Qwen3-ST-The-Next-Generation-v1-256k-ctx-6B-qx86-hi-mlx

Benchmark	                      qx86-hi	bf16	Δ
Arc_Challenge (Hard Science)	    0.460	0.456	+0.004
OpenBookQA (Rule-Based Knowledge)	0.414	0.404	+0.010

Method	Arc_Challenge OpenBookQA Arc_Easy BoolQ Overall Star Trek Fit
qx86-hi	        0.460	0.414	0.582	0.732	✅ Best for core tasks (science + knowledge)
bf16	        0.456	0.404	0.574	0.712	❌ Worse in key tasks
qx5-hi	        0.458	0.396	0.580	0.746	⚖️ Good science but weak OpenBookQA
qx6-hi	        0.456	0.418	0.581	0.727	⚖️ Strong OpenBookQA but lower Arc_Challenge
mxfp4	        0.441	0.402	0.557	0.811	❌ Poor science reasoning, only wins on BoolQ
qx64-hi	        0.443	0.416	0.562	0.772	⚖️ Subpar for hard reasoning despite good BoolQ/PiQA
qx86 (non-hi)	0.458	0.414	0.582	0.731	⚖️ Similar to qx86-hi but worse Arc_Challenge

Method	        Why It Fails
mxfp4	        Catastrophic drop in Arc_Challenge (0.441) — can't handle deep science problems
qx64-hi/qx64	Low-bit attention → weak reasoning for complex physics/ethics scenarios
qx86 non-hi	    Coarser attention precision → Arc_Challenge 0.458 vs qx86-hi's 0.460 (small but meaningful loss)

They're building agent code (in Haskell) that interfaces with Postgres-based workflow orchestrator
The ORM is designed to be agnostic - works for any kind of task: HTTP requests, file ops, tool calls
They care about logging to SQLite and using the DB via specific Postgres functions (agent_login, get_tasks, etc)

Get a minimum working version (MWS) of this agent out quickly
Give them control over it from the CLI
Include monitoring and crash safety details (because this fails too often)
Explicitly do not start the agent as a network service at first
Give an example of how to connect this agent with what's already in the archive

NO new PHP/Scala/Rust code
NO new Postgres functions → use our legacy ones
NO Haskell compiles → only direct Python commands from our archive/safe-python namespace

# Get a pre-existing Task Runner we built for the last Enterprise ship
/archive/safe-python/tasks/agent-runner.py \
  --target=database \
  --tool=ollama \
  --task="select 'STARTING RECOVERY'" from pg_node

🔥 What it does: