rafw007/ glm-4.7-flash-opencode

103 1 week ago

A family of custom models built on **GLM-4.7-Flash** (MoE, 30B total / 3B active), tuned to act as autonomous coding agents — **each variant targeting a specific harness**

tools thinking 
ollama run rafw007/glm-4.7-flash-opencode

Applications

Claude Code
Claude Code ollama launch claude --model rafw007/glm-4.7-flash-opencode
Codex App
Codex App ollama launch codex-app --model rafw007/glm-4.7-flash-opencode
OpenClaw
OpenClaw ollama launch openclaw --model rafw007/glm-4.7-flash-opencode
Hermes Agent
Hermes Agent ollama launch hermes --model rafw007/glm-4.7-flash-opencode
Codex
Codex ollama launch codex --model rafw007/glm-4.7-flash-opencode
OpenCode
OpenCode ollama launch opencode --model rafw007/glm-4.7-flash-opencode

Models

View all →

1 model

glm-4.7-flash-opencode:latest

19GB · 198K context window · Text · 1 week ago

Readme

GLM-4.7-Flash Coder — local family for agentic harnesses

A family of custom models built on GLM-4.7-Flash (MoE, 30B total / 3B active), tuned to act as autonomous coding agents — each variant targeting a specific harness (opencode, codex, and Claude Code soon). The models speak native tool-calling, so they fully run agentic coding tools locally — your code never leaves your machine, and your cloud token cost drops to zero.

In thinking mode the base GLM-4.7-Flash is very slow on Apple Silicon (several minutes for a simple reply). This family separates reasoning from output: thinking goes into a dedicated thinking field, while content carries only the clean answer and tool calls. The result: the model responds immediately instead of monologuing, and the output is clean for the harness parser.

Models in the family

Model Base Context Purpose
glm-4.7-flash-opencode GLM-4.7-Flash (MoE 30B / 3B active) 64K Published. Coding agent tuned for opencode (also works in codex). Clean content, no hallucinations, real tool-calling.
glm-4.7-flash-claude-code GLM-4.7-Flash 64K In progress. Variant for Claude Code (CC overrides thinking control — needs a dedicated renderer/template).

What it’s for

  • Agentic work in opencode / codex with native tool-calling.
  • Writing and editing agent code — files, edits, full agent loops.
  • Sysadmin/DevOps tasks in the terminal (disk, network, scripts).
  • Full privacy and offline operation — no code is sent to the cloud.

Quick start

ollama run rafw007/glm-4.7-flash-opencode

In opencode:

ollama launch opencode rafw007/glm-4.7-flash-opencode

Behavior tuning

  • Thinking out of the content. /nothink together with our SYSTEM prompt moves reasoning into a separate field — no monologue leaks into content.
  • No hallucinations. The model reports only values actually present in the tool output — no invented hosts, hardware, or numbers.
  • Acts, doesn’t ask. Inspect / scan / check / measure → it runs the command; its output is the answer.
  • macOS-aware. Uses df -h, du, nmap, system_profiler instead of Linux-only commands.

Sampling / context

  • temperature 1.0, num_ctx 65536 (64K).
  • GLM-4.7-Flash natively carries 128K context — you can raise it further on stronger hardware.

Test hardware

  • Mac Studio M2 (Apple Silicon), 32 GB class, macOS — performance reference.
  • Mac Mini M4, 32 GB RAM — works, though slower.
  • Ollama 0.24, GPU inference (Metal).

Measured behavior

Task Verdict
du / df Read the real disk output, no fabrication.
nmap Handled permission limits and returned 22 real hosts.
Tetris Full, working implementation — 396 lines (score, levels, next-piece preview, controls, game-over screen).

Performance (measured, same model, 64K context, 100% GPU, ~25 GB in memory):

Hardware Generation Prompt eval
Mac Studio M2 (32 GB) ~46 tok/s ~494 tok/s
Mac Mini M4 (32 GB) ~25 tok/s ~250 tok/s

The Studio is nearly 2× faster at identical quality — the difference comes from memory bandwidth, not the model.

Limitations

  • Claude Code incompatible — CC overrides thinking control and hangs the base build. A dedicated -claude-code variant is in the works.
  • In opencode the </tool_call> tag is sometimes printed as text (parser mismatch on the harness side).

How they were made

Designed, built, and tested with Claude Opus — the idea: the world’s best coding model builds smaller models in its own image that take over the work right on your desk. The system prompts, parameter choices, and context configuration come directly from that work.

License

MIT (inherited from the base GLM-4.7-Flash).

© 2026 rafw007