V1.1 state of open source- OS 4.4 months behind frontier [pdf]

THE STATE OF OPEN SOURCE AI A RECURRING MOZILLA ASSESSMENT Open Source AI 2026 Volume 1.1 September 2026 FOREWORD Raffi Krikorian Chief Technology Officer Mozilla In New Zealand's far north, a Māori broadcaster trains speech models for te reo, a language too small for any market, under a license that keeps the data with its people. In East Africa, farmers diagnose cassava disease with a model that runs on the phone itself, offline, in fields the cloud has never reached. In Switzerland, a public consortium trained a national model on public supercomputers and released all of it: weights, data, training code. None of them asked permission, and none of them could have rented this. They own it. That is the whole idea. Read what follows as a map: where open AI is winning — some numbers surprised even us — and where it is exposed. A case that hides its weak points is an advertisement. In the six weeks since our first draft, a major US lab returned to open weights and Washington debated a ban and declined it. The map moved while we drew it. Contents. 01 Capability and price Open models caught up: capability, cost, usage, size, efficiency 02 Ships easy, deploys hard Adoption, production, tooling, churn, the remaining gaps 03 Why they want the weights Optionality, cost control, the policy fight of 2026, sovereignty 04 Capital above the model Revenue, venture, corporates, consolidation 05 The harness Where the labs now compete: harness, wire format, standards, security 06 The new lock-in Memory, state, and the training loop 07 Opportunities Where to act while the layer is still open Sources: Mozilla / SlashData 2026 developer survey (n=1,410) · OpenRouter · Artificial Analysis · Epoch AI · METR · Linux Foundation / AAIF · EU AI Act & Commission guidance · public filings and reporting. FRONT MATTER · 04 "Open source" and "open weights" are different claims. MODEL LICENSE WEIGHTS BASE DATA CODE REPORT INFER. RESTRICTIONS Kimi K3 · Moonshot Custom Kimi K3 MaaS / UI attribution GLM-5.2 · Zhipu MIT None DeepSeek V4 Flash 0731 MIT None DeepSeek V4 Pro 0813 MIT None Qwen3.8-2.4T-A95B Custom Qwen3.8-Max Attribution + MaaS MiniMax-M3 MiniMax Community Commercial terms MiMo-V2.5-Pro · Xiaomi MIT None Inkling · Thinking Machines Apache 2.0 Separate AUP Nemotron 3 Ultra · NVIDIA OpenMDW 1.1 Dataset gating Muse Glimmer · Meta Apache 2.0 None Mistral Medium 3.5 Modified MIT Revenue carve-out Gemma 4 31B · Google Apache 2.0 None Nemotron 3 Super · NVIDIA Nemotron Open Model Custom; gating gpt-oss-120b · OpenAI Apache 2.0 None Nemotron 3.5 Lightning OpenMDW 1.1 Dataset gating Command A+ · Cohere Apache 2.0 None Yes Partial / gated No In this report, "open" means downloadable weights unless stated otherwise. 0 of 16 notable open releases ship the data recipe the OSI definition asks for. 8 license families: Apache 2.0, MIT, modified MIT, OpenMDW 1.1, Nemotron, and custom Kimi K3, Qwen3.8-Max, MiniMax. THE NUANCE Custom ≠ closed, Apache ≠ unconditional — Inkling-Small pairs Apache 2.0 weights with a separate, changeable Acceptable Use Policy. Data. No row ships a full corpus: three gate partial data, thirteen publish none. NVIDIA ungates a sample set; the remaining code, math and multilingual data needs gating and approval under its Data Access Agreement for Model Training. Gemma 4 31B is Apache 2.0 on its model card, which replaced the Gemma Terms of Use in April 2026. Qwen3.8 adds attribution above 100M MAU or $20M monthly revenue and a paid MaaS licence above $50M TTM. Mozilla analysis, Aug 2026 · model cards and LICENSE files · huggingface.co/moonshotai/Kimi-K3. SECTION 01 OF 07 Open models caught up on capability and price. CAPABILITY COST USAGE SIZE EFFICIENCY SECTION 01 · 06 What changed since 2023. That experience dates to Llama 2, two model generations ago. 2023 Llama 2 Quality issues, hallucinations Fine-tuning failures VERDICT Reverted to GPT-4 2024 Llama 3 era Major capability improvements Better reasoning, better context VERDICT A real alternative 2025 DeepSeek era Near-parity with the frontier Lower costs, strong economics VERDICT Competitive 2026 Production scale Enterprise adoption at scale Open models power ~⅓ of tokens; Meta returns to open (Aug) VERDICT A commercial ecosystem The decision has moved from the model to the harness, and that's where open is still hard to deploy. LMArena · DeepSeek-R1 release (MIT, Jan 20, 2025) · Mozilla/SlashData 2026 survey · Meta AI blog, "Introducing Muse Glimmer" + huggingface.co/blog/muse-glimmer (Aug 10, 2026). SECTION 01 · 07 The frontier's top four are closed. The next four are open. Artificial Analysis Intelligence Index, a 9-eval composite including Terminal-Bench 2.1, Humanity's Last Exam and GPQA Diamond. Top ten, v4.1.1. Claude Opus 5 Claude Fable 5 GPT-5.6 Sol Grok 4.6 Kimi K3 OPEN GLM-5.3 OPEN Qwen3.8 2.4T OPEN GLM-5.3 Flash OPEN Muse Spark 1.2 PROMISED GPT-5.6 Terra 3 pts gap from the best open model (K3, 60) to the closed leader (Opus 5, 63) #5 Kimi K3's rank, behind Opus 5, Fable 5, and the Sol / Grok 4.6 pair at 61 4 of the top ten are open — K3, GLM- 5.3, Qwen 3.8 and GLM-5.3 Flash. A fifth, Muse Spark 1.2, still has no weights, nor does 1.3 (Sept 2). MOVEMENT INTO SEPT 1 Kimi K3 60 now ties GLM-5.3. New open entries: GLM-5.3 60, GLM-5.3 Flash 57, and Meta's Muse Spark 1.2 57 — API live; open weights promised, not yet shipped. V4-Pro-0813 53. Artificial Analysis Intelligence Index v4.1.1 · Sept 1, 2026 data · vendor pricing pages. Figures here are v4.1.1; v4.2 announced Sept 4, 2026. 63 62 61 61 60 TOP OPEN 60 58 57 57 57 SECTION 01 · 08 The best open model trails Fable 5 by two points at 30% of the price. Snapshot: v4.1.1, Sept 1, 2026. Fable 5 and Sol are scored as deployed systems with fallbacks and guards, so the gap compares a system to a bare model. MODEL INDEX SCORE (52–63) LIST PRICE · INPUT / OUTPUT Opus 5 63.0 $5 / $25 Fable 5 62.1 $10 / $50 Sol 61.0 $5 / $30 · Fast $10/$60 Grok 4.6 61.0 $2 / $6 <200K · $4/$12 above Kimi K3 · OPEN 60.0 $3 / $15 GLM-5.3 · OPEN 60.0 $1.15 / $3.50 Qwen 3.8 Max · OPEN 58.0 $2 / $6 GLM-5.3 Flash · OPEN 57.0 $0.075 / $0.25 Terra 57.0 $2 / $12 · cut 20% Jul 30 Sonnet 5 53.4 $3 / $15 · $2/$10 promo K3 AGAINST FABLE 5 −2.1 index points 30% of the input price K3 AGAINST OPUS 5 −3.0 points · 60% of the input price THE CHEAPEST ROUTE K3 lists at $3.00/M input; the cheapest OpenRouter routes are reported at $2.60–$2.80/M. Sonnet 5 sits at the same $3/$15 list, nearly seven points lower on the index. Artificial Analysis Intelligence Index v4.1.1, Sept 1, 2026 · list pricing from each vendor's pricing page (Anthropic, OpenAI, xAI, Moonshot, Z.ai, Alibaba Cloud), verified Aug 2026. Artificial Analysis figures here are v4.1.1 · Artificial Analysis, "Announcing Artificial Analysis Intelligence Index v4.2," Sept 4, 2026. SECTION 01 · 09 The open-closed gap is five ECI points and it resets every release cycle. Epoch Capabilities Index by release date — the best open model (Kimi K3, 157) against the closed frontier (Claude Fable 5 and Opus 5, 162). Claude Fable 5 · Anthropic 162 Claude Opus 5 · Anthropic 162 GPT-5.6 Sol · OpenAI 161 GPT-5.6 Terra · OpenAI 159 Gemini 3.7 Flash · Google 157 Kimi K3 · OPEN 157 Grok 4.6 · xAI 156 Qwen 3.8 Max · OPEN 156 GPT-5.6 Luna · OpenAI 156 Claude Sonnet 5 · Anthropic 156 ECI scale · axis truncated 148 153 158 163 168 Squares mark point estimates; whiskers show 90% confidence intervals. 5 ptsFable 5 and Opus 5 at 162, Kimi K3 at 157 — below Epoch's measured average of 8 points since January (90% CI 7–11). WHY IT DOESN'T COMPOUND Every model here was released between June and August 2026. The margin between leaders resets each release cycle rather than accumulating, and Epoch translates the average 8-point gap to about four months. K3 and Gemini 3.7 Flash now sit level at 157, intervals fully overlapping. Epoch notes two reasons this gap may be understated: open models optimize on benchmarks more aggressively, and unpublished closed models are absent from the baseline. The August releases are now scored: Qwen 3.8 Max enters at 156, V4-Pro-0813 at 155. Epoch AI, Epoch Capabilities Index (CC-BY) · Sept 1, 2026 data · Epoch AI, "Open models lag state-of-the-art closed models by 4 months" (May 2026): average gap 8 ECI pts, 90% CI 7–11. SECTION 01 · 10 Closed models handle 8-to-12-hour tasks. Open gets there four months later. The METR doubling law converts the score gap into a band of tasks — and a calendar. Two independent estimates now agree: 4 months from scores (Epoch), ≈4.4 months from measurement (Mozilla's computation on METR's raw data). 50% TIME HORIZON BY RELEASE DATE · LOG SCALE 8–12 h band 2 m 10 m 1 h 4 h 12 h Jun '24 Jan '25 Jun '25 Jan '26 ≈ 4.4 months Opus 4.6 · ~12 h Opus 4.5 · 230 min · TH1.1 re-estimate K2 Thinking · 59.5 gpt-oss-120b · 51 R1 · 30.3 Closed · doubling 5.5 mo Open · doubling 3.9 mo 4.4 months measured lag 1.74× task-length ratio THE PAY RULE Under 8 hours, either model does the job. Past 12, neither does. The ~5× per-task premium pays only inside that band — and the band moves, because a task at the closed frontier today is at the open frontier four months later. Pay when four months on the marginal task is worth 5×: deadline work yes, routine work no. DERIVATION METR TH1.1 puts the doubling at 6.3 months all-time, 4.3 months from 2023 on, and ~3 months from 2024 on. Against METR's published doublings (6.3 and 4.3 months), a 4-month lag converts to a 1.5×–1.9× task-length band. Against Mozilla's own fitted doublings (closed 5.5 mo, open 3.9 mo) the same lag gives 1.7×–2.0×. Mozilla's measured fit lands at 1.74×, inside both. METHOD NOTE METR's latest published point is an early Claude Mythos Preview snapshot (Mar 2026, published May 8): a 50% horizon of ≥16 h (95% CI 8.5–55 h), which METR flags as at the edge of its task suite, so it is omitted here. It also flags measurements above 16 hours as unreliable, and puts error bars at ~2×. Opus 4.6 drops ~40% to 7 h 11 m on private- only