Verdict: GLM-5.3-Flash wins for agentic coding, DeepSeek V4 Flash wins on price, and MiniCPM5-2B wins on-device. The best open-source LLM for coding in September 2026 depends on where the work runs, and the split is now clean. GLM-5.3-Flash wins agentic coding, where the model has to drive a terminal and finish a multi-step task. DeepSeek V4 Flash wins on cost per token for high-volume work. And MiniCPM5-2B, released on 7 September 2026, wins on-device: a 2.52B dense model that scores 69.1 on LiveCodeBench v6 and 46.4 on SWE-bench Verified, according to its Hugging Face model card. Qwen3.8-Flash-Next sits close behind GLM on general intelligence and is the pick if you need very long context on a permissive-enough licence.
For a coding agent that runs commands, edits files and iterates until tests pass, GLM-5.3-Flash from Z.ai is the strongest open-weights option. Released on 26 August 2026, it is a 320B mixture-of-experts model with around 18B active parameters, a 1M-token context window and an MIT licence. On the flash-tier comparison published on 4 September 2026 it posts an Artificial Analysis Intelligence Index of 57, Terminal-Bench 2.1 of 84.3 and DeepSWE v1.1 of 63.4, the best agentic coding results of the three (regolo.ai). The practical reason it leads is not raw reasoning. Agentic benchmarks punish models that lose track of state across twenty tool calls, and GLM-5.3-Flash holds its plan longer. If you have used the GLM line for automation before, the workflow carries over largely unchanged; our GLM-5.2 automation guide covers the same harness pattern.
Throughput sits at about 49 tokens per second, and the launch promotion priced it at $0.075 per million input tokens and $0.25 per million output until 9 September 2026, according to the regolo.ai comparison. Check current pricing before you budget on it.
Qwen3.8-Flash-Next, Alibaba's late-August release, is the choice when context length and multilingual code matter more than terminal-driving. It runs roughly 6B active parameters with a 262K native window that extends to 1M using YaRN, under the Qwen Community licence rather than a straight OSI licence. Its Artificial Analysis Intelligence Index of 56 is within a point of GLM's, and it reports SWE-bench Pro 62.5, SWE-bench Multilingual 81.0 and LiveCodeBench v6 of 91.9 (same comparison source).
Treat that 91.9 with care. It is vendor-reported, and LiveCodeBench scores of that magnitude have historically compressed when independent evaluators rerun them on held-out problem sets. The SWE-bench Multilingual result is the more useful signal for real teams: if your repository is not English-commented Python, it is the number that predicts your experience. For a deeper look at how the Qwen flash variants differ from the dense siblings, see our Qwen3.8-Flash-Next comparison and the Qwen3.6-35B-A3B MoE guide.
DeepSeek V4 Flash, refreshed on 31 July 2026, is a 284B MoE with about 13B active parameters, roughly 1M context and an MIT licence. It has the lowest cost per token of the three flash models, and its coding scores are respectable rather than leading: SWE-bench Verified 79.0 from the April model card, Terminal-Bench 79 in Artificial Analysis's independent run, and an Intelligence Index of 50 (same comparison source).
That Terminal-Bench figure is the single most instructive number in this comparison. The vendor reported 82.7; the independent run returned 79. The gap is small, but it is directional, and it is the reason to weight independently measured results above model cards when you are choosing what to put in production.
Where DeepSeek V4 Flash pays off is volume. If you are running batch refactors, test generation or CI-time review across thousands of files, a two-to-three point benchmark deficit costs less than the token bill difference. If you are debugging one hard concurrency bug, it does not.
MiniCPM5-2B from OpenBMB, released 7 September 2026, is the interesting release because it moves the floor rather than the ceiling. It is a dense LlamaForCausalLM with 2,516,756,480 total parameters (1.98B non-embedding), 42 layers, grouped-query attention with 16 query and 2 KV heads, a 131,072-token context window and an Apache 2.0 licence (Hugging Face card, MarkTechPost).
The coding numbers are what stand out at that size. It scores 69.1 on LiveCodeBench v6 against 56.4 for Qwen3.5-4B, and 46.4 on SWE-bench Verified against 33.6, while averaging 53.9 across 34 benchmarks versus 51.1 for the larger Qwen3.5-4B (same sources). Tool use holds up too: 97.1 on τ²-Bench Telecom and 66.6 on BFCL v4. Artificial Analysis places it at 15 on Intelligence Index v4.2, the highest of any open-weights model under 4B total parameters (Artificial Analysis).
Deployment is unfussy. It runs under vLLM, SGLang, llama.cpp, Ollama, LM Studio, MLX and FlagOS, which means a laptop or a Mac with unified memory is a viable host. A SWE-bench Verified score of 46.4 will not replace a flash-tier model on hard tickets, but it is enough for autocomplete, commit messages, small refactors and offline work on code you are not allowed to send to an API.
Pick by constraint, not by leaderboard position.
| Constraint | Pick | Key number |
|---|---|---|
| Agent drives a terminal | GLM-5.3-Flash | Terminal-Bench 2.1: 84.3 |
| High-volume batch work | DeepSeek V4 Flash | Lowest cost per token of the trio | | Very long context, multilingual repos | Qwen3.8-Flash-Next | SWE-bench Multilingual: 81.0 | | Offline, laptop or edge | MiniCPM5-2B | SWE-bench Verified: 46.4 |
| Strict OSI licence required | GLM-5.3-Flash or DeepSeek V4 Flash (MIT), MiniCPM5-2B (Apache 2.0) | Qwen uses a community licence | A sensible setup is two models, not one: MiniCPM5-2B locally for fast, private, low-stakes edits, and one flash-tier model behind an API for the tasks that need to finish autonomously. If you want a comparison against the closed-weights tier, our coding benchmark roundup covers that side, and the free Claude Code setup guide shows how to wire an open model into an agent harness.
Q: Which open-source model is best for agentic coding in 2026?
A: GLM-5.3-Flash. It leads the flash tier on Terminal-Bench 2.1 at 84.3 and DeepSWE v1.1 at 63.4, both of which measure multi-step task completion rather than single-shot code generation.
Q: Can a 2.5B model really be useful for coding?
A: For narrow tasks, yes. MiniCPM5-2B scores 69.1 on LiveCodeBench v6 and 46.4 on SWE-bench Verified, which supports autocomplete, small refactors and offline use, but not autonomous work on hard bugs.
Q: Are these benchmark numbers independently verified?
A: Partly. Artificial Analysis figures are independent; SWE-bench and LiveCodeBench results quoted from model cards are vendor-reported. The independent Terminal-Bench run scored DeepSeek V4 Flash at 79 against a vendor-reported 82.7.
Q: Which of these models has the most permissive licence?
A: GLM-5.3-Flash and DeepSeek V4 Flash ship under MIT, and MiniCPM5-2B under Apache 2.0. Qwen3.8-Flash-Next uses the Qwen Community licence, which carries additional conditions worth reading before commercial use.
Q: How much context do these models handle?
A: GLM-5.3-Flash and DeepSeek V4 Flash offer roughly 1M tokens, Qwen3.8-Flash-Next 262K natively and up to 1M with YaRN, and MiniCPM5-2B 131,072 tokens.
Q: What hardware do I need to run MiniCPM5-2B?
A: It runs through llama.cpp, Ollama, LM Studio and MLX, so a modern laptop is a realistic host at common quantisation levels.