Ornith-1.5 Goes Open Under MIT, 1-Bit Qwen Runs on 8GB, and Agent Harnesses Become the Real Battleground

Ornith-1.5 Arrives Under MIT

The biggest release of the day is Ornith-1.5, a new open-weight family spanning 9B dense, 35B mixture-of-experts and 397B mixture-of-experts variants, all under the permissive MIT licence and shipping in FP8, GGUF, MLX and NVFP4 formats. The headline claim is end-to-end self-improvement: the model proposes its own tasks, builds scaffolds and generates reinforcement learning rollouts to create fresh training experiences. Reported agentic and coding scores are strong, including 86.1 on Terminal-Bench 2.1 and 86 on SWE-Bench Verified. vLLM and Ollama wired it into their serving stacks almost immediately.

Compression Keeps Getting Braver

Unsloth shipped new Qwen3.8-27B GGUFs built on its Dynamic v3 quantisation, claiming roughly 10% higher accuracy at the same file size, plus 1-bit quants that reportedly retain about 77% of BF16 accuracy while running in around 8GB of RAM. A new Divergence-300 metric extends top-1% greedy accuracy measurement across longer generations. The community wants deeper diagnostics: per-category and KV-cache quantisation KLD numbers, and a direct comparison line against the previous generation of quants.

Harnesses, Not Just Models

TrueFoundry open-sourced TrueForge, an MIT-licensed, self-hostable, vendor-neutral harness for production agents with tool orchestration, context management, subagents, code sandboxes, human approvals and traces. On a 14-task enterprise benchmark it matched Claude Managed Agents on Opus 4.8 while using around 30% fewer tokens, and routing to GLM-5.2 cut cost roughly 75% with accuracy preserved. Meanwhile DeepSeek's harness is being read as an open agent runtime where everything, including the agent loop itself, is a plugin.

Reinforcement Learning Through the Harness

Microsoft's Agent Lightning 1.0 connects arbitrary harnesses to reinforcement learning through an endpoint proxy, handling retokenization, sample merging and advantage calculation. With roughly 6,000 training examples and modest compute it reportedly moves Qwen3.5-9B on SWE-Bench Verified from 41.8% to 56.4%.

Gemini 3.7 Flash Tops the Analyst Agent Board

Google's Gemini 3.7 Flash took first place on the AA-AnalystAgent benchmark with 60.0% pass^5 and 70.5% pass@1 across 80 spreadsheet and document-heavy quantitative tasks, at 1.32 seconds per task and roughly $0.54 average cost. Google also pushed it into Search's AI Mode, Gemini chat and AI Studio.

Cheap Frontier and Private Safety

Replit launched a Free Mode powered by GPT-5.6 Luna, a striking sign that near-state-of-the-art capability is now cheap enough to give away. OpenAI also introduced Private Safety Processing, aiming to preserve Zero Data Retention for frontier models while still detecting cross-interaction safety risks without human access to content.

Retrieval Gets Filter-Aware

Qdrant argued filtered approximate nearest neighbour search belongs in the index itself, not just at query time. Its filterable HNSW reports 99.8% recall at 1.0ms versus 67.7% at 4.7ms for ACORN on a 1% filter over 1M vectors. Separately, Sentence Transformers v6.0 leans into multi-vector, late-interaction retrieval.

Claude Designs Proteins That Actually Work

Anthropic reports Claude autonomously designing disease-targeting proteins validated in real wet-lab assays, with a 35% success rate against a human baseline of roughly 10-15%. And a study instrumenting ten agentic applications found non-model components dominate latency in half of them, with task-aware serving cutting latency 29-40%.