freedom-intelligence/briefs/2026/09/2026-09-13.md
tegwick a78187c33e Make the sensing loop durable and pin the reserve to Scaleway
The weekday brief only counts when the file is on origin/main. Hub
events without git objects are failures. The open-weight SoT moves
from the VAULT 850 GiB quota to a dedicated Scaleway bucket (One Zone
IA for R, Glacier for S) so V4-Flash and K3 are economically in-scope.

Catalogs the real DeepSeek-V4-Flash-0731 (MIT ~167 GiB MoE, not 12B)
and nominates Qwen3.8-27B. Bucket create remains FI-WP-0004-T08.

Assistant: grok
Assistant-Session: 01a09c6a-1cfc-75b1-a78d-c13eaf22241d
2026-09-13 22:45:43 +02:00

81 lines
5.7 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

---
date: 2026-09-13
timezone: Europe/Berlin
author: grok
status: final
sources_checked:
- docs/sources-allowlist.md
- inventory/RESERVE-STATUS.md
- huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731
- huggingface.co/Qwen/Qwen3.8-27B (card identity via public trackers)
- https://www.scaleway.com/en/pricing/storage/
---
# Freedom Intelligence Daily Brief — 2026-09-13
> **Month catch-up.** Last file on `origin/main` is 2026-08-14. Hub
> `fi_daily_brief` events exist for 2026-09-07…11 with `committed=True`
> and no git objects. Those days are **not** recovered here. This file
> is a field + ops delta from 2026-08-14, not a resurrection of the
> missing executor commits.
## Headline deltas
- **Loop failure:** sensing ran locally on railiance and cleared `due` without publishing to origin. Rein-aharness dropped `git push` on 2026-09-04 as an integrity guard; FI now re-grants origin publication as a named completion requirement (FI-WP-0004).
- **DeepSeek-V4-Flash-0731 is not 12B.** Official card: MIT, 284B base / ~13B active (304B with DSpark), ~167 GiB HF tree, 2026-07-31. Automated briefs 08-04…14 re-announced it as a new 12B dense model (license flipped MIT/Apache) and 08-07 claimed it was already on disk. Catalog now records the real model as **S, approved, not collected**.
- **Qwen3.8-27B** (Apache-2.0 dense multimodal, 2026-08-14) and **Qwen3.8-2.4T-A95B** (custom Max licence, multi-TB, ~08-12) shipped in the hole. 27B is a W/R candidate; Max is S-watch only until euro budget is live.
- **Reserve facility:** VAULT HD 850 GiB quota is retired as SoT. Scaleway `nl-ams` object storage is the reserve (One Zone IA for R, Glacier for S). Kimi K3 (~1454 GiB) becomes a few euros/month on Glacier instead of a quota wall.
- **Next pull** after the bucket exists: V4-Flash official tree into `strategic/` Glacier — it always fitted even the old 850 GiB quota.
## Frontier & commercial (axis A)
| Item | Delta | Sources | Lab relevance |
| ---- | ----- | ------- | ------------- |
| DeepSeek-V4-Flash-0731 | Official open MIT MoE; API `deepseek-v4-flash`. Card scores Terminal Bench 2.1 82.7. | [HF card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731), [arXiv 2606.19348](https://arxiv.org/abs/2606.19348) | Closed-bar comparison via llm-connect; open identity is the S pull |
| Qwen3.8-Max family | API GA early Aug; open 2.4T-A95B under **bespoke** licence (revenue threshold), not Apache | Qwen card / trackers (verify licence text at pin) | Do not collect the 2.4T tree until T08 budget + licence read |
| GLM-5.3-Flash (late Aug) | 320B-A18B, MIT, 1M ctx, reported $0.15 / $0.50 per MTok | Z.ai release notes (verify) | Watch vs V4-Flash as MIT S alternative; not displacing V4-Flash as first pull |
## Edge / local / open (axis B)
| Item | Delta | Sources | Lab relevance |
| ---- | ----- | ------- | ------------- |
| DeepSeek-V4-Flash-0731 size class | ~167 GiB official tree, ~13B active — **S, not R** | HF file listing ~167 GB; card “304B params” | First Scaleway Glacier collect |
| Qwen3.8-27B | Dense multimodal Apache-2.0, ~27B, 2026-08-14 — same calendar day as last origin brief | [HF Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) | Candidate R-spine upgrade; cataloged `candidate` |
| Muse Glimmer 30B (Meta, ~08-10) | Apache-2.0 dense multimodal, ~2432 GB class | Meta/HF trackers (verify card) | Watch; do not pull until V4-Flash and 27B decision |
| Kimi K3 | Unchanged identity; capacity objection is now euros not GiB | existing catalog | Glacier ~€3.69/mo; still deferred until V4-Flash lands |
## Training & specialization (axis C)
| Item | Delta | Sources | Lab relevance |
| ---- | ----- | ------- | ------------- |
| Qwen3.8-27B as FT base | Dense 27B Apache-2.0 is a more serious QLoRA target than 8B if T2T3 VRAM exists | Qwen3.8-27B card | Hardware envelope hosts still TBD — do not pretend we can FT it today |
| No new homelab recipe displacement | Unsloth / TRL / Axolotl still the working set from baseline | allowlist | No stack change |
## Harness & fleet (axis D)
| Item | Delta | Sources | Lab relevance |
| ---- | ----- | ------- | ------------- |
| FI executor publication | 2026-09-04 rein-aharness removed `git push`; hub still cleared `due` | `rein-aharness` REINAH-WP-0003, hub events 2026-09-07…11 | **Broken durable memory.** FI-WP-0004 re-grants origin push and verifier |
| V4-Flash agentic stack | Official vLLM DSpark + SGLang recipes on the card | HF card “How to Run” | Axis D watch; serve is not this repo |
## Collection candidates
| id | org | name | priority | reason | approx size | license |
| -- | --- | ---- | -------- | ------ | ----------- | ------- |
| deepseek-ai__DeepSeek-V4-Flash-0731__strategic | deepseek-ai | DeepSeek-V4-Flash-0731 | high | Corrected S1; MIT; ~167 GiB; first Scaleway pull | ~167 GiB | MIT |
| Qwen__Qwen3.8-27B__candidate | Qwen | Qwen3.8-27B | medium | Apache-2.0 dense 27B R-upgrade | ~56 GiB class | Apache-2.0 |
Already in catalog, **not** new: Kimi K3, DeepSeek-V3, Qwen3-8B, R1-Distill-14B, BGE-M3, nomic-embed.
## Lab implications
- Finish FI-WP-0004: origin publish in the executor, durability verifier in CI/ops, Scaleway bucket (T08), then collect V4-Flash (T09).
- Do not treat hub `fi_daily_brief` as evidence the brief exists until `scripts/verify_brief_durability.py` is green.
- Recovering 2026-09-07…11 files from the railiance clone is optional; they were produced by the same failing loop and are not trusted.
- Hardware envelope hosts are still TBD — R-upgrade of Qwen3.8-27B waits on that as much as on disk.
## Notes / caveats
- Parameter counts and €/GB from primary cards/pricing pages today; re-measure at pull.
- Qwen3.8-Max open tree licence is **not** Apache-2.0; do not copy the 27B licence onto it.
- `DeepSeek-V4-Flash-0731-12B` is not a real repo id.