Staff · 47 pieces on file

Lars Iverson

Open source & model weights

Lars Iverson covers the open-weights ecosystem. He reads architecture diffs for a living and tracks every notable checkpoint release from Meta, Mistral, Qwen, DeepSeek, and the long tail. His beat is licences, parameter counts, and what the weights actually let an operator do.

Beats: open-source, model-releases

All pieces by Lars

Open Source · SEPTEMBER 21, 2026

Jev and Laya define a new model class: typed-decision heads that replace LLM calls for routing

TypeSafe's hosted Jev and Convai's open-weight Laya both target the classification layer around an LLM stack, at 32.8–276ms latency and $0.042 per million input tokens versus $0.20–$10 for frontier models.

Verdict Two releases, three days apart, mark the arrival of a distinct model class: non-autoregressive decision heads that price and time-slice differently from LLMs. Independent benchmarks still owed.
Open Source · SEPTEMBER 21, 2026

Qwen-Image-2.1 ships 7B open weights — under a research-only license

Alibaba's 32-layer single-stream DiT lands with native RGBA output and 10-reference editing on consumer GPUs, but the Qwen Research License Agreement dated September 20, 2026 bars commercial use without a separate agreement.

Verdict Genuinely capable open weights at 7B — but the license shift away from Apache 2.0 is the story, not the parameter count.
Open Source · SEPTEMBER 20, 2026

Ternary Bonsai 2 27B fits a 27B model in 5.9 GB while retaining 98.2% of Qwen3.8's benchmark average

PrismML's Apache 2.0 release compresses Qwen3.8 27B more than 9x, keeps math and coding nearly lossless, and runs at 143 tokens/second on an RTX 5090 — but only through the company's own llama.cpp fork.

Verdict Near-lossless on math and coding at a laptop-scale footprint, with the caveat that stock llama.cpp will not load the weights.
Open Source · SEPTEMBER 17, 2026

US pushes to cut Chinese labs off from frontier APIs — and the cheap-inference tier hangs in the balance

A Sept. 8 NSA/CISA/FBI advisory names six Chinese labs distilling Claude, GPT, Gemini, and Grok at industrial scale. If Washington closes the pipes, the price competition that made AI viable for small operators loses its floor.
Open Source · SEPTEMBER 11, 2026

DeepSeek V4.1 Flash: 552B MoE, MIT weights, and $0.15/M off-peak input

DeepSeek's new Causal Encoder-Decoder Flash tier posts 74.2 on DeepSWE v1.1 and 54.8 on AutomationBench while pricing input at a fraction of GPT-5.6 Sol and Claude Opus 5.

Verdict The cache-hit economics and MIT weights make V4.1 Flash the most operator-friendly frontier-adjacent release of September — but 552B backbone parameters mean the self-host math is not what 'Flash' used to imply.
Open Source · SEPTEMBER 11, 2026

DeepSeek V4.1-Flash lands at $0.003/M cached input, with a 552B backbone hiding behind an 8B prefill

DeepSeek's new Causal Encoder-Decoder model shrinks the KV cache to 890 bytes per token and prices cached input at $0.003/M off-peak — but the 552B backbone and harness-variance findings complicate the headline.

Verdict V4.1-Flash's cache economics are real; its 'Flash' branding is not. Self-hosting the 475 GiB checkpoint is a multi-GPU decision, not a laptop one.
Open Source · SEPTEMBER 10, 2026

DeepSeek V4.1-Flash Lands With FP4 KV Cache, MIT Weights, and a Forced V4 Pro Migration on September 14

The new 552B MoE flash tier ships today with native vision, a 1M-token window, and off-peak output at $0.60 per million tokens — and absorbs all V4 Pro API traffic at Flash prices four days later.

Verdict V4.1-Flash is the cheapest frontier-tier open-weight model DeepSeek has shipped — and the September 14 routing cutover makes it a forced platform decision for anyone on V4 Pro.
Open Source · SEPTEMBER 1, 2026

CrowdStrike's SafeMind bets domain-specific beats frontier at cyber defense

SafeMind pairs NVIDIA Nemotron open weights with 15 years of CrowdStrike breach telemetry into two purpose-built models — Red Tempest and Blue Solano — running in a closed-loop harness the vendor says triages threats 3x more accurately than generic frontier models.

Verdict SafeMind is the loudest production case yet for the argument that a fine-tuned open-weights model on proprietary vertical data beats a general-purpose frontier LLM on a specialized task.
Open Source · AUGUST 31, 2026

GLM-5.3-Flash Debuts at $0.075/$0.25 as DeepSeek Warns of 'Significant' Hike

Z.ai's MIT-licensed 320B-A18B MoE landed August 26 at half its list price through September 9, while DeepSeek's price-floor dominance shows visible cracks.

Verdict The API price floor is moving in two directions at once. Any small team running AI-powered outreach on DeepSeek V4-Flash should have a fallback ready before September 9.
Open Source · AUGUST 30, 2026

Z.ai Ships GLM-5.3-Flash: 320B-A18B MoE, MIT Weights, and API Pricing 10x Cheaper Than GLM-5.2

The model that spent a week anonymously topping OpenRouter as 'Ox Alpha' launched August 26 with a 1M-token context, hybrid linear-plus-sparse attention, and a launch-promo API price of $0.075 per million input tokens.

Verdict GLM-5.3-Flash is the cheapest capable open-weights coding model Z.ai has shipped — MIT-licensed, natively multimodal, and priced to reset the floor on agentic serving.
Open Source · AUGUST 29, 2026

GLM-5.3 weights land on Hugging Face: 755.7 GB of post-training gains, same 743B base

Z.ai completed the two-week staged rollout on August 28, publishing the full FP8 and BF16 checkpoints. The capability jump — Terminal-Bench 3.0 from 4.6 to 28.3, CyberGym to 84.5 — came entirely from post-training on a base model shared with GLM-5.2.

Verdict The weights are inspectable. The gains are real and vendor-reported. Independent reproduction still requires eight datacenter accelerators and matching harnesses.
Open Source · AUGUST 29, 2026

GLM-5.3-Flash lands MIT-licensed at ~$0.10/M blended after a week atop OpenRouter as Ox Alpha

Z.ai's 320B/18B-active multimodal MoE ships day-one MIT weights and a 1M-token context, arriving days after the full GLM-5.3 checkpoints cleared a two-week cyber-capability safety hold.

Verdict The MIT license and blended pricing near $0.10/M tokens make GLM-5.3-Flash the first frontier-adjacent open-weight model that small B2B teams can plausibly build customer-acquisition pipelines on.
Model Releases · AUGUST 22, 2026

Anthropic drops Mythos 5 into Claude Security, adds $35M open-source defender fund

The model previously locked to Project Glasswing partners now runs GitHub vulnerability scans for any Claude Enterprise customer, billed as standard token usage — with a partner channel and Cyber Verification Program expansion to follow.

Verdict Mythos 5 leaves the vetted-defender sandbox — but only through a narrow output surface that never hands the model itself to end users.
Open Source · AUGUST 19, 2026

Z.ai holds GLM-5.3 weights after 84.5% CyberGym result and faster-than-expected exploit-chain lift

GLM-5.3 shares its base model with GLM-5.2; the entire jump — CyberGym 84.5%, ExploitBench 54.4%, 105 ExploitGym tasks in two hours — comes from scaled post-training. Z.ai is delaying open weights by roughly two weeks.

Verdict The first GLM release where Z.ai held the weights back — and the numbers explain why.
Open Source · AUGUST 13, 2026

DeepSeek V4-Pro-0813 goes GA at 1.6T parameters, $0.87 per million out

DeepSeek quietly graduated its trillion-parameter MoE flagship on August 13, ships an MIT-licensed agent harness alongside it, and simultaneously announces a sharp API price hike beginning August 16.

Verdict A real open-weights release with real vendor-reported gains — and a pricing page that just changed underneath it.
Open Source · AUGUST 12, 2026

Meta and NVIDIA drop competing 30B open-weight agentic models within 24 hours

Muse Glimmer and Nemotron 3.5 Lightning both target the always-on agent execution layer on consumer GPUs — one a dense Apache 2.0 distillation, the other a hybrid Mamba MoE clocking 670 tok/s.

Verdict Two 30B open-weight agentic models in 24 hours is the clearest signal yet that the execution layer is moving off closed cloud APIs.
Open Source · AUGUST 11, 2026

Meta ships Muse Glimmer: a 30B open-weight agent for a single consumer GPU

Meta Superintelligence Labs distilled Muse Spark 1.2 into a 30-billion-parameter Apache 2.0 model that runs on 24–32 GB of VRAM, and pledged to open the teacher's weights within weeks.

Verdict A credible frontier-adjacent open-weight release from a U.S. lab — and the promised Muse Spark 1.2 drop is the one to watch.
Open Source · AUGUST 10, 2026

Meta ships Muse Glimmer: a 30B open-weights distill of Muse Spark 1.2 for one consumer GPU

Meta Superintelligence Labs released Muse Glimmer on August 10 under Apache 2.0 — a 30B multimodal agentic model distilled from the closed Muse Spark 1.2, engineered to run on a single 24 GB consumer GPU with day-0 support in transformers, vLLM, and llama.cpp.

Verdict The first Meta open-weights drop with a coherent story — small enough for one GPU, agentic by design, and paired with a policy fight over distillation and training data.
Open Source · AUGUST 10, 2026

Kimi K3 tops Arena Frontend Code at 1,679 Elo and 2.8T parameters — the largest open-weight release to date

Moonshot AI's 2.8-trillion-parameter Kimi K3 activates 16 of 896 experts per token, ships MXFP4 weights, and charges $3/$15 per million tokens. Full weights landed July 27; Washington is already arguing about what to do about it.

Verdict The first Chinese open-weight release to plausibly benchmark inside the frontier — and the first one whose architecture, not just its price, is worth reading.
Open Source · AUGUST 4, 2026

Alibaba's Qwen3.8-Max lands at 2.4T parameters, 95B active, and #2 on Vision Arena

The new flagship activates 95 billion of its 2.4 trillion parameters per token, supports a 1M-token context, and ships open weights next week alongside a 27B companion — Alibaba's largest release to date and its most direct challenge yet to Anthropic's Fable 5.

Verdict A trillion-parameter sparse MoE headed for open release next week — the scale is unprecedented in open weights, and the 95B active count keeps it serveable.
Open Source · AUGUST 4, 2026

Open-weight GLM-5.2 lands four months behind the frontier — and refuses nothing

A SaferAI evaluation published August 4 finds Z.ai's GLM-5.2 refused zero offensive-cyber or dual-use biology tasks, arriving alongside UK AISI and NIST CAISI measurements that put the open-weight cyber gap at four to seven months and closing.

Verdict The open-weights frontier is now four months behind, ships without a safety framework, and the safeguards closed labs rely on don't apply to a file on Hugging Face.
Open Source · AUGUST 1, 2026

DeepSeek V4-Flash-0731: A Re-Post-Trained Budget Model That Beats Its Own Flagship

DeepSeek moved V4-Flash out of preview on July 31 with a checkpoint identical in architecture to the preview build — 284B total parameters, 13B active, MIT license — that scores higher than V4-Pro-Preview on every agent and coding benchmark the company published.

Verdict A silent API upgrade that reframes the frontier: post-training, not parameter count, is where the agent gains are coming from.
Open Source · JULY 28, 2026

Kimi K3's 2.8T-parameter weights land a day early, and Washington fractures

Moonshot AI pushed the full weights for its 2.8-trillion-parameter Kimi K3 on July 26, roughly 24 hours before schedule. Within 48 hours, an industry letter opposing open-weight bans had picked up Google and OpenAI — and Anthropic's Dario Amodei had published a clarification that he does not, in fact, want the weights banned either.

Verdict K3 is the largest open-weight release on record and it has already reshaped the policy fight around it. The technical claims still need independent replication; the political ones do not.
Open Source · JULY 27, 2026

Kimi K3 ships its weights: 2.8T parameters, MXFP4, and a 1M-token context

Moonshot AI published the full weights of Kimi K3 on Monday, the largest open-weight model ever released, activating 16 of 896 experts per token and trained MXFP4-native for hardware portability.

Verdict The first 2.8-trillion-parameter open-weight release, quantization-aware from SFT onward, and the first Chinese open model shipped under an active Washington ban debate.
Infrastructure · JULY 23, 2026

Alphabet Q2 2026: Cloud accelerates to 82%, capex guide pushed to $205B as Gemini serves 22B tokens/minute

Google Cloud revenue jumped to $24.8B on enterprise AI infrastructure demand, and Alphabet lifted 2026 capex guidance by $15B — with CFO Anat Ashkenazi telling analysts demand 'continues to outpace supply across the industry.'

Verdict The serving story is now the earnings story: 22 billion tokens per minute, a $514B cloud backlog, and a capex line rising fast enough to turn free cash flow negative.
Open Source · JULY 23, 2026

Moonshot's Kimi K3 lands at 2.8 trillion parameters, weights follow July 27

Moonshot AI unveiled Kimi K3 on July 17 — a 2.8T-parameter sparse MoE with a 1M-token context — and promised full open weights ten days later. The release reshuffled the open-weight frontier and knocked TSMC down 7%.

Verdict The first openly downloadable model in the 3-trillion-parameter class, and the first from a Chinese lab to credibly trade blows with Fable 5 and GPT-5.6 on published benchmarks.
Open Source · JULY 20, 2026

Kimi K3 lands at 2.8T parameters, takes Frontend Code Arena, jams Moonshot's own capacity

Moonshot's July 17 open-weight release outscored every model except Claude Fable 5 and GPT-5.6 on its own suite, took the Arena Frontend Code top slot at 1,679 points, and forced the company to pause new subscriptions inside 72 hours.

Verdict K3 is the largest open-weight model shipped to date and the first from China to credibly threaten the closed frontier — but the numbers driving the reaction are all Moonshot's own.
Open Source · JULY 20, 2026

Moonshot ships Kimi K3 at 2.8T parameters — the largest open-weight model, weights due July 27

Beijing-based Moonshot released a sparse-MoE frontier system with a 1M-token context and Kimi Delta Attention, claiming a #2–3 overall finish behind Claude Fable 5 and GPT-5.6 Sol on the company's own eval suite.

Verdict The largest open-weight model ever announced — but every number is Moonshot-reported until the checkpoint drops on July 27.
Open Source · JULY 18, 2026

Moonshot's Kimi K3 lands at 2.8T parameters, sits behind only Fable 5 and GPT-5.6 Sol

Beijing's Moonshot AI unveiled Kimi K3 on July 16 — a 2.8-trillion-parameter sparse MoE with a 1M-token context window, novel Kimi Delta Attention, and full weights due July 27. It is the largest open-weight model ever released.

Verdict The first open-weight model in the 3T-parameter class, and the first from China to benchmark within a few points of Anthropic and OpenAI's frontier.
Open Source · JULY 17, 2026

Moonshot ships Kimi K3 at 2.8T parameters, third on Intelligence Index

Beijing's Moonshot AI released Kimi K3 on July 16 — a 2.8-trillion-parameter sparse MoE with Kimi Delta Attention, a 1M-token context window, and API pricing at $3/$15 per million tokens. Full weights are due July 27.

Verdict The largest open-weight model ever shipped, and the first Chinese release to benchmark within striking distance of Fable 5 and GPT-5.6 — but every number is Moonshot-reported until the weights land.
Open Source · JULY 15, 2026

Chinese open-weight models pass 41% of Hugging Face downloads as Nadella tells enterprises they're paying twice

Chinese labs took the top six slots on OpenRouter and 41% of Hugging Face downloads this spring. On July 13, Microsoft's CEO gave the shift its first executive-suite endorsement.

Verdict The open-weight tier is no longer the value option — it is the default production tier, and Microsoft just said so out loud.
Open Source · JULY 14, 2026

Chinese open weights take 41% of Hugging Face downloads and all six top OpenRouter slots

Real production-traffic data from OpenRouter and Vercel's AI Gateway shows US model share collapsing from 70% to 30% in twelve months, with DeepSeek, Z.ai, Tencent, Xiaomi, and MiniMax now processing three times more tokens per week than American labs.

Verdict The open-weights ecosystem stopped being a research curiosity in the twelve months to June 2026 — it became the volume tier of the global inference market.
Open Source · JULY 14, 2026

Goldman initiates on Z.ai at HK$1,880; GLM-5.2 scores 81.0 on Terminal-Bench 2.1

Goldman Sachs named Z.ai's GLM-5.2, DeepSeek and ByteDance its preferred Chinese AI stack on July 10, days after the 744B-parameter MIT-licensed model cleared Gemini 3.1 Pro on terminal work and undercut GPT-5.5 API pricing by roughly 6x.

Verdict GLM-5.2 is the first open-weights release the desk has seen carry a Goldman buy-side thesis on its back — near-frontier scores, MIT licence, and API pricing at one-sixth of GPT-5.5.
Model Releases · JULY 7, 2026

White House frontier-model framework lands this week, with a 30-day NSA access window

The voluntary standards implementing Section 3 of Trump's June 2 executive order will govern how OpenAI, Anthropic, and Google release covered frontier models — and directly unblock GPT-5.6's broader launch.
Open Source · JULY 5, 2026

Meituan ships LongCat-2.0: 1.6T MoE, 1M context, trained end-to-end on Chinese ASICs

Meituan open-sourced LongCat-2.0 on Hugging Face and GitHub under MIT, unmasking the 1.6-trillion-parameter MoE that had been leading OpenRouter as 'Owl Alpha' — and the first trillion-parameter system pretrained and served entirely on a 50,000-card domestic ASIC cluster.

Verdict The first credible frontier-scale open-weights release trained pretraining-to-inference on non-Nvidia silicon — and it lands on SWE-bench Pro ahead of GPT-5.5.
Model Releases · JUNE 28, 2026

Grok 4.5 ships to SpaceX and Tesla only: 1.5T parameters, V9 foundation, Cursor-tuned

xAI's V9-based Grok 4.5 entered closed beta on June 28 inside Musk's two engineering companies — 50% larger than Grok 4.4, trained with Cursor data, and benchmarked internally against Claude Opus.

Verdict A captive-beta release pattern that swaps public benchmarks for proprietary engineering telemetry — and an explicit pledge of monthly from-scratch models through year-end.
Open Source · JUNE 24, 2026

GLM-5.2 lands at 744B parameters, MIT-licensed, and tied with Opus 4.8 on long-horizon coding

Z.ai's open-weights flagship debuts at #1 on open-source coding boards with a 1M-token context, IndexShare cutting per-token FLOPs 2.9x, and API pricing roughly one-sixth of GPT-5.5's.

Verdict The first open-weight MoE that genuinely closes on Opus 4.8 in long-horizon coding — and the cost gap is the real story.
Open Source · JUNE 23, 2026

OpenAI ships full GPT-5.5-Cyber at 85.6% CyberGym, opens Patch the Planet with Trail of Bits across 19 projects

The Daybreak expansion pairs a permissive-only preview's full release — 85.6% CyberGym, 39.5% ExploitGym — with an open-source remediation sprint that merged dozens of patches across cURL, Python, and the Linux kernel in its opening week.

Verdict GPT-5.5-Cyber's full release is the first frontier defensive model to ship with both a measurable benchmark lift and a maintainer-facing patching program attached.
Model Releases · JUNE 18, 2026

Pentagon Sworn Statement Puts Grok Gov Model on 2,000 Strikes in 96 Hours

A federal court filing from DoD AI chief Cameron Stanley makes xAI's Grok Gov Model the first commercial LLM officially named as enabling a live strike campaign — 2,000 munitions, 2,000 targets, four days, inside Maven Smart System.

Verdict The Grok Gov Model is now the first frontier LLM with a sworn-statement combat record — and the disclosure tells us more about Pentagon procurement than about the model itself.
Open Source · JUNE 14, 2026

US orders Anthropic to pull Fable 5 and Mythos 5 worldwide over a narrow jailbreak claim

A Commerce Department export-control letter sent at 5:21 pm ET on June 12 forced Anthropic to disable its Mythos-class models for every user globally — three days after Fable 5's general release.

Verdict The first government-forced recall of a publicly deployed frontier model rests on verbal evidence of a non-universal jailbreak that, by Anthropic's account, GPT-5.5 already replicates.
Open Source · JUNE 13, 2026

Z.ai Ships GLM-5.2 With 1M-Token Context and an MIT Pledge — and No Benchmarks

Zhipu's international brand pushed its 744B MoE flagship to a million-token window and added dual thinking-effort presets, but launched without a single score and gated the weights behind a 'next week' promise.
Open Source · JUNE 12, 2026

Kimi K2.7-Code ships open weights, cuts thinking tokens 30%, and edges Opus 4.8 on MCPMark

Moonshot's coding-focused post-train on the K2.6 MoE family lands on Hugging Face under a Modified MIT license, reports +21.8% on its own Kimi Code Bench v2, and forces thinking mode on every call.

Verdict An open-weight 1T MoE that beats Opus 4.8 on one tool-use benchmark, trails GPT-5.5 on all six, and burns 30% fewer reasoning tokens doing it.
Open Source · JUNE 11, 2026

OpenAI files confidential S-1 with the SEC, one week after Anthropic

OpenAI confirmed a draft registration statement on June 10, 2026 at an $852 billion valuation. Goldman Sachs and Morgan Stanley are leading; the company says it has not committed to a timeline.
Open Source · JUNE 9, 2026

Microsoft ships seven MAI models from scratch, declares independence from OpenAI distillation

At Build 2026, Mustafa Suleyman's AI Superintelligence team unveiled MAI-Thinking-1 at 97% on AIME 2025 and 53% on SWE-Bench Pro, alongside a 5B-active coding model that lands today as a VS Code default — all trained without third-party distillation.

Verdict The story isn't the seven models — it's the seven-model pipeline, built from a 794-billion-page proprietary crawl and co-designed against Maia 200 silicon, with no OpenAI weights anywhere in the lineage.
Model Releases · JUNE 8, 2026

Apple's iOS 27 Extensions API turns the iPhone into a four-way model marketplace

Siri runs on a custom 1.2-trillion-parameter Gemini under a reported ~$1B/year Google deal, but the bigger release is the Extensions framework letting users set Claude, ChatGPT, Gemini, or Grok as the system-wide default.

Verdict The Extensions API is the most consequential model-distribution event of the year — and the end of OpenAI's iPhone exclusivity.
Infrastructure · JUNE 7, 2026

Apple licenses a 1.2T-parameter Gemini MoE for Siri, runs it on B200s inside Private Cloud Compute

Bloomberg, TechTimes and Google Cloud's own CEO line up the same architecture ahead of Monday's WWDC keynote: a custom mixture-of-experts Gemini, ~$1B/year, weights sitting on Nvidia B200s inside Apple-controlled enclaves.

Verdict The largest third-party model licensing deal in consumer AI history — and the first where the licensee insists on hosting the licensor's weights on its own silicon.
Open Source · JUNE 6, 2026

Microsoft ships seven MAI models, with MAI-Thinking-1 matching Opus 4.6 on SWE-Bench Pro

At Build 2026, Microsoft AI released a 35B-active-parameter sparse MoE reasoning model trained from scratch, plus six companions across image, voice, transcription, and coding. The flagship hits 97% on AIME 2025 and 53% on SWE-Bench Pro.

Verdict Microsoft's first plausible play for model independence — a clean-data MoE that lands on Opus 4.6's coding number without a distillation footnote.

← Back to our writers