The 8 DeAI stories that mattered this week

The week open weights got a deadline on the most consequential drop yet. Moonshot AI put Kimi K3.1 Pro on the API and pinned the full weights of a 2.8-trillion-parameter model to November 4 — a date that, if held, would put the highest-scoring open-weight model in builder hands. Around it, the infrastructure layered up: You.com routed twelve open models behind an enterprise SLA it has not yet audited, Mubit Labs shipped a small verifier whose 45% answer rate became the week's sharpest argument about what a benchmark is for, and DeepSeek kept converting its own stack from a press release into inspectable code. Eight stories mattered; here is what each one changes.

Key facts

• November 4 — Moonshot's promised date for Kimi K3.1 Pro's full open weights; until then the model is API-only at $2/$10 per million tokens, scoring 55 on the Artificial Analysis index, ahead of GPT-6 Astra and Gemini 4 Argon (53 each) and behind Claude Opus 5.5's 58.
• 12 models, one endpoint — You.com's Inference Router launch, with a claimed 99.9% uptime SLA and zero data retention, both provider claims with no independent audit published.
• 779 of 1,731 — Banyan Sapling's calibrated-answer count on its own model card eval: every confident answer correct, 952 cases returned unsure, a 45% answer rate.
• ~1.5 TB — the BF16 memory floor for Meta's Muse Spark Open at roughly 816B total parameters, per the serving guide, against a 42B-active count that reads far smaller than it serves.
• 13.6k stars — DeepInfer, DeepSeek's Apache-2.0 serving framework, within four days of release; the same week saw DeepSeek report a roughly $15 billion raise, unconfirmed by the company.
• 2h41m — the window the compromised LiteLLM 1.107.0 wheels were live on PyPI, long enough to exfiltrate a proxy's entire provider-key table.

1. Moonshot dates the big one: Kimi K3.1 Pro, weights by November 4

The lead story is a promise, and this week's supply wave has taught us to read promises carefully — but this one comes from a lab that has held them before. Moonshot AI put Kimi K3.1 Pro on the Kimi API at $2.00/$10.00 per million tokens, a post-trained step up from July's Kimi K3, scoring 55 on Artificial Analysis' independent Intelligence Index. That is third among tracked models, ahead of GPT-6 Astra and Gemini 4 Argon (53 each) and behind Claude Opus 5.5's 58 — and, per The Decoder's framing, if the weights ship as planned, K3.1 Pro would become the highest-scoring open-weight model available. Two details deserve builder attention beyond the headline. First, the disclosure that Moonshot removed cyber-related data from the mid-training and RL mix ahead of the open-weight release — a first-of-its-kind pre-release data change for a major open-weight drop, meaning the November download will not be the model behind today's API scores on security-adjacent tasks. Second, the pricing: roughly a third below Kimi K3's July launch rates, with a Highspeed variant claiming 2-3× the speed. Why it matters: a dated commitment on a frontier-scale model is exactly the promissory-note structure this beat has been tracking — the gap between announcement and artifact — and November 4 is the next date that gap either closes or does not. Full story.

2. You.com routes 12 open-weight models behind one enterprise SLA

The rental layer got an enterprise face. You.com launched Inference Router on October 22: twelve open-weight models behind one OpenAI-compatible endpoint, US-hosted, with a claimed 99.9% uptime SLA and zero data retention. The announcement drew roughly 395,000 views in a day. The claim structure is what makes this worth watching rather than repeating: a hosted router's zero-retention promise says nothing about what the upstream inference provider logs, and no retention policy document or third-party attestation backs the claim yet. Against that, the launch is a real signal about the buy side — a procurement team can now put an open model behind a contract with credits attached instead of defending self-hosting to a risk committee. Why it matters: the router sits precisely at the fault line this beat covers, the rent-versus-own decision, and its SLA-plus-zero-retention pitch is an attempt to make the rental side look enterprise-grade without the audit that would make it verifiable. Full story.

3. Mubit's Banyan Sapling: a verifier that says unsure 45% of the time

The week's most argued-about model is also its smallest. Mubit Labs released Banyan Sapling, an Apache 2.0 23B verification model — a pruned Qwen3-30B-A3B backbone with BF16 and FP32 adapters, sized for a single 48GB-class GPU — that reads a policy, evidence, and a claim and returns a verdict: supported, violation, missing, contradicted, or unsure, with calibrated probabilities. The launch thread called it "perfect on 779 of 1,731," which is exactly half true: every one of the 779 calibrated answers was correct, and the model returned unsure on the other 952 cases, a 45% answer rate. On ContractNLI, where the domain gap is widest, coverage was 0.82%. The precision claim is a vendor self-report pending independent reproduction; the coverage number comes from Mubit's own card. Why it matters: while October's flagship drops chase the frontier at 501B-816B parameters, this is the open-weights wave going the other way — a small, auditable component built to sit in an agent trust pipeline rather than to write code — and whether a 45% answer rate clears a production bar depends entirely on the pipeline's tolerance for "unsure." Full story.

4. The serving bill: 816B parameters, 1.5 TB of memory

The week's sharpest serving lesson came from Meta. Muse Spark Open's gated Hugging Face release puts roughly 816 billion parameters of multimodal weights in builder reach — text, image, audio and video inputs with text output, a context window just over a million tokens — with about 42 billion parameters active per token. The gap between those two numbers is the deployment story: Hugging Face describes the BF16 checkpoint at about 1.5 TB, and reports 18 tokens per second for a calibrated 2-bit llama.cpp quantization on an M3 Ultra with 512 GB of memory, using about 304 GB. Why it matters: active-parameter counts are a compute description, not a memory budget — a model that "activates" 42B can still need its entire expert set resident, and any ownership decision made against the smaller number is a hardware decision made against the wrong figure. Full story.

5. DeepSeek keeps shipping the stack, not just the weights

Two DeepSeek stories this week point the same direction. The company shipped V4.1-Flash-1017 under MIT with higher reported agent and coding scores and, notably, no Jinja chat template — the release ships a Python reference encoder and Rust libraries for request conversion and output parsing instead, handing over the thing teams usually reverse-engineer. Then it open-sourced DeepInfer, the Apache-2.0 serving framework it describes as powering large-scale production V4.1-Flash traffic, at 13.6k GitHub stars within four days — with the honest caveat that it covers offline inference only, no KV cache system, no prefill/decode disaggregation, no load balancing. Around the artifacts, the money and the policy moved too: Bloomberg reported a roughly $15 billion round from Tencent and CATL, unconfirmed by the company, three days after the DHS order excluded DeepSeek products — including third-party products built on DeepSeek implementations — from federal civilian IT systems. Why it matters: publishing the serving layer is how a vendor makes its stack inspectable, and the same week's funding and procurement signals show the capital and the compliance questions scaling alongside the code. Full story, the funding, the DHS order.

6. The key vault was the soft target — again

The week's most consequential failure was not a model but the plumbing between them. LiteLLM 1.107.0 published to PyPI and Docker on October 21 carrying a credential stealer and a remote access trojan; the malicious wheels were live under three hours and the Docker tags about two and a half. In that window the stealer read the proxy's own stored-key table — every provider API key a team owns — along with AWS, GCP and Azure credentials, SSH and GPG keys, and local tool configs. Wiz reports abuse of stolen credentials about 4.5 hours after exfiltration. It is the second LiteLLM PyPI compromise of 2026. Why it matters: the proxy is a single point of failure in the strongest sense — one compromised install crosses every provider boundary at once, and owning the stack means owning its dependency graph, not only its weights. Full story.

7. The local tier keeps getting real — and keeps needing measurement

Two smaller releases sharpened the on-device case. Liquid AI's LFM2.5-24B-A2B puts 2.45B active parameters to work on device-class hardware, with a reported 184 tokens per second on an M5 Max under Q4_K_M quantization at 4K context — self-reported, with the quantization and context configuration stated, which is the minimum diligence for a speed claim. Recursive's Flint targets the laptop coding case at 39.8B total and about 2.5B active, shipping twelve quantized variants for llama.cpp, LM Studio, Jan and Ollama the day after release, reporting 125+ tokens per second on an M4 Pro — also self-measured. Why it matters: both are plausible and neither has an independent run behind it yet, and the honest boundary on the local story is the one our LFM2.5 coverage draws — local inference keeps model requests on your hardware, but web tools, external services, logs and the agent harness can still send or retain data. "Runs locally" is a property of the model call, not of the workflow. Full story, Flint.

8. The rental side kept repricing

The week's pricing ledger moved on both ends. Cerebras priced gpt-oss-120b at $0.35/$0.75 per million tokens on the fastest endpoint it tracks and documented the cache economics alongside it — the specialist-silicon tier now has a public anchor. And Kyutai shipped CADRE as a full artifact set — paper, code, weights, and a playable demo — for interactive world models reporting sub-230 ms action latency at 20 frames per second on one H100, Kyutai's own measurement, not yet reproduced. Why it matters: the rent-versus-own arbitrage re-prices itself every quarter, and the release cadence this beat tracks is what keeps the own side of that math honest — the fixed-capability price decline only counts if the weights actually arrive, which is why item 1's November 4 date is the number to watch. Full story, CADRE.

Also on the tape

Four items that did not make the top eight but belong in your feed. Apertus 1.5's October technical report documents training stages, data mixtures and intermediate checkpoints for models whose Apache 2.0 weights shipped in July — a reproducibility contribution, not a new release. Fish Audio's Drama 3 is the license case: editable multi-speaker speech under a community license that separates low revenue from hosting permission. The week's daily briefs carry the day-by-day record — Nemotron 3.5, Mercury 3's throughput claim, Voxtral 3, OpenAI on Google Cloud Model Garden, NVIDIA Confidential Bare Metal, and llm-d 1.0. And last week's week-ahead calendar is worth a look in hindsight, having flagged this crowded release window before any of it landed.

The Friday hub, This Week in DeAI: The Weights Came Due — and the Bill for Owning Them, ties stories 4 through 7 into the full arc with complete sourcing, and the daily briefs from Oct 20 through Oct 24 carry the item-by-item record. Last week's edition: The 8 DeAI stories that mattered, week 41.