A 1T open-weight model you cannot download yet, and an open-weight rival ahead of it on two rows
Mistral announced Large 4 on 6 October: one trillion total parameters, 49 billion active per task, and the in-house nickname "Le Chonk". It takes text and images as input and emits text only. What shipped yesterday is an API preview — Mistral's own announcement says the weights come "by the end of the month", while TNW puts the date at 27 October. The lab says it trained the model from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European datacentres; TNW adds a two-month run drawing about 10 megawatts, neither of which appears in the announcement. The licence the weights will carry is not named in that announcement, and for a European team planning to self-host it is the detail that decides whether any of this is usable.
Independent measurement exists, and it is not flattering. Artificial Analysis scored the preview at 38 on its Intelligence Index on launch day, and Mistral had the third-party evaluator vals.ai run its legal and finance tests. The competitor scores in the launch material are Mistral's own, not a reporter's reconstruction, and that is what makes them worth reading: its DeepSWE v1.1 chart — run privately by Artificial Analysis before that harness launched publicly — puts Large 4 at 62 (61.7% in the text) against 68 for Kimi K3, 61 for GLM-5.3, 57 for DeepSeek-V4-Pro-0813, 51 for Qwen3.8 Max and 44 for Beam's self-reported figure. The one-point lead over GLM-5.3 is inside the noise of a single harness; the six points to Kimi K3 are not. The finance row goes the same way: on Mistral's Finch (FinWorkBench) chart Large 4 scores 67.4%, exactly level with DeepSeek-V4-Pro-0813 and nearly ten points behind Kimi K3 at 77.3%.
Where Large 4 does lead is law and vision. On Mistral's chart of vals.ai's Harvey legal agent benchmark it scores 15.8% against 12.9% for Kimi K3, 10.4% for Qwen3.8 Max, 8.3% for GLM-5.3, 7.5% for DeepSeek-V4-Pro-0813 and 5.4% for GPT-6 Astra — a clear win in a field where mid-teens is the state of the art, which says more about the benchmark than the model. On visual grounding Mistral reports Dense 200 at 42% against 41% for GPT-6 Astra, and TNW renders the same pair as a 42–42 tie. Mistral also puts the model's combined Coding Agent Index at 49.8%, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. The honest summary for a European buyer is narrower than the launch framing: a European lab will shortly put trillion-parameter weights in your hands, under a licence it has not yet named, and on its own numbers the best open weights for agentic coding and finance work are still Chinese.
A second Langflow batch, five days old on IBM's site and only now in the CVE feed
IBM published a security bulletin on 2 October titled "Langflow OSS is affected by multiple vulnerabilities", covering 25 CVEs in versions 1.0.0 through 1.12.2, fixed in 1.12.3, with the workaround field reading "None". Twenty-four of those records reached the National Vulnerability Database overnight, between 00:17 and 01:16 UTC on 7 October; CVE-2026-97677 (8.1, authenticated path traversal) is the one still absent. This is a separate batch from the five Langflow flaws published to NVD on Monday 5 October, which were fixed in 1.9.0 through 1.10.3 — the new affected range runs well past those releases, so a team that patched on the basis of last week's advisory is not covered.
Fourteen of the twenty-four NVD records state arbitrary code or OS command execution outright, counted from the descriptions, and a fifteenth (CVE-2026-93447, 7.5) is a deserialisation RCE that needs the server secret and Redis write access. Three need no credentials at all, and two of those need no user interaction either: CVE-2026-104334 (9.8, improper control of code generation) and CVE-2026-93674 (9.8, OS command injection), both AV:N/AC:L/PR:N/UI:N. The third, CVE-2026-93675 (8.8), is a dependency confusion that needs a user to act. The rest are authenticated: a sandbox escape in the code executor (CVE-2026-97676, 8.8), an incomplete blocklist in Langflow's own code security scanner (CVE-2026-97655, 8.8), credential exposure (CVE-2026-101331, 7.7), and cross-tenant reads in the vertex result cache (CVE-2026-97680, 8.3). Neither IBM nor the records say whether any of this is exploited. The operational point is the five-day gap: the fix existed on 2 October and the CVE feed most teams actually watch only learned about it this morning.
China's open-weight labs are raising their last private rounds before going public
DeepSeek is close to raising at least 80 billion yuan, about $12 billion, in a round that could approach 100 billion yuan, with Tencent and the battery maker CATL among the largest contributors — Bloomberg's reporting, relayed by TNW on 6 October. That follows a roughly 50-billion-yuan round earlier this year at a $52–59 billion valuation, in which founder Liang Wenfeng put in about 20 billion yuan, Tencent about 10 billion and CATL about 5 billion, alongside JD.com, NetEase and China's national AI investment fund. Reuters reported last month that DeepSeek has engaged CITIC Securities to prepare for a possible listing on Shanghai's STAR Market, with timing, valuation and scale all undetermined.
Moonshot AI has closed what is described as its final private round at a valuation of roughly $50 billion and is targeting a Hong Kong IPO in the first quarter of 2027, with preliminary investor meetings possibly starting this month — again Bloomberg, via Investing.com. For anyone running Kimi or DeepSeek weights on their own infrastructure, the thing that changes is the incentive structure behind the next release. A listed lab with quarterly disclosure has reasons to keep its best weights behind an API, to price hosted access defensively, and to deprecate on a schedule that suits a reporting calendar rather than a self-hosting customer. Open weights already in your possession stay yours; the question is what the 2027 releases look like.
Quick Hits
- A malicious MCP server could collect a client's OAuth secrets with no user interaction. CVE-2026-104850 (7.5), published 6 October, affects the MCP TypeScript SDK from 1.12.0: stored credentials were not bound to the authorization server they belong to, so a server naming its own authorization server in its protected-resource metadata receives the client's stored
refresh_tokenandclient_secret. Fixed in@modelcontextprotocol/sdk1.31.0 and@modelcontextprotocol/client2.2.0; on 1.x the only workaround is to connect OAuth-enabled clients to trusted servers only. - Two flaws in Amazon's Bedrock AgentCore starter toolkit fire when you import someone else's agent. CVE-2026-105812 (8.8 on CVSS 4.0, 9.0 on CVSS 3.1) puts unescaped configuration values into generated Python, and CVE-2026-106032 (5.7) turns OpenAPI schema processing into SSRF plus local file read. Both need a same-account actor; CVE-2026-105812 also needs the victim to run or deploy the imported agent, while CVE-2026-106032 fires on import alone. AWS has deprecated the starter toolkit in favour of
@aws/agentcore. Upgrading to 0.3.14 is not enough on its own for CVE-2026-105812: AWS says agents imported with an affected version must be re-imported on 0.3.14 or later and their local and deployed output artifacts replaced. - LangChain patched two tenant-boundary bugs in its storage integrations. In
@langchain/redis, CVE-2026-105799 (2.3, fixed 1.1.1) lets attacker-controlled values inject RediSearch syntax and broaden a filter used as an access boundary; in@langchain/mongodb, CVE-2026-106119 (6.0, fixed 1.3.1) lets a structured session identifier be read as a MongoDB query condition in a shared chat-history collection. Both scores are CVSS 4.0. - South Korea will spend 4.7 trillion won on a state frontier model from March 2027. Reuters reports the Ministry of Science and ICT will pick a lead developer by competitive tender once parliament approves the 2027 budget in December, with the winner possibly chosen as early as February, open to single companies, consortiums or special-purpose vehicles. About $3.5 billion, and separate from the existing home-grown foundation model programme.