OpenCode 2.0 released
OpenCode overrides Qwen model sampler settings to incorrect values, sending top-p=1.0 instead of the recommended 0.95 (thinking) or 0.80 (no-thinking). The issue is invisible with llama.cpp but exposed by ninfer's logging. A commit added the change without explanation.
OpenCode overrides Qwen model sampler settings to incorrect values, sending top-p=1.0 instead of the recommended 0.95 (thinking) or 0.80 (no-thinking). The issue is invisible with llama.cpp but exposed by ninfer's logging. A commit added the change without explanation.
Jin is a minimalist AI coding agent the author is building after finding Claude Code, OpenCode, pi, and omp didn't fit their needs. It supports the three most popular API standards and routes newer models through a proxy instead of requiring plugin updates. It is aimed at developers who want a simple, boring agent rather than broad provider support.
Jin is a minimalist AI coding agent the author is building after finding Claude Code, OpenCode, pi, and omp didn't fit their needs. It supports the three most popular API standards and routes newer models through a proxy instead of requiring plugin updates. It is aimed at developers who want a simple, boring agent rather than broad provider support.
A user comparing coding agents PI Agent and OpenCode on Qwen 3.8 27b reports PI Agent produces better results in the agent environment, uses fewer tokens, and lacks the 32k output hard limit. The post is the second and final comparison in a series.
A user comparing coding agents PI Agent and OpenCode on Qwen 3.8 27b reports PI Agent produces better results in the agent environment, uses fewer tokens, and lacks the 32k output hard limit. The post is the second and final comparison in a series.
Meta released Muse Spark 1.3 on 2026-09-02, targeting improved performance on agentic and coding tasks and real-world usability. The model rolled out in Muse Code and the Meta Model API, with the Contributor variant available on OpenCode Go and a max-reasoning mode planned after safety testing. A Reddit post reports pricing of $0.10 input / $0.20 output per million tokens, matching Fable 5.
Meta released Muse Spark 1.3 on 2026-09-02, targeting improved performance on agentic and coding tasks and real-world usability. The model rolled out in Muse Code and the Meta Model API, with the Contributor variant available on OpenCode Go and a max-reasoning mode planned after safety testing. A Reddit post reports pricing of $0.10 input / $0.20 output per million tokens, matching Fable 5.
A Reddit user praises Reasonix, a coding harness they switched to from OpenCode, highlighting its first-class DeepSeek integration, support for generic OpenAI- and Anthropic-compatible endpoints, and built-in sandbox and security review skills. The post asks the community about other lesser-known tools, indicating growing interest in alternative coding agents.
A Reddit user praises Reasonix, a coding harness they switched to from OpenCode, highlighting its first-class DeepSeek integration, support for generic OpenAI- and Anthropic-compatible endpoints, and built-in sandbox and security review skills. The post asks the community about other lesser-known tools, indicating growing interest in alternative coding agents.