The open-weights frontier, mid-2026

Sometime in the first half of 2026, the open-weights tier stopped being a discount rack. On the benchmarks that used to separate the frontier — graduate-level science, competition math — the best open models now sit inside the closed-frontier band. The gap that remains is narrower and more specific than the "open is a generation behind" story suggests. After taking apart Inkling, I lined up the five open models that actually matter this summer and checked the receipts.

The cast

Five models, five very different bets. Quick orientation before the numbers:

• GLM 5.2 (Zhipu AI, MIT) — the current open intelligence leader. First open model to top the Artificial Analysis Intelligence Index (51). Text-only, tuned hard for agentic coding. ~753B/40B MoE.
• Kimi K2.7 (Moonshot, modified-MIT) — the trillion-param (1T/32B) autonomy specialist: 12-hour agent runs, 300-agent swarms. K2.7 is a coding-focused point release trained with MuonClip.
• DeepSeek V4 (MIT) — the efficiency and factuality king. A 1.6T/49B "Pro" and a self-hostable 285B/13B "Flash," built around genuinely cheap 1M-token context.
• Nemotron 3 Ultra (NVIDIA, OpenMDW) — the architecture outlier: a 550B/55B hybrid Mamba-Attention MoE, pretrained natively in NVFP4, sold on throughput and token efficiency. Fully open — data and recipes too.
• Inkling (Thinking Machines, Apache 2.0) — the only true multimodal one: 975B/41B, native audio, published safety, and Tinker fine-tuning.

Who wins what

Raw intelligence → GLM 5.2

On the one genuinely independent cross-model yardstick — Artificial Analysis's Intelligence Index — the ranking is GLM 5.2 (51) > Nemotron 3 Ultra (48) > Kimi K2.6 (44) > Inkling (41). GLM is the first open model ever to lead that board, and it trails only the closed frontier.¹1 For reference, the closed leaders on the same index: Claude Opus 4.8 at 56, GPT-5.5 at 53. DeepSeek V4 is a preview and wasn't cleanly indexed at the time, but its component scores land it in the same top cluster.

Agentic coding → GLM 5.2 (with an asterisk)

This is where I had to walk back my own starting assumption. "Chinese labs own open coding" is directionally true — but only for GLM. GLM leads Terminal-Bench 2.1²2 77.9 measured independently by Artificial Analysis vs 81.0 vendor-reported — a textbook provenance gap. and SWE-bench Pro (62.1). But the benchmarks don't line up cleanly: GLM never published SWE-bench Verified, and Kimi K2.7's Verified is only 60.4 — below both Nemotron (71.9) and Inkling (77.6). So "Kimi K2.7 beats the field on coding" is just wrong on the one apples-to-apples number that exists. Kimi's real edge is autonomy — 12-hour runs and agent swarms — not raw SWE-bench.

Math & science → a near-tie at the ceiling

Effectively saturated. GLM posts AIME 2026 99.2 and GPQA Diamond 91.2; Kimi 96.4 / 90.5; DeepSeek — / 90.1; Inkling 97.1 / 87.2. The headline here isn't the winner — it's that open GPQA scores now overlap the closed band. GLM (91.2) and DeepSeek (90.1) match Claude Opus 4.8 (~92) and trail only Gemini 3.1 Pro (94.3). The "distinct lower tier" is gone on knowledge benchmarks.

Factuality & calibration → DeepSeek V4 (and a correction)

DeepSeek V4 Pro posts the best raw open factuality by a wide margin — SimpleQA-Verified 57.9 vs Inkling's 43.9 — roughly 20 points above prior open baselines. On the softer "does it hallucinate" axis (AA-Omniscience), the order is Kimi K2.6 (6.0) > GLM 5.2 (4.0) > Inkling (+2.1) > Nemotron (−1).

Multimodality & audio → Inkling, uncontested

The one row nobody else shows up for. GLM, DeepSeek, and Nemotron are text-only; Kimi K2.6 has a vision encoder (MoonViT, MMMU-Pro 79.4) but no audio. Inkling is the only natively omni-modal model of the five — text, image, audio, and video in — and the only one with a published safety suite (FORTRESS 78.0). If your product touches voice or images, this isn't a benchmark comparison; it's the only option on the list.

Efficiency, footprint & openness → DeepSeek / Nemotron

Two different kinds of efficient. DeepSeek wins cost — $0.87/M output³3 That $0.87 is a promotional rate; list price is $3.48/M — still the cheapest on the board. and genuinely cheap 1M context (~10% of V3's KV cache), and its 285B Flash is the only variant here you can realistically self-host. Nemotron wins throughput and footprint via its hybrid Mamba design and native NVFP4, and it's the most open of the five — weights, training data, and recipes all under OpenMDW. Everyone else is a ~1–2 TB-VRAM datacenter animal.

The scoreboard

Best open value in each row is bold. "v" = vendor-reported, "ind" = independent. Read it with the caveats above — this is a map, not a photo finish.

* Kimi index/benchmarks are K2.6 (K2.7 has no independent numbers). DeepSeek $0.87 was a promo; list is $3.48. "n/p" = not published (often because the model is text-only or the vendor reported a different benchmark).

How to choose

• GLM 5.2 — the default pick for the strongest all-round open model and the best open agentic coder, if you can feed a ~1TB-VRAM box. Caveat: very verbose (real token cost is high despite the low per-token price), text-only.
• DeepSeek V4 — best cost-per-token and best open factuality, with cheap 1M context. Flash (285B/13B) is the one you can actually run yourself.
• Kimi K2.7 — reach for it when the job is long-horizon autonomous agents (multi-hour runs, swarms). Trust-but-verify: no independent numbers yet.
• Nemotron 3 Ultra — best throughput and self-host footprint, fully open (data + recipes), and the most interesting architecture bet. NVIDIA's answer, and the strongest US open model after GLM.
• Inkling — the pick when you need multimodality or native audio, published safety, and cheap LoRA specialization via Tinker — not when you're chasing raw coding SOTA.

---