Model Releases
31 analyses from Devence Lab
· 2 min read
Gemini 4 Hydrogen makes model routing a measurable deployment decision
Google's faster Gemini 4 tier comes with claims about agent task success, speed and quota use. Teams should test complete workflows and permission failures before routing production work to it.

· 2 min read
BioBasicsBench makes failed controls the test for lab agents

· 2 min read
Mistral Large 4 open weights broaden defensive AI deployment choices

· 3 min read
Voxtral 3 makes voice-agent tool authority a real-time control
· 2 min read
Nemotron 3.5 Super VL needs workload evidence beyond its throughput headline
· 2 min read
Gemini 4 Argon's benchmark spread makes local acceptance tests mandatory
· 2 min read
Haiku 5.5 makes model routing an economics control, not a quality compromise
· 2 min read
Nemotron's two gold-level results make post-training the capability, not the checkpoint
· 2 min read
EmbeddingGemma 2 moves multimodal retrieval onto the device. The security boundary moves with it.
· 2 min read
Beam's 501B headline is not the deployment fact. The weights are still missing.
· 2 min read
GLM 5.3 on Bedrock makes the security harness part of the model evaluation

· 2 min read
Data residency is now a model-routing constraint. Bedrock’s India expansion makes it enforceable.

· 2 min read
GPT-6 Sol is 50% cheaper. That makes routing discipline more important, not less.

· 2 min read
A 500K context window is capacity, not memory. Budget context like infrastructure.

· 2 min read
Lower cost per task is not a cheaper token. Route work by task shape, not model tier.

· 2 min read
Four reasoning levels turn model choice into a runtime policy. Benchmark the effort, not just the model.

· 2 min read
A 1-million-token context window is not a memory strategy. Cache the stable prefix.

· 2 min read
13 inference launches do not make one serving stack. Optimise for the bottleneck you can measure.

· 2 min read
SSO is not an MCP privilege boundary. Authorise every tool call again.

· 2 min read
A vector store is not a RAG default. Retrieval economics belong in the architecture decision.

· 2 min read
95.2% recall is not proof that PII is gone. Redaction needs a reconciliation gate.

· 2 min read
Six misalignment reports are not a model card. They are an incident-response template.

· 2 min read
90% cheaper cached input is not a pricing trick. It changes what belongs in the prompt.
· 2 min read
Agent consent is not an OAuth screen. It is an execution boundary.
· 2 min read
A 7% pass rate is not a step change. It is a benchmark with no floor yet.
· 2 min read
Release notes just became compliance artifacts
· 2 min read
Capability thresholds are becoming a disclosure norm. Deployers should read them as a handoff.
· 2 min read
Your model's deprecation date is a risk you do not control
· 2 min read
A model scored 100% on ExploitBench. That tells you about the benchmark.
· 2 min read
GPT-6, Grok 4.7 and Gemini 3.8 shipped inside ten days. Your qualification cycle did not.
· 2 min read
The first cyber-defence model shipped as a product, not a research artefact