July 2026 Open-Weight Wave: K3, Inkling and the New Gap

The open-weight model wave of July 2026 compressed five significant moves into a single two-week window: Moonshot’s Kimi K3 launched at 2.8 trillion parameters, Thinking Machines shipped Inkling under Apache 2.0, MiniMax’s 2.7T “M3 Pro” plan surfaced in press reports, Mistral opened early access to a new sparse MoE family, and DeepSeek’s official V4 release sat on a mid-July schedule. No single vendor’s coverage connects them — this tracker does.

What’s at stake is bigger than any one launch. For most of 2026 the open-weight race read as a Chinese-lab phenomenon, and the gap to closed frontier models was described in vague, multi-generation terms. This window changes both framings: a credible US-built open model now exists at frontier-adjacent scale, and the measurable gap on the hardest coding benchmarks has narrowed to single digits for the best open entrant.

This is the news read of the week, not a forecast — our Q3 open-weight forecast covers the scenario work this roundup feeds, and the first-half open-weight recap covers January through June. Below: what shipped, what was only announced, a scorecard no one else has assembled, and what buyers should actually do this week.

Between July 4 and July 17, the open-weight landscape moved five times. Two were real ships — one Chinese, one American. Two were forward announcements with no product attached. One was a scheduled graduation from preview to official that, as of publication, has not been confirmed as landed. Most coverage treats each as an isolated story; read together, they describe a market where open releases now arrive in overlapping waves rather than isolated events.

The distribution matters as much as the count. Two ships and three announcements is a very different market signal than five ships — and the two ships bracket the strategic spectrum. K3 is the maximum-scale, benchmark-forward play; Inkling explicitly positions itself as customizable rather than strongest. Both strategies now have flagship-scale open hardware behind them, on two continents.

Moonshot’s Kimi K3 is the wave’s anchor event, and we covered the launch in depth in K3’s own July 17 release post — this section only places it in the wave’s context. The essentials: 2.8 trillion total parameters in a Stable LatentMoE configuration with 16 of 896 experts active, a 1M-token context window, native multimodal input, and API pricing of $3.00 per million input tokens on a cache miss ($0.30 on a cache hit) and $15.00 per million output — flat across the full context window, with no length-based tiering.

The benchmark story is where the “new gap” thesis gets its numbers. On Moonshot’s own fourteen-benchmark launch chart — vendor-reported, every model at max or xhigh thinking effort — K3 posts the top score of any model on five rows, Claude Fable 5 takes seven, and GPT-5.6 Sol takes two. The full breakdown lives in our K3 vs Fable 5 comparison; the selected rows below show the shape.

Two honesty notes. First, these are vendor-reported figures from Moonshot’s launch materials, not independent evaluations — treat the direction as informative and the decimals as marketing. Second, Moonshot itself lists a “noticeable gap in user experience” versus Fable 5 and GPT-5.6 Sol among K3’s stated limitations — an unusually candid caveat for a launch, and a reminder that benchmark rows and day-to-day usability are different axes.

Two days before K3, Thinking Machines Lab — the startup Mira Murati founded after leaving OpenAI, which raised a record $2 billion seed at a $12 billion valuation before shipping anything — released its first foundational model. Inkling is a 975B-total, 41B-active MoE, natively multimodal on input (text, image, audio, video) with text-only output including code and structured data, trained on 45 trillion tokens, released under Apache 2.0 with weights live on Hugging Face at launch. A lighter 276B Inkling-Small was previewed alongside, its weights to follow after testing.

The positioning is the interesting part. Thinking Machines says plainly that Inkling “is not the strongest overall model available today, open or closed” — a deliberate customizable-over-smartest bet. The differentiators it claims instead: a “controllable thinking effort” dial that lets developers programmatically scale reasoning budget from 0.2 to 0.99, trading token spend for depth; a stated design goal of answering directly on topics that may be subject to censorship; and Tinker, the fine-tuning and customization API that is the actual revenue vehicle — Inkling itself is not directly monetized, and Bridgewater Associates is already named as a Tinker customer for financial tasks.

On efficiency, Thinking Machines says Inkling uses roughly one-third the tokens of Nvidia’s Nemotron 3 Ultra for equivalent coding performance — a vendor claim relayed by a single outlet, not an independent benchmark, so treat it as a thesis to test rather than a fact to cite. The lab’s own compiled comparisons (as relayed by VentureBeat) put Inkling at 77.6% on SWE-bench Verified, 97.1% on AIME 2026, 63.8% on Terminal Bench 2.1, and 74.1% on MCP Atlas — beating Nemotron 3 Ultra on most rows while trailing DeepSeek V4 Pro on SWE-bench Verified (80.6%) and trailing the closed frontier everywhere hard.

Strategically, Inkling matters beyond its scores. It was trained on Nvidia’s latest AI infrastructure — Nvidia is an investor — and used other open-weight models, including Moonshot’s Kimi K2.5, to generate early post-training data. And per Axios’s reporting, Murati’s openness is case-by-case rather than a blanket commitment: she was at OpenAI in 2019 when it withheld full GPT-2 over misuse fears, and later Thinking Machines models may not all be open. The US re-entry is real; its permanence is not guaranteed.

The other three moves in the window are commitments of varying firmness, and the difference between them is worth being precise about.

MiniMax “M3 Pro” — a reported plan

Per a report in The Information relayed by The Decoder on July 8, MiniMax is planning a 2.7-trillion-parameter open-weight model, internally called “M3 Pro” (the name may change), targeted for later in 2026. If it ships at that scale, it would be positioned as the largest Chinese open-weight model on the market at release. Every part of that sentence is press-reported through a single chain with no MiniMax primary-source confirmation — a reported plan, not a roadmap commitment. For scale context: the current MiniMax M3 shipped May 31 with its total parameter count undisclosed, so the 2.7T figure cannot even be compared against an official baseline. The same report notes Chinese regulators may be considering tighter controls on future model releases — a live watch-item for any late-2026 timing.

Mistral’s sparse MoE — a teaser with early access

Mistral CEO Arthur Mensch confirmed a new open-weight model family in a LinkedIn essay and X posts around July 4, with early access opening in July for key partners in research, government, and industry. He separately described the family as “fat but sparse” — MoE phrasing pointing at a much larger total parameter count than Mistral Large 3 (675B total / 41B active, Apache 2.0, December 2025) while keeping active-parameter compute comparable. No parameter count, benchmark, license terms, or exact ship date has been confirmed.

One disambiguation this coverage cycle genuinely needs: a hoax called “Le Chaton Fat,” claiming a 24T–30T-parameter Mistral model, circulated June 14–16 and was debunked — Mensch himself joked publicly that the name would not be used. The July 4 announcement is a different, genuine event. Do not carry the meme’s parameter figures into any analysis of the real model; nothing about its size is confirmed. The regulatory backdrop is also part of Mistral’s framing: EU AI Act enforcement powers — information requests, model access, recall — activate August 2, 2026, and Mistral has signed the EU’s General Purpose AI Code of Practice alongside roughly two dozen providers. With ARR above $400 million in early 2026 and reported fundraise talks at a valuation above $23 billion, Mistral is framing this release around European strategic autonomy.

DeepSeek V4 — scheduled, and only scheduled

DeepSeek’s V4 shipped as a Preview Release on April 24 — V4-Pro at 1.6T total / 49B active and V4-Flash at 284B / 13B, with 1M context now default across official DeepSeek services. On June 30, DeepSeek announced the official, non-preview V4 release is scheduled for mid-July, together with its first time-of-day API surcharge: prices double during daily peak windows (9am–12pm and 2pm–6pm), with off-peak pricing unchanged. As of July 17, no DeepSeek primary source or independent outlet we reviewed confirms the official launch has actually landed — so this tracker counts it as a scheduled move, not a completed one. The harder deadline is unambiguous, though: the legacy deepseek-chat and deepseek-reasoner endpoints retire — fully inaccessible — after July 24, 2026, 15:59 UTC. Both currently route to V4-Flash in the interim.

No single published source lines up the current open-weight field in one table — each vendor’s press covers only its own launch. So we built it. The table groups by lab home market because that is the wave’s real story: the field is now US-plus-China (plus a pending European entry), not China-only. Alongside the July arrivals, the incumbents matter for context — Z.ai’s GLM-5.2, Tencent’s Hy3 release, and Nvidia’s Nemotron 3 Ultra all shipped within the past six weeks.

Demand-side data backs the supply-side picture. On OpenRouter’s weekly leaderboard snapshot from July 8, Chinese open-weight labs dominate token volume — DeepSeek alone at roughly 5.4 trillion tokens per week, with Xiaomi/MiMo, MiniMax, Z.ai, and Tencent all in the top ranks and Anthropic the only Western lab holding multiple top-10 slots. The steepest weekly growth: Tencent at +37% on the Hy3 launch, DeepSeek +26%, Z.ai +25%. Open-weight supply is being pulled by real workload demand, not just released into a void.

“Open models are one generation behind” is usually asserted, not shown. The tracker below makes it falsifiable: for each open-weight model, the point spread to the best closed model on the same vendor chart. The pattern that emerges is a split. The frontier-adjacent open entrant (K3) is now inside single digits on the hardest software-engineering benchmarks — and inside a single point on some. The generalist open entrants (Inkling) still trail by double digits on hard coding, by design and by their own admission.

The interpretation we would defend: the “gap” is no longer one number. For maximum-effort agentic coding, the best open model now sits roughly one model generation behind the best closed one — close enough that price, weights access, and deployment control can outweigh the score difference for specific workloads. For generalist, customization-first open models, the gap on hard coding remains wide and acknowledged. Buyers should stop asking whether open models have caught up and start asking which open model category their workload actually needs.

One name hangs over this whole window without being part of it. Gemini 3.5 Pro has not launched as of July 17, 2026: Google has published no model card, no pricing page, and no API listing for it. What is confirmed is thin — the model exists and runs internally, and it slipped from a June target. Sundar Pichai said on stage at Google I/O on May 19, “Give us until next month to get it to you,” and that month passed. Third-party reporting citing unnamed sources claims Google scrapped a near-complete base model and restarted pre-training over failures in complex SVG generation and recursive tool-calling stability, with a rumored July 17 target — none of it confirmed by Google.

The rumored specs — a 2-million-token context window, a “Deep Think Reasoning Layer,” autonomous multi-file workflow capability — are leak-grade claims from unnamed sources, and we deliberately exclude them from every table in this post. For calibration, the confirmed contrast: Gemini 3.5 Flash has shipped, and Google’s own launch figures put it ahead of Gemini 3.1 Pro on Terminal-Bench 2.1 (76.2% vs 70.3%) and MCP Atlas (83.6% vs 78.2%). Whenever 3.5 Pro does land, it resets the closed-frontier reference line that every gap number in this post is measured against — which is the one concrete reason to track the rumor at all.

A wave like this rewards a small amount of immediate housekeeping and punishes both overreaction and inaction. Four moves, in priority order:

Two of those actions have dedicated companions on this blog: the second-source playbook walks through standing up an open-weight fallback in practice, and the K3 July 27 readiness checklist covers what to prepare — hosting scale, license review, evaluation harness — before the weights drop. For teams that want the comparative evaluation run for them, our AI transformation engagements start with exactly this open-vs-closed routing analysis on your own workloads.