tegwick 70cbfe0c6c Catch up field brief, catalog Kimi K3, and wire live daily recurrence.

Add 2026-08-03 week catch-up brief with Kimi K3 as S1 strategic (collection
deferred on ~1454 GiB vs 850 GiB soft quota), document recurrence ops, and add a
workstation due-notice consumer cron helper. Production schedule was applied to
railiance ConfigMap and Temporal (weekdays 07:30 Europe/Berlin).

2026-08-03 17:37:53 +02:00

3.8 KiB

Raw Blame History

date

timezone

author

status

sources_checked

2026-08-03

Europe/Berlin

grok

final

research/2026-07-24-baseline-field-survey.md

briefs/2026/07/2026-07-28.md

public trackers (llm-stats, Artificial Analysis notes, HF moonshotai)

huggingface.co/moonshotai/Kimi-K3

Freedom Intelligence Daily Brief — 2026-08-03

Week catch-up from 2026-07-28 (recurring automation did not fire).
Delta overview, not a full resurvey.

Headline deltas

Kimi K3 open weights (Moonshot, ~Jul 27): 2.8T MoE, ~1.45 TiB official tree, custom Kimi K3 License — currently #1 open-weight on several boards; primary strategic S1 candidate (storage exceeds 850 GiB soft quota as full tree).
DeepSeek V4-Flash-0731 (Jul 31): same cheap price/size class, large capability jump — strong open price-performance; revisit as mid-size S/W vs our R spine.
Closed frontier churn: Claude Fable 5 back after export pause; Claude Opus 5 / GPT-5.6 family / Grok 4.5 / Gemini Flash tiers all active in trackers.
Lab: FI daily brief still due since Jul 28 (only two prior events); R spine remains on VAULT; recurrence wiring resumed this session.
Reserve action: nominate moonshotai/Kimi-K3 as S1 strategic with collection deferred pending capacity decision (full pull ≈1.45 TiB).

Frontier & commercial (axis A)

Item	Delta	Sources	Lab relevance
Claude Fable 5 / Opus 5	Fable restored; Opus competitive on coding/agentic	trackers / AA notes	Closed bar for harness evals via llm-connect
GPT-5.6 Sol/Terra/Luna	GA; Luna deep discount end-Jul	OpenAI pricing notes	Budget routing tiers
Grok 4.5	Public coding-focused tier	xAI product notes	Compare agent coding bar
Gemini 3.5/3.6 Flash class	Fast tiers shipping	Google product notes	Cheap closed automation

Edge / local / open (axis B)

Item	Delta	Sources	Lab relevance
Kimi K3	Best open general/agentic class; 2.8T / 1M ctx / vision	HF moonshotai/Kimi-K3	S1 identity; too large for 850 GiB soft quota as full tree
DeepSeek V4-Flash-0731	Open MIT-class value leader after Jul 31 retrain	DeepSeek release trackers	Candidate later mid-tier collect
GLM-5.2	Still “hostable strong open” for many teams	trackers	Alternative S dense if we skip full K3
Qwen 3.7/3.8-class	Continued family updates	Alibaba trackers	Watch vs Qwen3-8B R default

Training & specialization (axis C)

(none material for lab this window — no new FT stack displacement vs Unsloth/axolotl baseline)

Harness & fleet (axis D)

Item	Delta	Sources	Lab relevance
FI daily brief gap	No automated brief since 2026-07-28	State Hub progress	Wire recurrence + consumer this session
Agent swarm narrative (Kimi)	Product emphasis on multi-agent / long-horizon	Kimi product	Align axis D watchlist with K3 agent demos

Collection candidates

id	org	name	priority	reason	approx size	license
moonshotai__Kimi-K3__strategic	moonshotai	Kimi-K3	high	S1 best open general/agentic; displace V3 as primary strategic identity	~1454 GiB official	custom (Kimi K3 License) — review commercial thresholds

Decision: catalog + defer download until soft-quota / disk exception. Prefer official compressed if/when published under soft quota.

Lab implications

Keep R spine (Qwen3-8B, R1-Distill-14B, embeds) as daily local defaults.
Update strategic plan: S1 = Kimi K3 (identity), not only DeepSeek-V3.
Unblock weekday schedule on activity-core (ConfigMap apply + sync + schedule check).
Llama-3.2-3B still blocked on HF token (unchanged).

3.8 KiB Raw Blame History