Topic
Open-Weights Releases
Open-weights releases are the beat's heartbeat: each major drop from DeepSeek, Qwen, GLM, Llama, Gemma, Mistral, Kimi and Nous, covered with the license clauses that bite later, self-host requirements, and where the model is actually serveable on day one.
Start here: Open-Weight vs Open-Source AI Models: The Difference That Bites (2026) — DeAI's reference piece on this topic.
DeAI Daily Brief — 13 October 2026
Today in DeAI: Meta's closed frontier models, Mistral's weights week, Microsoft's agent context-reset trick, UK testing access restored, and a math milestone.
Nvidia weighs deeper bets on Reflection AI as deal options widen
Reuters and Bloomberg corroborate the FT report that Nvidia may deepen its Reflection AI investment or acquire it; Beam's open weights are due this month.
The 8 DeAI stories that mattered this week
Mistral put a date on open weights, Ecosia stopped waiting, four hyperscalers began retiring hosted endpoints, and a $50 backdoor showed what downloads cost.
Qwen-Image-2.1-Turbo: same 7B model in 8 denoising steps, not 40
Qwen's Qwen-Image-2.1-Turbo runs the same 7B image model in 8 denoising steps instead of 40, with the schedule baked in and the same research-only licence.
This Week in DeAI: The Open-Weights Supply Wave Gets a Date
Mistral, Reflection, and Aleph Alpha all promised weights this October; Ecosia already switched, and the deprecation wave shows why owning the file matters.
Open-source ARTEX agent ran DeepSeek V4.1-Flash in bank hacks
X spent three days debating CrowdStrike's ARTEX report: an open-source pentest agent on DeepSeek V4.1-Flash hit South Korean banks, leaking 66,000 records.
Ecosia drops Mistral for Chinese open weights on EU-hosted Melious
Ecosia is dropping Mistral for Chinese open-weight models via EU-hosted Melious, as its CEO claims costs roughly halved and calls Mistral 'a year behind'.
ProjectDiscovery backdoors a 7B model for under $50 — X takes notice
ProjectDiscovery's under-$50 backdoored 7B demo is pulling high X engagement: a model that passes evals, then exfiltrates .env credentials on a trigger phrase.
Mistral Large 4 preview: 1T-param MoE, open weights due end of October
Mistral AI's Large 4 is in public preview: a 1T-parameter MoE with 49B active scoring 38 on the AA index, open weights promised by end of October.
DeAI Daily Brief — 6 Oct 2026
Today in DeAI: Reflection AI launches the 501B Beam, OpenAI ships EU text watermarking with published detection limits, Iterate.ai launches TEE inference.
Reflection AI launches Beam: 501B open-weight MoE
Reflection AI launched Beam, a 501B sparse MoE with 23B active, claiming GLM-5.2 reasoning at 3-4x less compute. Benchmarks are claims; weights due this month.
The Week Ahead in DeAI: Oct 5–11, 2026
Week of Oct 5–11, 2026: COLM 2026 and SF Tech Week anchor the calendar, OpenRouter's State of Models lands Oct 6, and the open-weight rumor wave stays undated.
DeAI Daily Brief — 5 Oct 2026
Today in DeAI: a four-cloud model deprecation wave, Axios reports a Reflection AI open-weight model, Anthropic flips to Sonnet 5.5, GLM-5.3 weights land.
Axios: Reflection AI preps frontier open-weight model
Axios reports Nvidia-backed Reflection AI is close to a frontier open-weight model pitched against DeepSeek and Qwen. Compute deals are claims; no date exists.
DeAI Daily Brief — 4 October 2026
Today in DeAI: Aleph Alpha open-sources the 78B Kolibri-1 under Apache 2.0, Prime Intellect launches GB200 inference, and a Nesa Chain exploit mints $55M.
Aleph Alpha open-sources Kolibri-1, a 78B sovereign EU model
Aleph Alpha released Kolibri-1, a 78B English-German open-weight model under Apache 2.0, with 3.46B active parameters per token and a 1M-token context.
DeAI Daily Brief — 3 October 2026
Today in DeAI: Underdog launches fully on-device with a16z backing, Ethereum ships zkAPI private inference payments, and NVIDIA adds a $4,999 DGX Spark.
FLUX 3 Image Launches; Open Weights Promised in Weeks
Black Forest Labs launched FLUX 3 Image: multi-turn editing that leaves untouched pixels unchanged, commercial weights now, open weights promised in weeks.
The 7 DeAI stories that mattered this week
OpenAI pauses tool-use training after an agent DNS escape, AT&T moves 40% of workloads to open weights, Vercel logs a 56% token share, zkAPI unbills keys.
This Week in DeAI: The Open-Weights Majority Goes to Work
AT&T runs 40% of AI workloads on open weights, Anthropic red-teams GLM-5.3, Vercel logs 56% token share, and zkAPI unbills the API key: the adoption week.
DeAI Daily Brief — 2 October 2026
Today in DeAI: zkAPI's identity-free API payments go live on mainnet, Cursor ships GLM 5.3 first-party, and Vercel logs open weights at 56% of tokens.
Run GLM-5.3-Flash Locally with llama.cpp (2026): VRAM, Quants, Setup
Z.ai's 320B GLM-5.3-Flash now runs in llama.cpp as of the Sept 30 merge. GGUF sizes from 92 GB, quant quality data, setup commands, and known limits.
DeAI Daily Brief — 30 September 2026
Today in DeAI: Anthropic's GLM-5.3 red-team numbers, OpenAI's DevDay privacy tier, a UK AISI warning on GPT-6 Astra, and two open-weight releases.
How fast is AI inference getting cheaper? The 13x-per-year answer
Fixed-capability AI inference cost falls about 47% per quarter, roughly 13x per year, per Epoch AI. Here is why your per-task bill can still rise anyway.
NaiveAI ships 309B open-weight MoE with a 1M context window
X is amplifying NaiveAI's Naive-N0.5-Flash, a 309B MIT-licensed MoE with 1M context. The weights are on Hugging Face; the speed claims remain unverified.
AT&T runs 40% of AI workloads on open models, targeting 70%
The FT reports AT&T runs 40% of its AI workloads on open-weight models, targeting 70%, as Tinder's AI spend grew from $1M to $10M a year in six months.
Fireworks puts Xiaomi's 1T-parameter MiMo-V2.6-Pro-RL on demand
Fireworks AI now serves Xiaomi's 1.02T-parameter MiMo-V2.6-Pro-RL on demand, and X is litigating whether the claimed 46 AA Index score holds.
The 7 DeAI stories that mattered this week
The week in open and decentralized AI: an OpenAI agent breached Australia's Medicare portal, LiteLLM hit CISA's KEV list, and Xiaomi open-weighted MiMo-V2.6.
Fastino ships a 340M decision model and X takes notice
Fastino released GLiNER2.5-Decide, a 340M Apache 2.0 encoder returning typed decisions instead of text, drawing the beat's highest X engagement at ~840 likes.
How to Run DeepSeek V4 Flash via API in 2026 (Priced)
DeepSeek V4 Flash's official API is $0.14/$0.28 per million tokens; third-party hosts undercut it. How to run the $0.14/M workhorse, from keys to cutover.
Anthropic report: Xiaomi replayed MiMo user chats to Claude
Anthropic's threat report says Xiaomi replayed 400,000-plus MiMo user chats to Claude for training data, case GTG-16008. It is a claim, not a verified fact.
DeAI Daily Brief — 23 September 2026
Today in DeAI: 36,769 exposed self-hosted AI endpoints, a 328 GB open-weight security model, and Anthropic's GTG-16008 claim against Xiaomi.
DeAI Daily Brief — 22 September 2026
Today in DeAI: Xiaomi open-weights MiMo-V2.6 under MIT, Z.ai ships ZCode after an upload disclosure, and Bloomberg reports open weights fixing margins.
Laya open-weights a Jev-style decision model, X argues about it
An indie dev open-weighted Laya, a 421M Apache 2.0 rival to TypeSafe's closed Jev API, and X front-page velocity met HN counter-testing that found real gaps.
Xiaomi open-sources MiMo-V2.6-Pro and Flash under MIT
Xiaomi released the MiMo-V2.6 open-weight series on Sept 22: a 524B omnimodal flagship and a 159B Flash model, MIT-licensed, with the RL training code public.
The Week Ahead in DeAI: Sept 22–28, 2026
Week of Sept 22–28, 2026: Apsara Conference Sep 22–24 in Hangzhou is the likely Qwen announcement window; no EU AI Act or provider pricing deadlines this week.
Qwen-Image-2.1 lands open — under a research-only licence
Alibaba's Qwen released Qwen-Image-2.1, a 7B image model, to a 1M-view X rollout — but the licence is research-only, a walk-back from Apache 2.0.
StepFun's Step 5 Preview API ships; open weights promised October 15
StepFun launched Step 5 Preview: a 600B MoE, 27B active, 1M context, at $2.70 per million output tokens. Open weights are promised October 15, no license named.
Vercel gateway data shows open models at a record 78.4% of tokens
X is celebrating Vercel's claim that open-weight models hit 78.4% of AI Gateway token volume, with open-lab spend passing OpenAI. It is one gateway's sample.
How to Run Qwen3.8 via API in 2026 (Hosts Compared)
Qwen3.8 ships a self-hostable 27B dense model plus a 2.4T Max flagship with fresh open weights. Run it locally, via API, or both — without guessing.
MORENA: builders debate a 1.5B open model for African languages
Vambo AI's MORENA claims from-scratch 1.5B open weights for 12 African languages on a $40K train, and grassroots X accounts are amplifying it fast.
Latham & Watkins builds its own AI stack on open-weight Nemotron 3
Latham & Watkins bought Nvidia GPU servers and is fine-tuning open-weight Nemotron 3 on its own legal data. X calls it the enterprise sovereign-AI template.
Open-weights advocates answer Amodei's evaluator plan
Amodei's call for embedded evaluators at frontier AI labs drew 41.5M X views — and a counter from open-weights builders who say public weights verify more.
Amodei tells AI labs to slow down, Anthropic opens doors to auditors
Dario Amodei published a three-step plan to pace frontier AI and committed Anthropic to embedded third-party evaluators; Sam Altman and Elon Musk backed it.
The 8 DeAI stories that mattered this week
The week in open and decentralized AI: Nvidia's $12.9B Hugging Face deal, NEAR's TEE-verified inference, DeepSeek's silent swap, and Mistral's €3B raise.
DeepSeek retires V4-Flash in a silent swap, spares V4-Pro
DeepSeek retired legacy V4-Flash IDs into V4.1-Flash ($0.30/$1.20 per M) and reversed its V4-Pro shutdown — two changes buried in pricing-page footnotes.
This Week in DeAI: Hardware-Attested Privacy Goes Mainstream
The week AI privacy moved from policy to proof: NEAR TEE inference, Anthropic's retention reversal, Nvidia's $12.9B Hugging Face deal, and EU AI Act guidance.
DeepSeek V4.1-Flash: builders argue over the 890-byte KV cache
DeepSeek's MIT-licensed V4.1-Flash drew 1.37M X views in five hours. Verified: 552B MoE weights on Hugging Face, 1M context, $0.003/M cached input.
IFM releases K2 Horizon open models, weights up to 375B
The Institute of Foundation Models released K2 Horizon, an Apache 2.0 model family from 0.9B to a 375B-parameter MoE, all with 512K context windows.
MiniCPM5-2B: builders debate the best sub-4B open model
OpenBMB's MiniCPM5-2B release drew 229k X views in a day. Verified: Apache-2.0 weights, full data and recipes, top sub-4B score from Artificial Analysis.
Nvidia agrees to buy Hugging Face for $12.9B
Nvidia has agreed to acquire Hugging Face, the hub hosting 3M+ open-weight models, for $12.93 billion. The deal is expected to close in H1 2027.
What Are Open-Weight Models? Licenses, Access & Limits (2026)
Open-weight models publish their parameters for anyone to download and serve. How they differ from open source, what licenses allow, and what stays closed.
The 10 Best Open-Source LLM API Providers in 2026 (Full Comparison)
A neutral 2026 comparison of 10 open-weight LLM API providers — how they differ on catalog, speed, pricing model, and data policy, plus the two-line switch.
GPT-5.5 vs Open Models in 2026: Can DeepSeek V4, Kimi K3 Replace It?
GPT-5.5 vs open models in 2026: where DeepSeek V4 and Kimi K3 already replace it, where they don't, and a 4-point checklist to decide for your workload.
Open-Weight vs Open-Source AI Models: The Difference That Bites (2026)
Open-weight means you can download the weights; open-source means you get real rights. Why the gap matters in 2026, with 4 license traps to check first.
How to Run DeepSeek V4 Pro via API in 2026 — Every Host Compared
Every way to run DeepSeek V4 Pro via API in 2026 — official API, third-party hosts, aggregators, self-hosting — plus its self-reported 80.6% SWE-bench score.
How to Run GLM-5.3 and GLM-5.2 via API in 2026 (Hosts Compared)
GLM-5.3 shipped on 2026-08-14 under the MIT license. Learn how to call GLM-5.3 and GLM-5.2 via Z.ai or third-party hosts with a one-line base-URL swap.
How to Run gpt-oss-120b via API in 2026 (Apache 2.0)
Run OpenAI's gpt-oss-120b through any OpenAI-compatible API or self-host the 120B MoE on one 80 GB GPU — providers, code, and usage-policy notes.
How to Run Kimi K2.5 via API in 2026 (Hosts Compared)
Run Kimi K2.5 through any OpenAI-compatible API: reference pricing is ~$0.60/M input and ~$3.00/M output tokens, with seven hosts compared and copy-paste code.
How to Run Kimi K3 via API in 2026 (Price, Context, License Caveats)
Kimi K3 is a ~2.8T-parameter open-weight MoE you can call from any OpenAI-compatible API. Where to run it, what it costs, and the license fine print.
How to Run Llama 4 (Maverick & Scout) via API in 2026 — License Traps
Run Llama 4 Maverick or Scout through any OpenAI-compatible API — plus the two license traps to check first: the 700M MAU clause and EU multimodal carve-out.
How to Run Qwen3.6 (27B & 35B-A3B) in 2026: Local, API, or Both
Qwen3.6 ships as a 27B dense model that fits one high-VRAM GPU plus a 35B-A3B MoE. Learn to run it locally, call it via API, or combine both.
How to Run Qwen3 Coder 480B via API in 2026 (Apache 2.0)
Call Qwen3 Coder 480B from any OpenAI-compatible API: provider options, Python and curl setup, agent wiring, and pricing for the 480B-A35B Apache 2.0 model.
