Practical notes: The 2B Model That Beats 4B Rivals—Until You Test It as an Agent

Use this as an operator-facing rebuild of the ideas in “The 2B Model That Beats 4B Rivals — Until You Test It as an Agent”: clear stages, ordered code slots, and recovery notes that survive a handoff.

Why multi-turn agentic harnesses tell a different story than single-shot index scores — and how to test it yourself.

The Why multi-turn agentic harnesses stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Budget tokens per turn and per session. Agentic tools expand context aggressively; hard caps keep demos from becoming surprise invoices.

The Claim: A 2.5B Model That Beats 4B Rivals

The The Claim A 2 stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Budget tokens per turn and per session. Agentic tools expand context aggressively; hard caps keep demos from becoming surprise invoices.

What’s Actually Under the Hood

The What s Actually Under stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Budget tokens per turn and per session. Agentic tools expand context aggressively; hard caps keep demos from becoming surprise invoices.

The First Crack: “23 vs 15” on the Artificial Analysis Index

The The First Crack 23 stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Budget tokens per turn and per session. Agentic tools expand context aggressively; hard caps keep demos from becoming surprise invoices.

The Real Gap: Agentic Harnesses Tell a Different Story

The The Real Gap Agentic stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Budget tokens per turn and per session. Agentic tools expand context aggressively; hard caps keep demos from becoming surprise invoices. The The Real Gap Agentic stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline.

Reproducing It Locally: Quantization, Throughput, and Failure Modes

For the Reproducing It Locally Quantization stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Prefer structured outputs with schema validation over free-form prose when the next step is code or a tool call.

# macOS
brew install ollama

# Linux
curl -fsSL https://ollama.com/install.sh | sh

# Recommended flags for flash attention + quantized KV cache
OLLAMA_FLASH_ATTENTION=1 OLLAMA_KV_CACHE_TYPE=q8_0 ollama serve &

# Run directly from Ollama's registry
ollama run openbmb/minicpm5-2b

huggingface-cli download openbmb/MiniCPM5-2B-GGUF \
  MiniCPM5-2B-Q4_K_M.gguf \
  --local-dir ./minicpm5-2b-gguf

pip install vllm

vllm serve openbmb/MiniCPM5-2B \
  --max-model-len 131072 \
  --dtype bfloat16 \
  --port 8000

# Query it like any OpenAI chat completion endpoint
curl http://localhost:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model": "openbmb/MiniCPM5-2B", "messages": [{"role": "user", "content": "Write a quicksort in Rust"}]}'

So When Is This Model Actually the Right Choice?

For the So When Is This stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Prefer structured outputs with schema validation over free-form prose when the next step is code or a tool call.

The Bigger Lesson: Benchmark Claims Are Benchmark-Specific

For the The Bigger Lesson Benchmark stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Prefer structured outputs with schema validation over free-form prose when the next step is code or a tool call. For the The Bigger Lesson Benchmark stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline.

Closing Thoughts

When working through the Closing Thoughts stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Cache stable system instructions and tool schemas. Re-sending identical preamble is a common source of burn.

Resources and acknowledgments

When working through the Resources and acknowledgments stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Cache stable system instructions and tool schemas. Re-sending identical preamble is a common source of burn.

Operational checklist

When working through the Operational checklist stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest.

Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish.

Cache stable system instructions and tool schemas. Re-sending identical preamble is a common source of burn.

Keep graph state flat and typed. Nested blobs hide which node wrote which field and break resume after interrupts.

Add a smoke test that exercises the critical path in CI with fixtures, not live paid APIs, whenever budgets allow.

Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

Before promoting the stack, freeze versions, capture a golden transcript for the critical path, and confirm rollback steps. Shared environments need rate limits, tenancy checks, and a clear owner for secret rotation. Prefer boring reliability over clever one-off demos.

Batch note for 94ea12aa679c: keep provider keys out of the repo, set a per-session token ceiling, and store transcripts next to the eval fixtures so later model swaps stay comparable.

When working through the hardening note 0 stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph.

Hardening detail 0/865: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

The hardening note 1 stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline.

Hardening detail 1/865: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

For the hardening note 2 stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

Hardening detail 2/865: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

When working through the hardening note 3 stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish.

Hardening detail 3/865: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

The hardening note 4 stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion.

Hardening detail 4/865: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

For the hardening note 5 stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph.

Hardening detail 5/865: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

When working through the hardening note 6 stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline.

Hardening detail 6/865: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

The hardening note 7 stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

Hardening detail 7/865: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

The hardening note 0 stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion.

Hardening detail 0/884: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

For the hardening note 1 stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph.

Hardening detail 1/884: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.