AI Coding Live News

LightOn open-sources ColGrep-Agent-2B: local 2B code-search agent cuts Claude Code API tokens ~34% and cache-aware cost ~40%

On 2026-10-12 LightOn AI released ColGrep-Agent-2B (fine-tuned from MiniCPM5-2B, Apache-2.0): a local 2B code-localization agent that searches with ColGrep multi-vector retrieval plus a read-only terminal, then returns file paths and line ranges to a frontier coding agent. On 52 code-search questions across 13 unseen repos (624 Claude Code / Sonnet 5.5 sessions), delegating search cut API tokens ~34.3% and cache-aware cost ~40% vs grep, with 80.8% vs 83.3% accuracy (not statistically significant). On SWE-bench Lite localization it posted 19.92 line F1—best among models evaluated in the same ColGrep harness, including several >10× larger. Ships brew/cargo/GGUF installers and hooks for Claude Code, Codex, and OpenCode.

LightOn open-sources ColGrep-Agent-2B: local 2B code-search agent cuts Claude Code API tokens ~34% and cache-aware cost ~40%

Key Takeaways

  • Official: HF Blog — ColGrep-Agent (2026-10-12); model lightonai/colgrep-agent-2B; repo lightonai/next-plaid
  • Vendor Claude Code + Sonnet 5.5 study (52×3): −34.3% tokens vs grep (95% CI 28–40), ~−40% cache-aware cost (7.3¢→4.4¢/q); accuracy 80.8% vs 83.3%
  • Localization: SWE-bench Lite file F1 76.18 / line F1 19.92; Multi-SWE-bench Flash file F1 62.57 / line F1 15.93 (5-seed avg)
  • Train: MiniCPM5-2B → SFT (Qwen3.5-27B teacher) → GRPO 500 steps; tools: colgrep / read-only terminal / finish; 10-turn budget
  • Install: brew install lightonai/tap/colgrep && colgrep --install-agent; default GGUF ~2.5GB; hooks for Claude Code / Codex / OpenCode
Read full story

vLLM: day-0 MiniCPM5-2B on stable — 131K context, Think/No-Think, tool calling

vLLM (@vllm_project) announced day-0 support for OpenBMB MiniCPM5-2B on stable vLLM: 2.6B dense with native 131K context, Think/No-Think from one checkpoint, and tool calling via vLLM’s minicpm5 parser. Built on stock LlamaForCausalLM with training data opened alongside weights.

Key Takeaways

  • Engine: day-0 MiniCPM5-2B on stable vLLM.
  • Capabilities: native 131K context; Think/No-Think from one checkpoint; tool calling via minicpm5 parser.
  • Ecosystem: stock LlamaForCausalLM; training data opened with weights.
Read full story
Linked Models & Agents:
⚡ Openbmb Minicpm5 2b🌐 vllm🌐 openbmb

MiniCPM5-2B scores 15 on AA Intelligence Index v4.2 — top open-weights model under 4B

Artificial Analysis (@ArtificialAnlys) scored OpenBMB MiniCPM5-2B (2.6B dense, 131k context, Apache 2.0) at 15 on Intelligence Index v4.2 — highest among open-weights models under 4B total parameters. Strong small-model agentic results (GDPval-AA / τ³-Banking); coding/long-context benches like Terminal-Bench lag larger peers.

Key Takeaways

  • Model: OpenBMB MiniCPM5-2B — 2.6B dense, 131k context, Apache 2.0.
  • Score: AA Intelligence Index v4.2 = 15, highest open-weights under 4B (vs Granite 4.2 3B at 11).
  • Tradeoff: strong small-model agentic benches; weaker on Terminal-Bench / HLE-style coding and knowledge.
Read full story
Linked Models & Agents:
⚡ Openbmb Minicpm5 2b🌐 openbmb