AI Model Tracker: Kimi K3 Weights, Astra Math Advances

Published August 2, 2026 · Research window: July 26 – August 2, 2026 (UTC) · Author context: Essa Mamdani, AI/full-stack developer

This week the frontier-model story split in two directions: a 2.8-trillion-parameter open-weight model became runnable on consumer hardware, and OpenAI previewed its next major model through ten claimed mathematics results. One is a shipping engineering artifact; the other is a research teaser with formal certificates but no external review.

The week at a glance

Kimi K3 weights are now open

Moonshot AI published the full Kimi K3 weights on July 26, 2026, a day ahead of its July 27 target. The Hugging Face model card lists 2.8 trillion total parameters, 104 billion activated parameters per token, a 1,048,576-token context window, and native vision via MoonViT-V2. It is a 93-layer MoE transformer built on Kimi Delta Attention (KDA) and Attention Residuals, with 896 experts and 16 selected per token.

This screenshot proves the weights are on an official distribution channel. The page also lists vendor-reported benchmark comparisons against Claude Fable 5, GPT-5.6 Sol, Claude Opus 4.8, GPT-5.5, and GLM-5.2.

The license is Moonshot’s own Kimi K3 License, so read the exact terms rather than assuming Apache 2.0.

WASTE: running a 2.8T model on a laptop

The engineering story of the week is sqliteai/waste, an embeddable C inference engine with no third-party runtime dependencies. WASTE keeps the shared model trunk in memory, streams only the selected experts from disk, and uses leftover RAM as a bounded expert cache. The result is that the full Kimi K3 runs on a 64 GB MacBook Pro at roughly 0.45–0.62 tokens per second.

Kimi K3 activates only about 4% of its parameters per token. WASTE reads the 11.83 MB expert chunks on demand, overlaps reads with compute, and applies 3-bit residual vector quantization to experts while keeping sensitive shared weights at 4 or 8 bits. KDA linear attention and a compressed latent KV cache also shrink memory: at 4K context the KV cache is about 0.21 GB instead of 11.25 GB.

The practical floor is 29.06 GB of RAM to open the model, plus about 982 GB of internal NVMe storage for the converted container. A 32 GB machine can open K3 but will page heavily. Conversion takes roughly 4.7 hours with three workers and needs another 1.42 TB of temporary staging storage.

The counter-intuitive cache chart

WASTE’s efficiency docs contain a warning every local-inference builder should see: giving the process more memory is not always faster.

The chart shows throughput rising from 0.57 tok/s at 3.32 GB cache to 0.63 tok/s at 17.32 GB cache, then dropping to about 0.08 tok/s at 23.32 GB and 29.32 GB. The hit rate keeps climbing, but the operating system starts paging the larger cache, so every cache hit becomes a page fault. Memory budgeting is a systems problem, not a bigger-is-better parameter.

OpenAI Astra: ten math claims with caveats

On August 1, OpenAI published “Ten advances in mathematics and theoretical computer science” and said an internal version of Astra produced the results. The claimed advances span sphere packing, binary and spherical codes, non-sofic groups, Connes’s rigidity conjecture, arithmetic circuit complexity, quantum parallel repetition, closest vector problem hardness, Ehrhart’s volume conjecture, multicolor Ramsey numbers, and extremal combinatorics.

This screenshot shows the verification artifact, not just the announcement. OpenAI released a 249-page manuscript, reasoning walkthroughs, and Lean 4 certificates, and estimates the token cost to find the solutions at roughly $2,000 at Sol API rates.

Still, classify this as “vendor research claim with formal certificates, pending peer review.” Lean certificates check the formalized arguments, but they do not validate that the theorem matches the intended statement, that the search process was sound, or that human post-processing introduced no errors.

What I would pick this week

For running a frontier open-weight model locally, Kimi K3 + WASTE is now the most credible path. It is not fast at 0.5 tok/s, but it proves the full weights are usable on hardware that fits under a desk. To experiment cheaply, start with Kimi-Linear through WASTE: 19 GB container, 1.28 GB RAM, about 10.7 tok/s.

For paid tooling, this tracker does not change last week’s read. Claude Opus 5 and Grok 4.5 are still the models to watch, as covered in the July 29 tracker. Pair model choice with AI coding agents: skills, plugins, MCP, ACP, memory and workflows and structured outputs for reliable AI APIs.

For reasoning-heavy research, OpenAI Astra is the signal to watch, but wait for peer review and API access before changing roadmaps. For leaner K3 local runs, see the earlier 1-bit Kimi K3 GGUF guide.

FAQ

Are the Kimi K3 benchmark numbers on Hugging Face independent?

No. They are vendor-reported results from Moonshot AI, even though they are published on an open model-card page. Wait for independent reproductions on the open weights before treating them as settled.

Can I run Kimi K3 on a 32 GB machine?

Technically WASTE can open the model with 29.06 GB, but the project warns that a 32 GB machine will page heavily and be slow. A 64 GB machine with fast internal NVMe is the practical minimum.

Is WASTE production-ready?

It is an early open-source engine with measured docs and validation against a PyTorch reference. It is promising for research and local experimentation, but treat it like any new inference stack: test on your workload before shipping to users.

Do the OpenAI Astra math proofs mean the model is superhuman?

They mean the model generated arguments that humans then prepared and formalized in Lean. That is a genuine capability signal, but peer review is needed to confirm the mathematical claims and to separate model contribution from human curation.

Why does more cache hurt throughput in WASTE?

Because the operating system starts paging the larger cache. A cache hit that triggers a page fault is slower than a cache miss that streams directly from NVMe. Memory budgeting must leave enough room for the OS and other resident pages.

Methodology and scope

I searched vendor pages, GitHub repositories, and project docs published between July 26 and August 2, 2026. Kimi K3 details come from the Hugging Face model card and Moonshot AI homepage. WASTE measurements come from the sqliteai/waste repository, README, and docs/EFFICIENCY.md. OpenAI Astra claims come from the official announcement and the linked openai/ten-proofs repository.

Vendor-reported benchmarks are kept separate from WASTE engineering measurements, which are project-reported but reproducible by anyone with the hardware. Astra results are research claims with formal certificates, pending peer review. I did not merge these evidence classes into a single leaderboard.

Source notes

• Moonshot AI: Kimi K3 on Hugging Face
• Moonshot AI homepage
• sqliteai/waste: WASTE inference engine
• sqliteai/waste: docs/EFFICIENCY.md
• OpenAI: Ten advances in mathematics and theoretical computer science
• OpenAI: ten-proofs GitHub repository
• Hacker News front page, August 1, 2026 (discovery source for WASTE)

Author context

I write about AI systems and full-stack engineering from the point where product behavior meets production reality. My bias is toward explicit contracts, narrow tool surfaces, and enough telemetry to explain failures without exporting private data everywhere.

If you want help turning this kind of tracker into a production eval or local-inference stack, I’m available through /hire.