What changed since April
Eleven weeks ago the open-weight conversation had a single, fairly settled answer. It does not any more. Four flagship releases have landed back to back since April 2026, and the "best open-weight model" title has changed hands three times in that window. Here is the shape of it before we get into the detail:
- DeepSeek V4 Pro (24 April) opened the wave and briefly topped the open-weight field on Artificial Analysis's Intelligence Index.
- Kimi K2.6 (late April) from Moonshot AI landed within days, pushing hard on long-horizon agentic coding.
- MiniMax M3 (1 June) was the first open-weight model to pair frontier coding with a genuinely usable 1M-token context and native multimodality.
- Kimi K2.7 Code (12 June) followed less than a fortnight later, a coding-specialised refresh built around the Model Context Protocol.
- GLM-5.2 (13 June) from Z.ai overtook the field on the general Intelligence Index and holds the open-weight lead as of this writing, in July 2026.
None of that list is exhaustive — Mistral, Qwen and Llama all shipped updates in the same window — but these five are the releases that actually moved the top of the leaderboard, and they are the ones India- and UK-based teams keep asking us about.
The wave, in order
DeepSeek V4 Pro arrived first and set the tone. A 1.6-trillion-parameter Mixture-of-Experts model with 49B active parameters per token, released under the plain MIT licence with a 1M-token context window, it posted 80.6% on SWE-bench Verified — the highest score any open-weight model had reached on that benchmark at the time — and roughly 90 on GPQA Diamond. For a few weeks it was simply the model to beat, and we covered why that mattered for builders deciding between open and closed in detail.
Kimi K2.6 followed almost immediately with a different emphasis: not raw benchmark maximisation but agentic reliability. Moonshot AI's release cut the hallucination rate on AA-Omniscience from 65% (in K2.5) to 39%, and post-trained the model to coordinate up to 300 parallel sub-agents across thousands of steps — a calibration story that mattered more to production teams than another point on a leaderboard.
Then came a genuine change of shape. MiniMax M3, released 1 June, was the first open-weight model to combine frontier-tier coding performance, a working 1M-token context, and native multimodal input — image, video, even desktop computer-use — in one checkpoint. It posted 59.0% on the demanding SWE-Bench Pro set, ahead of its open-weight peers at the time, which we walked through in our June coding shoot-out. The catch, which we will come back to, is its licence.
Kimi K2.7 Code landed on 12 June as a narrower, coding-focused follow-up to K2.6, built around Model Context Protocol tool-calling and claiming a 30% cut in reasoning-token usage for the same task quality. It is worth being precise about what is independently verified here and what is not — a distinction we set out at length in our Kimi K2.7 Code deep dive.
GLM-5.2 shipped a day later, on 13 June, and is the reason the leaderboard conversation has a clear current answer. Z.ai's 744B-parameter MoE (roughly 40B active) scores 51 on the Artificial Analysis Intelligence Index v4.1 — the highest of any open-weight model — under an unrestricted MIT licence. We covered the self-hosting economics of GLM-5.2 specifically in a separate piece; this article is about the wider pattern.
"Leads the leaderboard" is true of a specific index, on a specific date, often on a specific reasoning-effort setting. Several of these models publish multiple variants — reasoning-on versus off, different quantisation levels — that score meaningfully differently on the same benchmark suite. Quote the variant, not just the model name, when you cite a score internally.
How the four current flagships stack up
Put side by side, the differences that matter for a self-hosting decision are licence, context window and how much of the benchmark story is independently verified versus vendor-reported.
| Model | Released | Params (total / active) | Context window | Licence | AA Intelligence Index (v4.1) | Headline coding score |
|---|---|---|---|---|---|---|
| DeepSeek V4 Pro | 24 Apr 2026 | 1.6T / 49B | 1M tokens | MIT | 44 | SWE-bench Verified 80.6% |
| Kimi K2.7 Code | 12 Jun 2026 | 1T / 32B | 256K tokens | Modified MIT | 42 | Kimi Code Bench v2: +21.8% vs K2.6 (vendor-only) |
| GLM-5.2 | 13 Jun 2026 | 744B / 40B | 1M tokens | MIT | 51 (current leader) | SWE-bench Pro 62.1% |
| MiniMax M3 | 1 Jun 2026 | ~428B / ~23B | 1M tokens | MiniMax Community Licence | not consistently reported* | SWE-bench Pro 59.0% |
*Third-party trackers publish different Intelligence Index figures for MiniMax M3 depending on which reasoning-effort variant is scored, from the mid-40s to mid-50s. We are not citing a single number until that settles.
Licence fine print: three different open-weight deals
"Open-weight" is doing a lot of work in that sentence, and it covers real variation. DeepSeek V4 Pro and GLM-5.2 both ship under the plain MIT licence — OSI-approved, no revenue clauses, no attribution notice required, about as close to unconditional as a model licence gets. Kimi K2.7 Code (and K2.6 before it) use a Modified MIT licence, which starts from the same base but layers on Moonshot-specific terms, so read the actual licence file rather than assuming MIT-equivalence from the name. MiniMax M3 is the outlier: its MiniMax Community Licence permits commercial use but requires a visible "Built with MiniMax M3" notice, and adds a separate-authorisation requirement once a product built on it crosses a stated annual revenue threshold.
None of the four releases its full training-data pipeline or training recipe, so none of them meets a strict open-source-software definition even where the weights themselves carry an OSI-approved licence. Open-weight is the accurate term for all four models; treat "open-source model" as marketing shorthand and verify the actual licence terms before you build a commercial product on top of any of them.
Do not copy a licence summary from a blog post or a comparison table — including this one — into a compliance document. Licence terms are the one place in this story where the primary source is non-negotiable: pull the actual LICENSE file from the model's Hugging Face repository and have someone read it before you ship a paid product on top of it.
Why GLM-5.2 is on top right now
Two things separate GLM-5.2 from simply being "the newest release." First, the Intelligence Index gain is real and independently measured — Artificial Analysis runs the same evaluation suite across every model it scores, so the 51-versus-44-versus-43 gap against DeepSeek V4 Pro and Kimi K2.6 is not a vendor claim. Second, the architecture backs up the number: GLM-5.2's IndexShare sparse-attention mechanism cuts per-token compute by roughly 2.9x at the full 1M-token context length, which is the difference between a 1M window being a marketing bullet point and one you can actually afford to serve at volume. GPQA Diamond sits around 89 on Z.ai's own reporting, and SWE-bench Pro at 62.1% is a meaningful step above DeepSeek V4 Pro's equivalent score on the same harder benchmark set.
That said, "leads the aggregate index" and "wins every task" are different claims. DeepSeek V4 Pro still holds a narrow edge on raw SWE-bench Verified (80.6% versus GLM-5.2's SWE-bench Pro figures, which are not directly comparable since Pro is the harder variant). Pick the benchmark closest to your actual workload, not the model with the biggest headline number.
The Kimi K2.7 Code caveat: benchmarks nobody else has checked
Kimi K2.7 Code deserves a specific flag. Every performance number Moonshot AI has published for it — the 21.8% jump on Kimi Code Bench v2 and the 30% token-efficiency claim — comes from Moonshot's own benchmark suite. As of writing there is no independent submission to SWE-bench Verified, LiveCodeBench or GPQA Diamond, the standard public suites every other model in this piece has been measured against by a third party. Its Intelligence Index score of 42, on the one benchmark an outside party has run, does not even improve on K2.6's 43 — Moonshot's own framing is that K2.7 Code trades a small intelligence step-down for meaningfully better token efficiency, and that is a reasonable trade for a coding-specific tool, but it is a different story from "the new leader."
When a lab publishes benchmark numbers on a suite it invented itself, treat the delta (how much better than its own predecessor) as more trustworthy than the absolute score. Deltas on an internal suite are harder to game accidentally than a single headline figure, and Moonshot's own framing here — a step down on the independent index in exchange for efficiency — is the kind of honest trade-off worth taking at face value.
What the churn means for India and UK builders
The practical question is not "which model is best" — that answer has changed three times since April and will change again. The practical question is what kind of engineering decision you should actually make given that instability.
For teams in India, the calculus includes IndiaAI Mission compute access and DPDP data-residency obligations; running any of these four models on an in-region node keeps regulated data inside Indian jurisdiction while giving you a hedge against per-token US-API pricing. For UK teams, a UK South or sovereign-compute deployment carries the same data-protection logic under UK GDPR, alongside a genuine cost argument as GPU pricing in both markets keeps falling. Neither market has a structural reason to prefer one of these four models over another on regulatory grounds — the decision is almost entirely about workload shape, licence terms and how much engineering time you want to spend re-benchmarking every six weeks.
- Long-context document or compliance work — GLM-5.2, DeepSeek V4 Pro and MiniMax M3 all offer a genuinely usable 1M-token window; Kimi K2.7 Code's 256K window rules it out for this use case specifically.
- Agentic coding in a CI/CD pipeline — Kimi K2.7 Code's MCP-first design makes it a reasonable pick if you can tolerate the vendor-only benchmark caveat and validate independently on your own tasks first.
- General-purpose reasoning plus coding — GLM-5.2 currently has the strongest independently-verified case, on both the aggregate index and SWE-bench Pro.
- Tight legal review timelines — DeepSeek V4 Pro and GLM-5.2's unconditional MIT terms clear faster than MiniMax M3's revenue-threshold clause or Kimi's modified terms.
Build your inference layer around a model-agnostic interface — the same prompt templates, evaluation harness and serving wrapper should work whichever open-weight model sits behind them. Given the top spot has changed three times in eleven weeks, the teams best positioned are not the ones that bet on today's leader, but the ones that can swap the backend in an afternoon when the leaderboard moves again.
Every article here is written by a Verified Builder. Want your name on the next one?
AI Tech Connect lists AI engineers, founders and researchers across India and the UK — and the people hiring browse it to find them. Adding your profile is free.
Become a Verified Builder →The bottom line
Four major open-weight releases in eleven weeks, three changes at the top of the leaderboard, and a fourth release (Kimi K2.7 Code) that is genuinely useful but not yet independently verified against the standard suites. That pace is the story, more than any single model's score. GLM-5.2 has the strongest, most independently-verified claim to "best open-weight model" as of July 2026 — a real Intelligence Index lead, a permissive MIT licence, and an architecture that makes its 1M-token context affordable to actually serve. But the honest advice for a builder in Bengaluru or London is the same either way: pick the model that fits this quarter's workload and licence requirements, keep the serving layer portable, and expect to revisit this comparison again before the year is out.