 | Did some test tasks with v4 flash. The context management, tool use accuracy and thinking traces all looked excellent. It is one of the few open-weights models I have tested that does not get confused with multi tool calls or complex native tool definitions It must have called at least 100 tool calls over multiple runs, not a single error, not even when editing many files at once Downside: slow token generation and takes a while to finish thinking (I have not shown but it thought for good few minutes for planning and execution) Read that deepseek is bringing a lot more capacity online in H2'26. Looking forward to it, LFG [link] [comments] |