TAI #214: Kimi K3 Brings Open Weight Closer to the Frontier

Also, Fable disproves an 87-year-old math conjecture, Qwen3.8-Max, Thinking Machine's Inking & more!

What happened this week in AI by Louie

After a wave of strong new closed models last week, Moonshot AI’s Kimi K3 is the closest a model headed for open release has come to the frontier since DeepSeek R1 in early 2025. This week also produced more evidence that LLM agents’ capabilities are turning recursive, with agents optimizing the infrastructure that trains their successors and producing publishable mathematics.

Moonshot says K3 combines 2.8 trillion total parameters, a one-million-token context window, native vision, and a Mixture-of-Experts design that routes each token through 16 of 896 experts. Kimi Delta Attention handles long contexts efficiently, while Attention Residuals lets each layer retrieve selected earlier-layer outputs, which Moonshot credits with roughly 25% better training efficiency for under 2% extra cost, part of a claimed 2.5x scaling-efficiency gain over K2. Quantization-aware training runs in four-bit MXFP4, and Moonshot recommends at least 64 accelerators for deployment. K3 is live in Kimi’s products and API, with weights promised by July 27.

Independent testing shows a consistent shape: close to the leaders overall, ahead in key agentic lanes. On Artificial Analysis’s Intelligence Index as of July 21, K3 scores 57.11, behind Claude Fable 5 at 59.86 and GPT-5.6 Sol at 58.89 but above Opus 4.8 at 55.69 and GLM-5.2 at 51.09. It sits within 0.25 points of Fable on the Coding Index at 76.24, leads AutomationBench-AA at 52.71%, and ranks first on Arena’s preliminary frontend leaderboard at 1,679 Elo. Its weakest lane is knowledge reliability, where its AA-Omniscience score of 18 trails Fable’s 40 and Opus 4.8’s 27, and Moonshot itself concedes K3 trails Fable and Sol overall, including on user experience.

Moonshot’s own coding table is strong and uneven. K3 narrowly trails Sol on Terminal-Bench 2.1, edges it on ProgramBench, and leads the SWE Marathon chart.

K3 also topped our internal writing benchmark, an early, narrow test of our editorial voice. K3 Thinking ranked first at 2,840 Elo, ahead of Fable 5 at 2,760, after K2.6 placed ninth at 2,214, and it costs us about $0.25 per script, roughly a fifth of Fable’s price. While it has now fallen behind Fable after an upgrade to our evaluations, the jump from K2.6 shows what a generational leap looks like.

The economics are less flattering than the leaderboards. While its per-token costs are below OpenAI and Anthropic, they are 3x higher than Kimi 2.6. This is a huge model and will also not be cheap to inference yourself once weights are available. On a task level, Artificial Analysis measures K3 at $0.95 per weighted Intelligence Index task, only 8% below Sol’s $1.04, despite token prices 40% to 50% lower, because K3 emits 54% more output tokens. Its first-party endpoint also generates slowly, at 39.5 output tokens per second and 67.5 seconds for a standardized 500-answer-token request. However, it streams its first reasoning token in about four seconds while Fable and Sol’s maximum-effort endpoints stay silent for around two minutes.

Source: Kimi. Total parameters of each company’s flagship model

But the demand for K3 has already overwhelmed the supply side. Kimi paused new subscriptions on July 19 after 48 hours of demand pushed its GPUs near capacity. To protect existing customers, they have paused it for new subscribers and confirmed that inference supply already binds adoption. Open weights do not repeal serving costs either: the raw files occupy roughly 1.5 terabytes in MXFP4 before runtime memory, and the key-value cache, and the footprint is nearly triple K2.6’s one trillion parameters, though active compute per token likely grows far less.

AI stocks sold off on Friday in what some investors called a smaller DeepSeek moment, with K3 one driver among several. The reaction reflects how quickly the frontier is tightening.

There is intense discussion about Chinese labs distilling US frontier models. My view is simple: of course they are, and so is everyone else. Frontier intelligence can enter a model through many routes. Labs can collect direct outputs, generate synthetic examples, or buy data from providers that used Fable to brainstorm reinforcement-learning tasks, benchmarks, rubrics, and adversarial tests. Elon Musk has even acknowledged xAI’s use of Claude-generated data. One way or another, it seems very likely that some of Fable’s intelligence helped bootstrap the major leaps from GLM-5.2, Grok, Meta, and K3 over the past few weeks.

A successful five-to-ten-hour Fable agent run can produce code, tests, tool traces, critiques, and a verified final artifact. That is exceptionally valuable reinforcement-learning data for extending agentic capabilities. Labs can gain enormous value without private reasoning traces or logits. The AI race is a commercial and geopolitical contest led by CEOs who believe their company should lead it. They will take easy capability gains when the consequences appear limited.

K3 remains a huge achievement. Using this data effectively requires a strong base model, sophisticated reinforcement-learning infrastructure, reliable environments, good graders, and extensive original research. These labs are also creating novel data, training tasks, architectures, and systems. K3 is far more than a distillation of Fable and already beats it in some areas. Moonshot says K3 autonomously optimized an Attention Residuals kernel, built a MiniTriton compiler, and completed a 48-hour chip-design experiment.

Chinese labs are currently producing more visible architecture-level innovation and appear to have at least comparable aggregate research talent. They need that strength to compensate for less capital, less compute, and weaker access to the advanced semiconductor ecosystem. US labs benefit from hyperscaler balance sheets, leading chips, high-bandwidth memory, and easier access throughout the deep ultraviolet and extreme ultraviolet lithography supply chain.

The largest US labs have also locked up enormous GPU capacity through long-term contracts. That is a major moat. They can win by maintaining the strongest frontier models. If base-model intelligence becomes commoditized, they can still win by controlling much of the inference capacity needed to deploy it. Codex and Claude Code create another moat as they become embedded in enterprises, accumulate workflow context and evaluations, and let the model and agent harness improve together.

Which brings us to recursion at a different level: many types of recursive AI improvement are now operational. First, coding agents accelerate the engineers building the next model across data pipelines, kernels, evaluations, and deployment; OpenAI says early GPT-5.3 Codex versions helped debug their own training and deployment, and K3’s kernel run is the same loop inside Moonshot. Second, stronger models manufacture the training pressure for their successors: tasks, synthetic data, verifiers, judges, and adversarial examples. Teams can spend orders of magnitude more inference through maximum reasoning, sampling, and agent fleets, then compress the best verified results into a cheaper baseline. OpenAI’s GPT-Red, which generated the attacks used to harden GPT-5.6, is that loop in production.

The third level, AI designing its successor’s architecture, remains unsolved but no longer looks distant. Thousands of agents could test candidate designs on small clusters against loss curves and benchmarks before the winners scale, if proxy experiments can be made to transfer. AlphaEvolve’s measured improvements to Google’s data-center scheduling, Tensor Processing Unit circuits, and matrix kernels used in AI training show part of the path, and K3’s experiments point the same way.

The multi-agent method is also starting to produce real mathematics. On Monday, Anthropic mathematician Levent Alpoge posted an explicit Fable-assisted counterexample to the Jacobian conjecture, open since 1939: a polynomial map with constant nonzero Jacobian determinant that sends three distinct points to the same output. Independent symbolic checks confirm the arithmetic. It disproves the conjecture in dimension three and above, while the two-dimensional case stays open and a formal paper is pending. The announcement was almost casual: Alpoge thanked his friend Akhil Mathew for asking the question and Fable for working through the World Cup final. Within a day, a separate Codex run produced an essentially equivalent construction under a published prompt demanding diverse multi-agent search and adversarial checking. Current models excel at intelligence-guided search under exact verifiers and at connecting distant specialized fields, and those strengths sit close to what machine learning research itself requires.

I would not count Anthropic out either. Mythos training was reportedly complete by late March, giving nearly four months for evaluation, post-training, and product work, so Claude could leap ahead again. The durable lesson of 2026 is that several labs can now make major capability jumps within weeks, and each generation supplies the tools and data that make the next arrive faster.

---

---

Why should you care?

Near-frontier open weights compress model API margins while expanding total compute demand, because cheaper intelligence makes longer tasks, more retries, and larger agent fleets economical. The winners shift toward scarce capacity, efficient serving, and distribution: hyperscalers and neoclouds can serve every model. At the same time, US AI labs defend their reserved inference capacity, economics through their coding harnesses, and enterprise governance. I would watch completed-task cost, output-token efficiency, and accelerator occupancy over leaderboard rank or parameter count. For builders, K3 is a credible option for workflows that previously required the most restricted US systems.

For researchers, the recursion carries the longest fuse, and its fuel is cheap, exact verification. Mathematics, code, kernels, and chip layouts reject bad candidates almost for free, which is why the Jacobian result landed there first, and why environments, simulators, formal checks, and high-quality evaluators are becoming core research infrastructure rather than tooling afterthoughts.

Governments face the hardest version of the trade-off, because capabilities that ship as weights run beyond monitored endpoints and cannot be recalled. If the weights land as promised, any restriction drafted after July 27 will be regulating files the world already holds.

— Louie Peters — Towards AI Co-founder and CEO

Hottest News

1. Moonshot Launched Kimi 3 and Paused it After Subscription Demand Hit GPU Limits

Moonshot AI launched Kimi K3, a 2.8 trillion-parameter MoE model with a 1M-token context window, built for long-horizon coding, reasoning, and agentic tasks. It is the largest open-weight model released to date, surpassing DeepSeek V4 Pro’s 1.6T parameters. It uses Stable LatentMoE, effectively activating 16 of 896 experts. The model supports text, image, and video inputs and is available through Kimi.com, Kimi Work, Kimi Code, and the Kimi API. API pricing is $3/$15 per million input/output tokens. On independent benchmarks, Kimi K3 placed third on the Artificial Analysis Intelligence Index behind Fable 5 and GPT-5.6 Sol and topped several coding and agent benchmarks.

Within 48 hours of launch, user demand exceeded Moonshot’s GPU capacity, forcing the company to pause new consumer subscriptions on July 19 to protect service quality for existing subscribers. Moonshot said it would reopen new spots in batches as capacity is added and plans to split memberships into two tiers: one for Kimi Web, App, and Work, and a separate Kimi Code membership for coding workflows. Full open-weight model releases are planned for July 27 on Hugging Face.

2. Anthropic Makes Claude Fable 5 a Standard Max and Team Premium Feature

Anthropic announced that beginning July 20, Claude Fable 5 will be permanently included in all Max and Team Premium plans at 50% of usage limits. This ends weeks of uncertainty over Fable 5’s availability, which had been extended in stages following the model’s redeployment on July 1 after the Commerce Department export control suspension. The eligible plans are Max 5x ($110/month), Max 20x ($220/month), and Team Premium ($100/month). Pro and Team Standard subscribers will not receive Fable 5 as part of their base plans but retain access through usage credits at standard API rates ($10/$50 per million input/output tokens). Pro and Team Standard users will receive a one-time $100 credit.

3. Alibaba’s Qwen Team Launches Qwen3.8-Max-Preview and Says Full Qwen3.8 Will Go Open-Weight

Alibaba’s Qwen team announced Qwen3.8, a 2.4 trillion-parameter multimodal model that the team describes as “one of the most powerful models available today, second only to Fable 5.” Qwen3.8-Max-Preview is available now on Alibaba’s Token Plan, Qoder, and QoderWork. The model accepts text, image, and video inputs and is compatible with both the OpenAI and Anthropic API protocols, allowing existing coding agents to connect without rebuilding their harnesses. Alibaba has committed to releasing Qwen3.8 with open weights. No benchmark results, model card, or active-parameter count have been disclosed. The “second only to Fable 5” claim is Alibaba’s own positioning, not an independently verified result. Token Plan pricing starts at approximately $6/month for the Lite tier; standalone per-token API pricing has not been announced. The launch comes three days after Moonshot AI’s Kimi K3, putting two trillion-scale Chinese models in the market within the same week.

4. SpaceXAI Open-Sourced Grok Build

SpaceXAI open-sourced Grok Build’s coding-agent harness and terminal interface under Apache 2.0. The release includes the full Rust agent loop, TUI, CLI shell, and tool-call dispatch layer, allowing developers to inspect exactly how the agent assembles context, calls models, and executes tools. The code is published as a squashed monorepo mirror with no prior commit history. Grok Build supports autonomous coding tasks, file editing, shell command execution, web search, and functionality extensions via Skills and MCP servers. The open-source release came roughly 72 hours after an AI-safety researcher published wire-level evidence that Grok Build had been silently uploading complete Git repositories to a Google Cloud Storage bucket run by xAI, including unredacted secrets and full commit history, even when the agent was not instructed to open files. SpaceXAI disabled default data retention for all users on July 12, committed to deleting all previously retained coding data, and stated that zero data retention had always been respected when users explicitly disabled upload.

5. Thinking Machines Lab Released Inkling

Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, released Inkling, its first production model trained from scratch. Inkling is a 975B-parameter MoE transformer with 41B active parameters, trained on 45 trillion tokens of text, images, audio, and video. It accepts text, image, and audio inputs and supports a context window of up to 1M tokens. The model includes controllable thinking effort from 0.2 to 0.99. On the Artificial Analysis Intelligence Index, Inkling debuted at 41, making it the leading US open-weights model, ahead of Nemotron 3 Ultra (38), Gemma 4 31B (29), and gpt-oss-120b (24). On coding, it scored 77.6% on SWE-bench Verified. Thinking Machines also previewed Inkling-Small, a 276B-parameter MoE with 12B active parameters that matches or exceeds Inkling on several benchmarks; weights will be released once testing is complete. Inkling is accessible via the Tinker platform API (256K context window), and weights are available on Hugging Face (1M context window) under Apache 2.0. The NVFP4 quantized checkpoint reduces aggregate VRAM requirements to approximately 600 GB.

6. OpenAI Introduced GPT-Red

OpenAI announced GPT-Red, an internal automated red-teaming model built to find prompt injection vulnerabilities at scale. GPT-Red uses self-play reinforcement learning: an attacker model and multiple defender models train against each other simultaneously, with each side earning rewards for successfully executing or resisting attacks. As defenders become more robust, GPT-Red generates progressively harder attacks, creating a training curriculum whose difficulty scales with the model’s own defenses. OpenAI used GPT-Red to adversarially train GPT-5.6, making it substantially more resistant to prompt injection. In head-to-head evaluations, GPT-Red beat human red-teamers 84% to 13% on prompt injection tasks. GPT-Red is kept separate from deployed models: attack capabilities remain internal, while the resulting robustness transfers to the defender. OpenAI acknowledges limitations: multi-turn attacks and image-based prompt injection still require human red-teamers. GPT-Red will not be released externally.

---

AI Tip of the Day

Structured Outputs make responses easier to plug into software, but they also change how the model generates its answer.

When we first wrote the lesson ‘Structuring Your Data’ for our Full Stack AI Engineering course, the main problem was getting models to return JSON your application could reliably parse. As structured outputs became standard across newer reasoning models and smaller models used in production, another problem became harder to ignore: the schema can be perfectly valid while the answer inside it gets worse.

This is especially important on tasks that already stretch the model’s reasoning. Requiring it to solve the problem within a rigid schema can reduce accuracy, particularly when the schema is complex, or the model has little capacity to spare.

Before adding a schema, run the task in free form and record the quality of the answers. Then introduce the smallest schema your application needs and repeat the same evaluation.

Measure answer quality and schema validity separately. A 100% parse rate only tells you that your software can read the response. It does not tell you whether the response is correct.

When structure reduces accuracy, simplify the schema. For tasks where the final formatting is deterministic, let the model work through the problem freely and package its answer into the required structure in code.

---

Five 5-minute reads/videos to keep you learning

1. How AI Agent Memory Actually Works — And How to Build It

Stateless LLM calls forget everything, so agents need a memory layer built on top. This guide maps the four memory types: working, episodic, semantic, and procedural, grounding them in the CoALA framework and Tulving’s research. It clarifies how agent memory differs from RAG, walks through the write-consolidate-retrieve-forget lifecycle, and includes a runnable, framework-agnostic Python implementation for each memory store.

2. Reward Design Is the Hard Part: Building Verifiable Rewards for Tool-Using Agents

RLVR’s clean verify-and-score loop breaks down once agents span 10 to 20 tool-calling turns, where a single terminal reward leaves most steps without signal. This article lays out a layered fix: turn-level verifiable rewards, rubric-based LLM judges, and guidance injection when sparsity stalls training entirely. It also covers agent-specific reward hacks like tool-call theater and retry gaming. The reward design here works less like ML and more like systems engineering, with its own test suite.

3. Building Stateful AI Agents That Survive Session Kills

Coding agents lose everything when a session dies. This article shows how to build a harness on Tensorlake’s Firecracker-backed MicroVM sandboxes that keeps state alive across days. These sandboxes suspend and resume with filesystem, memory, and running processes intact, while snapshots double as agent memory and fork into parallel workers for GSPO-style candidate ranking. The piece also covers isolated verifier sandboxes, custom baked images cutting cold starts from 12 seconds to 600 milliseconds, and native SSH workflows for persistent multi-day debugging.

4. Understand HNSW: Why Your Vector Search Returns Garbage (Build your own minimalist HNSW from scratch)

Vector search often fails at the index layer, not at the embedding layer. In this article, the author builds a minimal HNSW implementation in pure Python and numpy, then deliberately breaks it to show how weak build parameters cap recall at 0.44 while stronger settings reach 0.78. Measurements demonstrate that M and ef_construction set a hard ceiling no query-time tuning can break, while ef_search hits diminishing returns fast.

5. Context Engineering for Bedrock Agents: A Hands-On Guide Beyond Prompt Engineering

Prompt tuning alone can’t fix agents that hallucinate mid-loop, so this guide proposes a three-layer context architecture for AWS Bedrock: persistent identity and guardrails, TTL-bounded session state, and per-turn transient retrieval. The author builds a Python ContextAssembler that enforces token budgets, evicts expired layers, and logs every injection decision for debugging. It also covers wiring in Bedrock Knowledge Bases, semantic tool selection, session memory, cost guards, and log analysis.

Repositories & Tools

1. AirLLM runs 70B+ models on a single 4GB GPU by loading one transformer layer at a time, without quantization, distillation, or pruning.

2. Code-review-graph is a local-first code intelligence graph that parses codebases with Tree-sitter into a persistent SQLite knowledge graph.

3. AstrBot is an agentic chatbot platform that deploys LLM-powered assistants across mainstream instant messaging apps, with agent capabilities, RAG, MCP, knowledge bases, etc.

4. WrenAI is a generative BI engine that gives AI agents a governed context layer of business semantics, approved definitions, and memory.

5. DeepStream is a streaming analytics toolkit for AI-based video and image understanding, providing a GStreamer-based framework.

Top Papers of The Week

1. Function-Aware Fill-in-the-Middle as Mid-Training for Coding Agents

A coding agent’s action-observation-continuation loop is structurally identical to a function call site. This paper exploits that by masking functions selected via program dependency graphs during mid-training, forcing the model to predict continuations conditioned on values computed elsewhere. Despite training on Python only, the inductive bias transfers to non-Python coding and tool-use benchmarks. Mid-training improves SWE-Bench Verified by +2.8/+3.0 at 7B/14B and +3.2 on Qwen3–8B, while mitigating the capability erosion that agentic post-training otherwise inflicts on general coding and tool-use tasks.

2. GenCeption: Video Generators as General-Purpose Vision Learners

Google DeepMind repurposes a pre-trained video generative diffusion backbone into a single feed-forward perception model that handles depth, surface normals, camera pose, segmentation, and 3D keypoint prediction, all steered by text instructions. GenCeption matches or surpasses specialized models (DepthAnything3, SAM3, D4RT, VGGT-Omega) and achieves comparable performance with 7x to 500x less training data. A model trained exclusively on synthetic human videos generalizes to real-world footage and out-of-distribution categories, including animals and robots.

3. PalmClaw: Native On-Device Agent Framework for Mobile Phones

Existing mobile agents operate through GUI actions that form long, interface-dependent sequences. PalmClaw runs natively on mobile phones, exposing device capabilities (contacts, calendar, camera, sensors, apps) as structured tools the agent calls directly instead of simulating screen touches. The framework manages sessions, persistent memory, skills, and the full agent loop on-device, producing shorter action sequences with clearer execution boundaries.

4. RecGPT-V3 Technical Report

Running LLM-based recommendation at scale on Taobao exposed three bottlenecks: stateless user modeling, a lossy text-tag channel for item grounding, and verbose chain-of-thought latency. RecGPT-V3 addresses all three with a Memory Hub that cuts user-modeling compute by 55.8%, a Hybrid-modal Foundation Model that jointly reasons over text and Semantic IDs for direct item grounding, and Latent Intent Reasoning that compresses rationales into learnable latent tokens while lowering output token cost by 200x. In production A/B tests: GMV +3.97%, CTR +1.00%, with serving resource consumption down 52.4%.

5. Xiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World Trajectories

Xiaomi presents a foundational VLA model that can follow diverse language instructions and adapt to novel downstream tasks with minimal fine-tuning. It is pre-trained on over 100K hours of real-world manipulation trajectories collected via embodiment-free UMI devices. It uses a scalable auto-labeling pipeline to annotate trajectory clips with natural-language descriptions of scene state transitions. Post-training adapts the model to specific embodiments with minimal data. It establishes a new state-of-the-art with a 57.6% success rate on RoboCasa365.

Quick Links

1. Perplexity AI released WANDR, an open benchmark of 500 realistic data-collection tasks designed to evaluate whether AI research agents can search broadly and back every finding with specific evidence. Taken together, the 500 tasks call for 170,495 source-backed records. The tasks are ranked by 167 lower-, 166 middle-, and 167 higher-difficulty examples. Perplexity Search (Code) leads at 0.363 soft F1 and 0.133 hard F1. Anthropic is second at 0.249 and 0.072; every other system tops out at 0.121 soft F1 and 0.035 hard F1.

2. Zyphra released ZUNA1.1, an update to its 380M-parameter EEG foundation model that now accepts variable-length inputs from 0.5 to 30 seconds, up from the fixed 5-second segments in ZUNA1. The model reconstructs missing channels, denoises existing ones, and upsamples across arbitrary electrode layouts, trained on approximately 3.5M channel-hours of EEG data with realistic corruption patterns. Reconstruction quality matches or exceeds ZUNA1 while handling a far wider range of real-world recording conditions. Cloud-served inference is available in Zyphra’s EEG Playground with no GPU, installation, or coding required.

3. Mistral released Robostral Navigate, its first model optimized for embodied AI. The 8B vision-language model guides robots through complex environments using only a single standard RGB camera and natural-language instructions, eliminating the need for LiDAR, depth sensors, or multiple cameras. Trained entirely in simulation on 400,000 navigation trajectories across 6,000 scenes, it achieves 76.6% success on unseen R2R-CE environments, 9.7 points above the previous best single-camera model and 4.5 points above the best multi-sensor system.

4. PrismML released Bonsai 27B, a 27B-parameter multimodal model compressed from Qwen3.6 27B to 1-bit and ternary weights using end-to-end low-bit quantization. The 1-bit variant fits in 3.9 GB and runs on iPhone, iPad, and Mac via MLX; the ternary variant ships at 5.9 GB for laptops and consumer GPUs. Both retain the full 262K context window and support text, vision, tool calling, and agentic workflows. PrismML reports 90% and 95% intelligence retention, respectively. Speeds reach 163 tok/s (1-bit) and 134 tok/s (ternary) on an RTX 5090.

Who’s Hiring in AI

Senior Application Engineer @GovCIO (Remote)

AI Senior Project Manager @Huntington National Bank (Columbus, IN, USA)

AI Systems Engineer @ASG (Remote/USA)

Forward Deployed Engineer, Life Sciences @Domino Data Lab (Remote)

Forward Deployed Engineer @Twilio (Remote/USA)

Senior AI Engineer @Texas Sports Academy (Remote/USA)

Interested in sharing a job opportunity here? Contact sponsors@towardsai.net.

Think a friend would enjoy this too? Share the newsletter and let them join the conversation.