17. Serving paths reach merged upstream code (2026-09-13 UTC): vLLM merged DeepSelect sparse-indexer TopK and Engram DP-sharding/async CPU-offload/shared-memory tables for DeepSeek V4.1-Flash; SGLang merged a one-DGX-Spark / GB10 path for nvidia/Qwen3.8-Flash-Next-NVFP4 (200-question GSM8K 97.5%/97.0%, about 93/90 tok/s). Both remain on main and untagged, so this is continued open-weight serving adoption, not a TB/SWE reproduction on the weights (ev-20260910-01 / ev-20260826-04 updates).
Can open-weight models support coding agent workflows?
The archive tracks progress in open-weight models and serving tools. Selection still requires tests on your own tasks, hardware and toolchain; vendor benchmarks do not establish local workflow performance.
As of 2026-09-28 · Analysis based on archive evidence
Scope of the answer
For teams able to deploy and evaluate models. Licensing, hardware budgets and independent reproduction remain prerequisites.
Open questions
- Can the public weights reproduce results at comparable task and budget settings?
- Does deployment cost and maintenance compare favorably with hosted APIs?
Supporting material
18. vLLM 0.30.0 tagged stable (2026-09-22, ev-20260922-01): previously merged serving paths for DeepSeek V4.1, GLM-5.3, Qwen3.8 and other open-weight models enter a stable release alongside a resident GPU weight cache, HiSparse host KV tier and broad security hardening. Serving-adoption evidence strengthens, but this is still not a third-party Terminal-Bench or SWE reproduction on public weights.
Hy4's 30-day adoption review found 22,788 monthly downloads, while the full model and reproducible third-party evaluation remain absent.
More evidence · 12
GLM-5.3 weights landed: HF zai-org/GLM-5.3 (FP8, 153 files) + GLM-5.3-BF16 created 2026-08-25, 753B total params, ungated; the custom glm-5.3 license is MIT-style with a single security-review clause for >$10B-revenue MaaS businesses; day-0 vLLM/SGLang/TokenSpeed/KTransformers/Unsloth/vLLM-Ascend support (primary source, ev-20260814-02 update)
Independent eval: Artificial Analysis Intelligence Index — GLM-5.3 (max) at 60 (on par with Kimi K3, +7 over GLM-5.2); GLM-5.3-Flash at 57 (added 8/26); GLM-5.3-Flash open-sourced under plain MIT the same day (320B-A18B, new base, hybrid sparse-linear attention — same-org reinforcing signal, ev-20260825-01)
Qwen3.8-Flash-Next open-weight architecture preview (Qwen, dated 2026-08-26; HF repo created 8/24; coverage-gap recovery, ev-20260826-04): 125B/6B-activated + 51B n-gram embedding + 4B MTP (~180B), Gated DeltaNet linear attention + Qwen Sparse Attention every 4th layer, 262k native context (1M via YaRN), multimodal; custom qwen-community-1.0 license; 52k downloads and 4.2k likes in 4 days — a same-direction open-weight reinforcement from a different org, landing hybrid linear attention in the same week as GLM-5.3-Flash (cross-org technical convergence)
GLM-5.3-Flash download velocity (verified 2026-08-29): 189,793 downloads (~4 days after listing; 30d window = all-time) + 1,557 likes; GLM-5.3 FP8 at 8,804 (753B artifact); vLLM recipe page live, NVFP4 quantized-serving discussions appearing (ev-20260825-01 update)
Tencent Hy4 preview open weights (Tencent, announced 2026-08-28, HF repos created 8/27; coverage-gap recovery, ev-20260828-02): 770B total / 49B active MoE, context beyond 1M, Apache 2.0 (per HF tags), standard + FP8 weights with a day-0 vLLM recipe — the third Chinese lab to open-source major weights in one week, and the closest candidate yet for the same-tier-different-org criterion (b); however the only eval is a vendor-run blind test (163 experts / 203 tasks, 2.99 vs GLM-5.3 2.92, Kimi K3 2.94), there are no public benchmark scores, and it is a preview; ~1.4k downloads in 2 days
Download velocity confirmed (verified 2026-08-31): the HF counters frozen for two days all rolled — GLM-5.3-Flash at 346,516 downloads 6 days after listing (+83% vs the 8/29 sample) + 1,711 likes; Qwen3.8-Flash-Next at 121,976 (+133%) + 4,384 likes; the GLM-5.3 FP8 flagship repo 8,804 -> 50,116 (~5.7x — the 753B weights now pulled at scale) — cross-org (Z.ai / Qwen), cross-size (320B / 180B / 753B) growth simultaneously, confirming criterion (c) sustained download velocity; Hy4 preview modest (2,123 main + 1,469 FP8) (ev-20260825-01 / ev-20260826-04 / ev-20260814-02 / ev-20260828-02 updates)
DeepSeek V4-Flash-Vision-Exp open weights (2026-08-31, MIT, ev-20260831-02): the V4 family's first multimodal model, holding Flash-tier text-agent benchmarks (TB 2.1 83.9 vs Opus-4.8 85.0) while jumping multimodal agent benchmarks (ApexBench 36.5 vs 26.2 for the text-only sibling that ignores multimodal input) — the open-weight wave extending from text coding agents into multimodal agents; ~17.9k downloads + 452 likes within ~1 day of weights landing
Velocity continuation + community serving tooling (verified 2026-09-02): GLM-5.3-Flash 441,348 (+27% vs 8/31), Qwen3.8-Flash-Next 207,941 (+70%, plus 130,451 FP8), GLM-5.3 FP8 94,403 (+88%), Hy4 3,516 (+66%) — a second consecutive growth cycle; unsloth GGUF derivatives (Flash 63.7k / Flash-Next 431k); slotstream (Show HN, 147 pts) streams the 104GB 4-bit Flash-Next from SSD on a 48GB M5 Pro at ~12 tok/s warm decode (32GB peak, Ollama/OpenAI-compatible APIs); serving research catching up with hybrid architectures (Tail-Replay/DASC: prefix caching and state compression for hybrid linear attention, ev-20260901-04) (ev-20260825-01 / ev-20260826-04 / ev-20260814-02 / ev-20260828-02 updates)
Third consecutive velocity cycle + serving ecosystem maturing (verified 2026-09-05): GLM-5.3 FP8 94,403 -> 303,534 (+222% — the 753B weights pulled at scale), GLM-5.3-Flash 441,348 -> 654,957 (+48%), Qwen3.8-Flash-Next 207,941 -> 351,374 (+69%, plus FP8 186,676), Hy4 preview 3,516 -> 5,684 (+62%), DeepSeek V4-Flash-Vision-Exp 17.9k -> 133,024 (steepest relative pull-rate of the window); the unsloth Flash-Next GGUF derivative reached 702,251, exceeding Qwen's own main repo; NVIDIA published nvidia/Qwen3.8-Flash-Next-NVFP4 (9/2) and AMD published GLM-5.3 Quark quantizations — silicon vendors now ship official quantizations for this wave; serving research keeps catching up: an NVFP4 W4A4 whole-model recipe for hybrids (ev-20260905-03) and the random-KV-eviction negative result (ev-20260905-04); GLM-5.3-Flash discussion #38 carries the first third-party FP8 deployment throughput report (2x L40S) (ev-20260814-02 / ev-20260825-01 / ev-20260826-04 / ev-20260828-02 / ev-20260831-02 updates)
Fourth consecutive velocity cycle (verified 2026-09-08): GLM-5.3 FP8 303,534 -> 442,064 (+46%), GLM-5.3-Flash 654,957 -> 784,005 (+20%), Qwen3.8-Flash-Next 351,374 -> 474,693 (+35%, plus FP8 224,875), Hy4 preview 5,684 -> 6,705 (+18%), DeepSeek V4-Flash-Vision-Exp 133,024 -> 251,611 (+89%, steepest relative pull-rate for the second consecutive check); unsloth Flash-Next GGUF 702,251 -> 868,243 and NVIDIA's NVFP4 derivative at 18,068 — the 753B flagship and every derivative line still growing (ev-20260814-02 / ev-20260825-01 / ev-20260826-04 / ev-20260828-02 / ev-20260831-02 updates)
15. Criterion (b) met + fifth consecutive velocity cycle (checked 2026-09-11): DeepSeek V4.1 Flash open-weights (9/10, MIT, 763B total/552B backbone, TB 2.1 90.6 above every listed proprietary model, DeepSWE 74.2; 75.7k downloads in ~1.5 days) — a same-tier open-weight release from another organization at GLM-5.3's exact scale class (763B vs 753B); counters fifth cycle: GLM-5.3 FP8 597,626 (+35%), Flash 1,173,520 (+50%), Flash-Next 586,040 + FP8 349,091, Hy4 8,219, V4-Flash-Vision 443,954; serving diagnostics follow the hybrid architectures (SinkProbe, ev-20260910-14); safety spillover made concrete — a single-direction ablation strips refusal from the shipped block-FP8 GLM-5.3-Flash weights (ev-20260911-06)
16. Sixth consecutive velocity cycle + the nearest approach yet to a weights-side reproduction (checked 2026-09-13T00:03Z): GLM-5.3 FP8 635,504 (+6.3%), GLM-5.3-Flash 1,333,574 (+13.6%), Flash-Next 604,992 + FP8 369,963 (+3.2%/+6.0%), Hy4 8,936 (+8.7%), V4-Flash-Vision 484,422 (+9.1%), DeepSeek V4.1-Flash 140,636 (steepest, ~2.5 days after release); unsloth Flash-Next GGUF 1,160,057 and the NVIDIA NVFP4 derivative 89,924 (~5x in 4 days) - every line still growing, though the per-day slope on the Qwen/GLM Flash lines is below the prior 3-day cycles. Criterion (a) sees its nearest approach yet: HF discussion #50 (9/12) runs the public weights on a single DGX Spark (GB10, 128GB unified memory) with 2-bit routed experts, HumanEval 97.0% above the z.ai API's 95.1% (identical prompts, temperature 0; reproduction artifacts public) - but HumanEval is not Terminal-Bench/SWE, so the criterion remains unmet. Adjacent (vendor-reported, not counted): Z.ai's engineering post (9/12) discloses that the ox-alpha anonymous-testing phase ran entirely on a Chinese-AI-chip cluster (custom SGLang engine, EPD disaggregation, 3x end-to-end improvement, cost comparable to mainstream NVIDIA GPUs) (ev-20260825-01 update)
Counter-evidence
No direct counter-evidence is recorded. That alone does not establish the conclusion. Independent reproduction and actual deployments remain the next checks.