Fixing the deployment target splits the flash-tier coding leaderboard into three winners

Terminal-Bench 2.1 of 84.3 for GLM-5.3-Flash and Terminal-Bench of 79 for DeepSeek V4 Flash are not the same measurement ^[4][7]. DeepSeek's 79 came from an Artificial Analysis run, while its own card reported 82.7, so the independent harness took 3.7 points off the vendor figure ^[8][16]. The table does not include an independent rerun of GLM's 84.3. Subtract the same 3.7 and it lands at 80.6, or 1.6 points ahead of DeepSeek ^[17]. That subtraction only means something if both models were scored on the same harness version with the same scaffold.

If you redistribute weights, the licence column shortens the list before the score columns do. GLM-5.3-Flash is a 320B mixture-of-experts with about 18B active parameters under MIT ^[3]. DeepSeek V4 Flash is a 284B MoE with roughly 13B active, also MIT ^[6]. That is 5.6 per cent of parameters active in one and 4.6 per cent in the other ^[19]. Qwen3.8-Flash-Next runs around 6B active under Alibaba's Qwen Community licence, which the dev.to writeup notes is not a straight OSI licence ^[9]. MiniCPM5-2B is 2.52B dense under Apache 2.0, and runs in llama.cpp and Ollama ^[13].

The launch promotion priced GLM-5.3-Flash at $0.075 per million input tokens and $0.25 per million output until 9 September 2026, with throughput around 49 tokens per second ^[5]. The comparison's last-verified date is 8 September 2026 ^[14]. The promotion outlived the verification by one day ^[20]. No per-token price for DeepSeek V4 Flash appears in the comparison ^[22]. It claims the cost lead but prints no price.

On volume work the dev.to writeup argues that a two-to-three point benchmark deficit costs less than the token bill difference across batch refactors, test generation and CI-time review over thousands of files, and that it does not when you are debugging one hard concurrency bug ^[15].

MiniCPM5-2B, from OpenBMB, claims 46.4 on SWE-bench Verified, which is 32.6 points below the 79.0 on DeepSeek's April model card ^[12][7][18]. Both numbers come from vendor cards. The same MiniCPM card claims 69.1 on LiveCodeBench v6 and a 131,072-token context on a 2.52B dense model ^[12][13]. The comparison's own caveat is that many headline numbers are vendor-reported and independent runs have come in lower ^[14].

Of the published figures, Qwen3.8-Flash-Next's SWE-bench Multilingual of 81.0 is the one the writeup treats as predictive: it says that is the number that matches your experience if your repository is not English-commented Python ^[11]. The same source cautions that the model's LiveCodeBench v6 of 91.9 is vendor-reported, and that scores of that size have compressed historically when independent evaluators rerun them on held-out problem sets ^[10][11]. Its explanation for GLM's agentic lead is state tracking, not reasoning: agentic benchmarks punish models that lose track across twenty tool calls, and it says GLM-5.3-Flash holds its plan longer ^[21].