Three open-model releases, three different bets on what small means
Mistral shipped a big model with an enterprise platform, Fastino shipped five small models that decide instead of generate, and Nanbeige shipped one that declines most questions.
- Mistral released Mistral Large 4, codenamed Le Chonk, with advanced coding and defensive cybersecurity capabilities, alongside its Studio enterprise platform. Self-reported m:…8918
- Mistral's Studio supports deployment in cloud, virtual private cloud, or on-premises, and is already in production at BNP Paribas, ABANCA, and Accenture. Self-reported m:…8918
- Fastino released five open-weight decision models from 0.4B to 27B parameters; the 0.4B model answers in a median 7 milliseconds and the 27B in 112 milliseconds. Self-reported m:…5614
- Nanbeige released Nanbeige4.3-33B, a dense 33-billion-parameter reasoning model under an Apache 2.0 license, scoring 20 on the Artificial Analysis Intelligence Index, matching Qwen3.8 27B. m:…5123
- Nanbeige declines 71% of questions on the Omniscience test and has an 18% hallucination rate there, while scoring 30% on Terminal-Bench 4.0. m:…5123
What shipped
Mistral AI has launched Mistral Large 4, codenamed Le Chonk, alongside significant updates to its enterprise platform, Studio. The company says the model provides advanced coding capabilities and defensive cybersecurity capabilities, and Studio is designed to help organizations build and deploy agents, workflows, and applications while keeping control over their models, data, and infrastructure. Studio now includes managed agents and apps, and developers can build agents that carry out multi-step tasks inside it. m:…8918
Mistral's pitch is control, not capability. Studio supports deployment in cloud, virtual private cloud, or on-premises environments, and provides hosting, authentication, and access controls for enterprise apps. BNP Paribas, ABANCA, and Accenture already use Studio in production. The platform also includes OpenTelemetry traces that show each step of a run, including model calls and tool use, and evaluations that let teams test prompts, models, or tools against datasets before release. That is the boring infrastructure that decides whether a bank actually ships an agent. m:…8918
Fastino took the opposite approach to size. It released the GLiDE family, five open-weight decision models — 0.4B, 0.8B, 2B, 9B, and 27B parameters — designed for classification, routing, and safety tasks. Instead of generating text, GLiDE models output probability distributions for predefined options, doing one forward pass per question without generating tokens. All five share the same System One schema interface, so a team can swap sizes without rewriting the integration. m:…5614
Speed is the feature
The GLiDE numbers are the story. Median latency is 7 milliseconds for the 0.4B model, 18 milliseconds for the 2B, 39 milliseconds for the 9B, and 112 milliseconds for the 27B. Memory needs run from 0.8 gigabytes for the smallest to 52 gigabytes for the largest in bf16 precision. The 27B and 9B models use recurrent looped transformers to increase computational depth without adding parameters, while the 2B skips that technique because the latency increase outweighs the accuracy gain. Fastino says the 27B model is expected to score 73.3 on the Decision Index 0.4.1, a benchmark for decision tasks. m:…5614
Nanbeige, the foundation model research team at BOSS Zhipin, released Nanbeige4.3-33B, a dense 33-billion-parameter reasoning model under an Apache 2.0 license with a 524,000-token context window. It scores 20 on the Artificial Analysis Intelligence Index v5, matching Qwen3.8 27B, but with a different profile: stronger on terminal use and physics reasoning, weaker on knowledge-based tests. It scores 30% on Terminal-Bench 4.0 and 16% on the hard subset of CritPt v2. m:…5123
The most interesting Nanbeige number is not a score but a refusal rate. On the Omniscience test, the model declines 71% of questions and has an 18% hallucination rate. A model that says it does not know most of the time is a strange product pitch, but it is an honest one: for a coding or terminal agent, a refusal is cheaper than a confident wrong answer that deletes the wrong file. The trade-off is explicit in the data — the same model that matches a peer on the headline index is far more cautious in practice. m:…5123
What it means for builders
The three releases point at the same conclusion from different directions: the question is no longer which model is smartest, but which model fits the job. A routing decision that takes 7 milliseconds and never generates a token is a different product from a reasoning model with a 524,000-token window, and neither is a substitute for the other. Teams that keep reaching for the biggest general model for every task are paying frontier prices for decisions a small specialist could make in milliseconds. m:…5614 m:…8918
The second lesson is about evaluation. Mistral ships evaluations that test prompts, models, or tools against datasets before release, and Nanbeige's refusal rate is only visible because someone measured it. A builder who picks a model on a single headline score will miss the profile — the terminal strength, the knowledge weakness, the caution — that decides whether the model survives contact with real users. The week's releases reward the teams that read the whole scorecard. m:…8918 m:…5123
What is still unresolved
- Whether GLiDE 27B actually scores 73.3 on the Decision Index, since Fastino's figure is an expectation, not a measured result.
- How Nanbeige's 71% refusal rate plays in real coding work, where declining too often is its own failure.
- Whether Mistral's Studio controls are enough to win regulated-enterprise deployments against larger US platforms.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.