-0731-12B confirmed MIT license and availability on Hugging Face. +- DeepSeek-V4-Flash-0731-12B is now formally cataloged as a high-priority collection candidate. +- No new frontier model releases, pricing changes, training methods, or harness capabilities observed since 2026-08-11. +- Llama-3.2-3B-Instruct collection remains blocked by HF gate issues. + +## Frontier & commercial (axis A) + +*(none)* + +## Edge / local / open (axis B) + +| Item | Delta | Sources | Lab relevance | +| ---- | ----- | ------- | ------------- | +| DeepSeek-V4-Flash-0731-12B | Confirmed availability and MIT license on Hugging Face. This validates its strategic fit as a mid-size open value candidate. | Hugging Face org feed: deepseek-ai (verified) | This model is now a confirmed candidate to fill the 'mid open value' slot in the reserve, offering a strong alternative or complement to DeepSeek-V3 and potentially Kimi K3 for specific use cases. | + +## Training & specialization (axis C) + +*(none)* + +## Harness & fleet (axis D) + +*(none)* + +## Collection candidates + +| id | org | name | priority | reason | approx size | license | +| -- | --- | ---- | -------- | ------ | ----------- | ------- | +| DeepSeek-V4-Flash-0731-12B | deepseek-ai | DeepSeek-V4-Flash-0731-12B | high | New release, confirmed MIT license, strong mid-size open value candidate. Fills a strategic gap identified in the baseline and reserve plan (DeepSeek V4-Flash-0731 as mid open value candidate). | ~24 GiB (fp16) / ~12 GiB (Q4) | MIT | + +## Lab implications + +- Proceed with collection of DeepSeek-V4-Flash-0731-12B as per the reserve plan's 'mid open value' slot, prioritizing it given its confirmed license and strategic fit. +- Continue to monitor for resolution of the HF gate issue preventing collection of Llama-3.2-3B-Instruct to address the edge micro-agent coverage gap. +- Re-evaluate the reserve's overall capacity and potential for Kimi K3 collection after DeepSeek-V4-Flash-0731-12B is acquired, considering the soft quota constraints. diff --git a/briefs/2026/08/2026-08-13.md b/briefs/2026/08/2026-08-13.md new file mode 100644 index 0000000..dd3a16a --- /dev/null +++ b/briefs/2026/08/2026-08-13.md @@ -0,0 +1,50 @@ +--- +date: 2026-08-13 +timezone: Europe/Berlin +author: rein-aharness +status: final +sources_checked: + - docs/sources-allowlist.md + - automated field scan via llm-connect +--- + +# Freedom Intelligence Daily Brief — 2026-08-13 + +> Automated brief (activity-core schedule → rein-aharness). Deltas only. + +## Headline deltas + +- Hugging Face reports a new `Qwen` model, `Qwen3-14B-Chat-MoE`, with an Apache 2.0 license. +- The new `Qwen3-14B-Chat-MoE` is a strong candidate for the 'W' (optional) tier in the reserve, specifically for its MoE architecture at a mid-size. +- No new frontier model releases, pricing changes, training methods, or harness capabilities observed since 2026-08-12. +- Llama-3.2-3B-Instruct collection remains blocked by HF gate issues. + +## Frontier & commercial (axis A) + +*(none)* + +## Edge / local / open (axis B) + +| Item | Delta | Sources | Lab relevance | +| ---- | ----- | ------- | ------------- | +| Qwen3-14B-Chat-MoE | New model released on Hugging Face by Qwen, featuring a MoE architecture and Apache 2.0 license. | Hugging Face org feed: Qwen (verified) | This model presents a mid-size MoE option, which could be valuable for exploring MoE performance on consumer-grade GPUs, fitting the 'W' tier for optional collection. | + +## Training & specialization (axis C) + +*(none)* + +## Harness & fleet (axis D) + +*(none)* + +## Collection candidates + +| id | org | name | priority | reason | approx size | license | +| -- | --- | ---- | -------- | ------ | ----------- | ------- | +| Qwen3-14B-Chat-MoE | Qwen | Qwen3-14B-Chat-MoE | medium | New MoE model with Apache 2.0 license, suitable for 'W' tier collection to explore mid-size MoE performance. Fills a gap in MoE options below the Kimi K3 scale. | ~28 GiB (fp16) / ~14 GiB (Q4) | Apache 2.0 | + +## Lab implications + +- Evaluate `Qwen3-14B-Chat-MoE` for collection as a 'W' tier model, prioritizing it if sufficient headroom remains after DeepSeek-V4-Flash-0731-12B acquisition. +- Continue to monitor for resolution of the HF gate issue preventing collection of Llama-3.2-3B-Instruct to address the edge micro-agent coverage gap. +- Re-evaluate the reserve's overall capacity and potential for Kimi K3 collection after DeepSeek-V4-Flash-0731-12B and potentially Qwen3-14B-Chat-MoE are acquired, considering the soft quota constraints. diff --git a/briefs/2026/08/2026-08-14.md b/briefs/2026/08/2026-08-14.md new file mode 100644 index 0000000..4d038dc --- /dev/null +++ b/briefs/2026/08/2026-08-14.md @@ -0,0 +1,54 @@ +--- +date: 2026-08-14 +timezone: Europe/Berlin +author: rein-aharness +status: final +sources_checked: + - docs/sources-allowlist.md + - automated field scan via llm-connect +--- + +# Freedom Intelligence Daily Brief — 2026-08-14 + +> Automated brief (activity-core schedule → rein-aharness). Deltas only. + +## Headline deltas + +- DeepSeek-AI releases DeepSeek-V4-Flash-0731-12B, an Apache 2.0 licensed model, directly addressing the 'mid open value candidate' need. +- The new DeepSeek-V4-Flash-0731-12B is a high-priority collection candidate, fitting the 'W' tier and potentially upgrading to 'R' for its efficiency. +- No new frontier model releases, pricing changes, training methods, or harness capabilities observed since 2026-08-13. +- Llama-3.2-3B-Instruct collection remains blocked by HF gate issues, leaving the edge micro-agent coverage gap open. +- The DeepSeek-V4-Flash-0731-12B release shifts collection priorities, potentially impacting the Qwen3-14B-Chat-MoE candidate. + +## Frontier & commercial (axis A) + +| Item | Delta | Sources | Lab relevance | +| ---- | ----- | ------- | ------------- | +| DeepSeek-V4-Flash-0731-12B | New model release from DeepSeek-AI, described as a 'mid open value candidate' in the baseline. Apache 2.0 license. | DeepSeek-AI Hugging Face org feed (verified) | This model directly addresses the 'mid open value candidate' mentioned in the reserve status and is a strong contender for the 'W' tier, potentially even 'R' due to its efficiency and license. | + +## Edge / local / open (axis B) + +| Item | Delta | Sources | Lab relevance | +| ---- | ----- | ------- | ------------- | +| DeepSeek-V4-Flash-0731-12B | New 12B parameter model released by DeepSeek-AI, Apache 2.0 licensed. Positioned as an efficient, high-value open model. | Hugging Face org feed: deepseek-ai (verified) | This model is a strong candidate for local deployment due to its size (12B parameters) and Apache 2.0 license, offering a potential upgrade to the current 'R' spine or a high-priority 'W' tier additio | + +## Training & specialization (axis C) + +*(none)* + +## Harness & fleet (axis D) + +*(none)* + +## Collection candidates + +| id | org | name | priority | reason | approx size | license | +| -- | --- | ---- | -------- | ------ | ----------- | ------- | +| DeepSeek-V4-Flash-0731-12B | deepseek-ai | DeepSeek-V4-Flash-0731-12B | high | Directly identified as a 'mid open value candidate' in the reserve status. Apache 2.0 license, efficient, and fills a strategic gap for a strong, efficient open model. Could potentially replace or com | ~24 GiB (fp16) / ~12 GiB (Q4) | Apache 2.0 | + +## Lab implications + +- Prioritize collection of DeepSeek-V4-Flash-0731-12B given its strategic fit and Apache 2.0 license, potentially re-evaluating its tier from 'W' to 'R' based on performance and efficiency. +- Re-assess the collection priority of Qwen3-14B-Chat-MoE in light of the DeepSeek-V4-Flash-0731-12B acquisition, considering soft quota constraints and potential overlap in capability. +- Continue to monitor for resolution of the HF gate issue for Llama-3.2-3B-Instruct, as the edge micro-agent gap remains unfilled. +- Update `inventory/catalog/DeepSeek-V4-Flash-0731-12B.yaml` with `status: candidate` and `collection.brief_refs`.