It's been a relentless few weeks. A new frontier-grade open-weight model landing every couple of days, each one moving the bar for what you can run on hardware you own. By the end of the month the cadence was multiple drops a week.
Qwen 3.8-Max is the headline grabber, but it's only the lead item in a stack of six. DeepSeek pushed a V4-Flash revision, Moonshot shipped the first 3T-class open model, Thinking Machines put out its first production release, and Poolside landed a US-built coding specialist. All open weight. All runnable yourself.
The backdrop hasn't moved since June. Anthropic shipped Claude Fable 5, the US Commerce Department hit it with an export-control order three days later, and every customer everywhere lost access overnight. Weights you download and serve can't be revoked. And revocation isn't the only reason to own your weights any more. The open wave is within a few points of the frontier on independent benchmarks, and for most workloads that gap isn't one you'd feel.
Qwen 3.8-Max: 2.4T and a bold claim
Alibaba previewed Qwen 3.8-Max at WAIC on 19 July and put the open-weight announcement up on 2 August. It scales to 2.4 trillion parameters on the Qwen 3.5 architecture, it's multimodal, and it's the first Max-class Qwen that'll ship open weights - Apache 2.0, weights "soon." You don't have to wait to try it: the Max-Preview is live now on Alibaba's Token Plan and on Qoder, ahead of the weight drop.
Alibaba's claim is that it's second only to Claude Fable 5. Maybe. It isn't on any independent leaderboard yet, so we're taking their word for it for now. The top independently-scored Qwen is still Qwen3.7 Max at AA Index 46, with Qwen3.6 Plus at 40 and Qwen3.5-397B-A17B at 34. If 3.8-Max lands near Fable 5 (60) that's a real generational jump. If it lands near Qwen3.7 Max, it's a bigger footprint for incremental gain. Wait for the numbers.
Qwen's announcement and OpenLM's summary are both worth a read.
DeepSeek V4-Flash got an agent refresh
DeepSeek pushed the 0731 revision of V4-Flash on 31 July, into public beta on their Responses API. The headline change is agent capability; V4-Pro is unchanged. Specs: 284B total, 13B active, mixture-of-experts, text only, 1M context, MIT licence, weights on Hugging Face.
Here's the number that matters. Artificial Analysis scored the 0731 at 50 on their Intelligence Index - a 10-point jump over April's V4 Flash, and 6 points clear of V4 Pro, on the same architecture and the same pricing. That's a revision done right. Cost per task comes out around $0.03, cache hits are $0.003 per million tokens (cheapest on the leaderboard), and it pushes 113 tokens a second.
This is the model you point high-volume tool-calling traffic at. The value play of the wave. DeepSeek's Responses API guide has the wiring.
Kimi K3: the first open 3T-class model
Moonshot released Kimi K3 on 16 July, weights on 27 July. It's the first open model to reach the 3-trillion-parameter class: 2.8T total, 104B active per token (16 of 896 experts), built on Kimi Delta Attention and Attention Residuals with a Stable LatentMoE router. Native vision, 1M-token context, MXFP4 weights and MXFP8 activations. Moonshot's framing is that the new architecture buys roughly 2.5x the intelligence per unit of compute over K2 - not just more parameters - and they shipped the attention kernels alongside the weights.
It scores 57 on the AA Index at max effort, which makes it the strongest open model on the board and seventh overall, nipping at GPT-5.6 Sol (59). Pricing is $0.30 cache-hit input, $3.00 cache-miss input, $15 output per million tokens, with a cache hit rate above 90% on coding workloads. The full-precision footprint is 1.56 TB.
Read the release post or grab the weights.
The 1-bit Kimi K3: 594 GB and runnable
A 1.56 TB model isn't much use on hardware you own, which is where Unsloth came in. Their Dynamic 1-bit GGUF quant of Kimi K3 lands at 594 GB and holds onto about 78.9% top-1 accuracy. That's a community quant, not an official Moonshot release - worth being clear about. The floor to run it is roughly 650 GB of combined RAM and VRAM, which is a lot for a desktop but reasonable for a serious workstation or a small rack. If that's still too much, there's a REAP55 variant that prunes 55% of the experts and trades more quality for a smaller footprint.
This is the model that makes "frontier on your own metal" stop being a contradiction. Moonshot's own MXFP4 recipe wants a supernode of 64 or more accelerators; the Unsloth 1-bit GGUF is the realistic self-host path, and it runs in llama.cpp and vLLM.
Unsloth's run-it-locally guide and the GGUF weights are the places to start.
Inkling Small: Mira Murati's compact MoE
Thinking Machines - Mira Murati's lab - released Inkling Small on 30 July. It's 276B total with 12B active, mixture-of-experts, multimodal (text, image, and speech in, text out), 1M context, Apache 2.0, reasoning. It scores 40 on the AA Index, costs about $0.07 per task, and runs at 120 tokens a second. For a multimodal reasoning model at that price and speed, it's a serious little workhorse.
It's the small sibling. The flagship Inkling launched on 15 July at 975B total / 41B active - AA Index 41, and Artificial Analysis called it the leading US open-weight model when it dropped, Thinking Machines' first production release. Inkling Small is the one you'd actually self-host; Inkling is the one making the case for the lab.
Model page and weights.
Laguna S 2.1: the coding specialist
Poolside's Laguna S 2.1 came out on 20 July, and it's worth being precise about what it is: a coding-specialised model, not a general frontier one. 118B total, 8B active, mixture-of-experts, text-to-text, native reasoning, 1M context, OpenMDW-1.1. Poolside's positioning is that it rivals GLM-5.2, and it sits between the lighter Laguna XS 2.1 (33B-A3B, out 2 July) and the API-only Laguna M.1.
The thing worth noting is where it's built. Poolside is US-based and trains from scratch on its own data and infrastructure. After the Fable 5 export-control episode, a US-made open-weight coding model you can run yourself has a certain obvious appeal.
Poolside's release notes have the detail.
The comparison in one table
| Model | Params (total/active) | Context | Modality | AA Index | Cost/task | Licence | Weights |
|---|---|---|---|---|---|---|---|
| Qwen 3.8-Max | 2.4T / TBD | TBD | Multimodal | TBD (vendor-claimed 2nd to Fable 5) | TBD | Apache 2.0 | "soon" |
| DeepSeek V4-Flash 0731 | 284B / 13B | 1M | Text | 50 | $0.03 | MIT | HF |
| Kimi K3 | 2.8T / 104B | 1M | Text + vision | 57 | $0.86 | Kimi K3 | HF |
| Inkling Small | 276B / 12B | 1M | Text + image + speech | 40 | $0.07 | Apache 2.0 | HF |
| Laguna S 2.1 | 118B / 8B | 1M | Text | - | - | OpenMDW-1.1 | Poolside |
| Claude Fable 5 (reference) | - | 1M | Multimodal | 60 | $3.15 | proprietary | no |
| GPT-5.6 Sol (reference) | - | 1M | Multimodal | 59 | $1.23 | proprietary | no |
Intelligence Index and cost/task are from Artificial Analysis - their composite and their weighted cost per Index task, captured late July 2026. Both shift as models are re-scored, and cost/task is workload- and cache-dependent, so treat it as a rough comparator rather than a quote. Specs (params, context, modality, licence) come from vendor releases. The two reference rows show the proprietary bar the open wave is chasing. Qwen 3.8-Max has no independent benchmark or cost numbers yet; Laguna S 2.1 has no independent score yet (vendor claims GLM-5.2 parity).
Weights on disk still aren't an API
Here's the catch, same as it was in June. A model on disk isn't a service. To run these in production you need routing across machines, auth and keys, billing, failover when a box disappears, and the OpenAI-shaped endpoint your tooling already speaks.
That's what Infersec is for. Run Infersec's Conduit application on your hardware - Macs with Apple Silicon, NVIDIA rigs, AMD accelerators - and it dials home, pulls the model config, and serves. Point an endpoint at your sources and out comes an OpenAI- and Anthropic-compatible API. HuggingFace integration fetches the weights, MCP handles tool calls server-side, and Infersec itself processes customer data in the EEA. The Conduit application runs on your hardware, wherever that is. On the public inference path there's no prompt storage - prompts live in memory for the request and are gone when it ends. The Console's chat is its own thing and keeps what you submit.
Because the endpoint speaks OpenAI, everything you'd plug into a frontier model just works - opencode, hermes, mastra, LangGraph. Swap the base URL for your Infersec endpoint, drop in a key, and you're on weights nobody can take back.
July 2026 basically closed the open-weight gap. August is for running it.
Sign up for early access, or see what we recommend on the Recommended Models page.