Open weights and the routing market
Open-weight models are taking production routing share faster than their licences are settling.
-
Developers are routing production traffic to an anonymous provider that retains their prompts
Ox Alpha took 23.2 trillion OpenRouter tokens this week, first place by more than double the next model. Joseph W. Elstner, August 23, matched its tokenizer to Zhipu's released GLM-5 vocabulary across 95 probes at zero error, but cautioned this “establishes vocabulary identity, not ownership.” OpenRouter states the provider retains prompts and completions. Elstner
-
WATCHQwen open-sources Qwen3.8-Flash-Next tonight
a multimodal mixture-of-experts model previewing the Qwen4 architecture, with an FP8 variant. No specs and no benchmarks disclosed yet. TechNode
-
Open weights are winning the routing market before they even have a name
Z.ai confirmed August 26 that Ox Alpha, first on OpenRouter by routed tokens, was GLM-5.3-Flash, then released the weights under MIT: 320B total, 18B active. Simon Willison ran Qwen3.8-Flash-Next on a DGX Spark the day it shipped, 125B total but “only 6B active”. Daniel Han had Unsloth GGUFs ready day zero. The identity question filed here Monday moved, and it moved open. Willison
-
SHIPZ.ai opens GLM-5.3-Flash under MIT
320B total, 18B active, 1M context, multimodal. This is the model that ran anonymously as Ox Alpha, served on Chinese AI chips. Z.ai
-
Deep Cogito raised a $43M Series A to build open-weight models companies control themselves. WSJ
-
SHIPUnsloth v0.1.804-beta quantizes GLM-5.3-Flash and Qwen3.8-Flash-Next
GLM-5.3 fits 102GB of RAM plus VRAM, Qwen3.8 fits 75GB. It also fixes NVIDIA freezes under Wayland and AMD model-load crashes. Release
-
SHIPZ.ai puts full GLM-5.3 under a custom license
Companies above $10B revenue over any 12 months must pass Z.ai's security review before commercial use. GLM-5.2 was MIT, and Flash stays MIT. The New Stack
-
SHIPTencent opens Hy4 preview at 770B total, 49B active
1M context and 1.56TB of weights, with a chat template carrying two reasoning modes: high by default, no_think to switch off. Willison ran it August 29. Willison
-
SHIPDeepSeek open-sources its first multimodal model
V4-Flash-Vision-Exp shipped on Hugging Face under an MIT license August 31, ten days after the same weights went live on DeepSeek's API. DeepSeek says it beats the text-only V4-Flash on chart-reading and agentic benchmarks; the company calls its own eval set narrow. Hugging Face
-
ACTGLM-5.3-Flash still has no merged llama.cpp support
Three competing pull requests have sat open since August 26; the author of one flags that numerical validation against the original weights was never finished. Unsloth's own fork branch is the only working local path for now. llama.cpp
-
vLLM v0.28.0 shipped August 26 with Kimi K3 decode-context-parallel optimizations, 584 commits from 270 contributors. vLLM
-
Open-weight adoption is outrunning the tooling that runs it locally
GLM-5.3-Flash led OpenRouter's token volume at roughly 1.5 trillion tokens a day as of August 30, ahead of DeepSeek V4 Flash, while llama.cpp mainline still cannot load it. Three pull requests remain unmerged; Daniel Han's fork hit a new Metal integer-overflow bug September 1. A separate opencode issue found the model's reasoning collapsing into repeated exclamation marks under SGLang's tool-calling parser. OpenRouter · Han · opencode
-
WATCHZ.ai's H1 revenue jumps almost 400% but misses estimates
Revenue hit 954 million yuan, about $142 million, with API and open-platform business now 86.5% of the total. Net loss narrowed 12% to 2.07 billion yuan, against the backdrop of GLM-5.3's new revenue-gated license. SCMP
-
WATCHllama.cpp merges the Metal fix for the GLM-5.3-Flash overflow bug
PR #28210, merged September 2, promotes offset math to 64-bit and ends the long-context collapse Daniel Han flagged September 1. Three separate pull requests adding GLM-5.3-Flash support to llama.cpp mainline remain open. llama.cpp
-
GLM-5.3-Flash's OpenRouter volume rose to 1.9 trillion tokens a day; mainline llama.cpp support is still unmerged. OpenRouter
-
Enterprises are moving production workloads to open-weight models once they're good enough, not waiting for frontier quality
Gergely Orosz's September 3 Pulse newsletter reports companies shifting less-complex workloads to open models for roughly 50% lower infrastructure costs. The New York Times framed the same shift September 4 as a "good enough" threat to Anthropic and OpenAI's premium pricing, the same week Fable 5.1 itself got a 75% cache-read price cut. Orosz
-
WATCHvLLM's GLM-5.3 ROCm accuracy fix approved, not yet merged
The fix for vLLM's model-runner bug that collapsed GLM-5.3 accuracy on ROCm from 91.6 percent to 14.9 percent won reviewer approval September 2, closing in on landing; the VLLM_USE_V2_MODEL_RUNNER=1 workaround still applies until it merges. vLLM PR #54925
-
Three competing llama.cpp pull requests for GLM-5.3-Flash support remain unmerged; reviewers flag a 7 percent decode-throughput cost in one. llama.cpp PR #27754
-
Support for GLM-5.3-Flash in llama.cpp is consolidating around one implementation, not the one its most prominent contributor wrote
Three competing llama.cpp pull requests for GLM-5.3-Flash have been open since late August; reviewers are now converging on PR #27773 over Daniel Han's #27754, which still repeats tokens at deep context. Han pushed back: 'I know there is another impl, but this impl has been validated and utilized by many folks and works fine.' A third PR, #27752, is being abandoned in #27773's favor. llama.cpp PR #27773 · llama.cpp PR #27754
-
Unsloth's v0.1.806-beta doubled decode speed for GLM-5.3-Flash and Qwen3.8-Flash-Next via multi-token prediction, now on by default. Unsloth
-
OpenRouter rankings still show no reversal: DeepSeek V4 Flash leads GLM-5.3-Flash, 1.7 trillion tokens a day versus 1.5 trillion. OpenRouter
-
The production routing leaderboard for open-weight models reshuffled again this week, and the new leader is a model most readers haven't heard of yet
OpenRouter's rankings as of September 8 put Tencent's Hy4 Preview, released August 28, at 3.4 trillion tokens a day, more than any other model tracked, while GLM-5.3-Flash (1.8T) overtook DeepSeek V4 Flash (1.7T), a reversal from earlier in the week. Teortaxes, testing Hy4 directly, called it 'outside my expectations.' vLLM shipped support for both models on release day, September 9. OpenRouter · Teortaxes
-
SHIPvLLM 0.29.0 makes Model Runner V2 the default and adds day-one support for Hy4 and Qwen3.8-Flash-Next
The September 9 release, 594 commits from 277 contributors, turns on FlashInfer all-reduce by default, drops ten deprecated model architectures, and replaces the old module-based server invocation with vllm serve. vLLM v0.29.0
-
Nathan Lambert: Chinese open-model releases increasingly attach revenue thresholds and security reviews, while Western labs still default to Apache 2.0. Lambert
-
llama.cpp's three competing GLM-5.3-Flash pull requests remain unmerged; Daniel Han's PR #27754 still needs two maintainer approvals. llama.cpp PR #27754
-
Tencent's Hy4 Preview keeps its OpenRouter lead; OpenAI's GPT-5.6 Luna is the fastest riser, second ahead of GLM-5.3-Flash. OpenRouter
-
The blocking CUDA crash in GLM-5.3-Flash's llama.cpp port was fixed September 11; the merge still awaits maintainer review. llama.cpp PR #27754
-
SHIPllama.cpp v0.4.1 drops mmap and mlock for one load-mode flag
Georgi Gerganov's release adds Maple 20B-A1B and Tencent Hy4 support but removes --mmap, --mlock and --direct-io in favor of a single --load-mode flag — a breaking change for any script or config that used them. llama.cpp releases
-
Open-model routing rankings are turning over within days of a launch, not months
DeepSeek's V4.1 Flash — launched September 10 with a novel causal encoder-decoder architecture — climbed to challenge Tencent's Hy4 Preview atop OpenRouter's daily token-volume ranking within days, pushing GLM-5.3-Flash down the list, per Tokenmaxxing's tracking of OpenRouter's own usage data. Tokenmaxxing · Latent Space
-
WATCHGLM-5.3-Flash and a cut-rate OpenAI model climb OpenRouter's rankings
GLM-5.3-Flash entered OpenRouter's top rankings at #2 by weekly token volume; OpenAI's cheapest tier, GPT-5.6 Luna, jumped 129% week over week to #3, behind DeepSeek V4.1 Flash — widening the field of models taking real routing share. Tokenmaxxing
-
SHIPStepFun ships Step 5 Preview, a 600B open model matching Kimi K3 Max
The 600B-parameter sparse MoE model (27B active) adds 1M-token context and native image input, scoring 44 on Artificial Analysis's index. API pricing starts at $1/M input tokens; full weights open October 15. Pandaily · Artificial Analysis
-
SHIPQwen-Image-2.1 drops Apache 2.0 for a non-commercial-only license
Alibaba's new 7B open image-editing model adds native transparent-image output and up to 10 reference images, but ships under a Qwen Research License barring commercial use without a separate license from Alibaba. Hugging Face · GitHub LICENSE
-
SHIPXiaomi open-sources a frontier-tier model at a fraction of the cost
MiMo-V2.6-Pro ties Grok 4.7 at 46 on Artificial Analysis's Intelligence Index. Xiaomi released it under an MIT license along with its RL training code and environments, and says the full training run cost about $2.62 million. Artificial Analysis · VentureBeat
-
Chinese open-weight models have moved from catching up to outright leading their US counterparts
Nathan Lambert's Sept 21 congressional briefing found Chinese open models draw roughly 3.2 billion Hugging Face downloads against 1.6 billion for US open models, and OpenRouter's Chinese-model usage share grew from about 70% in early 2025 to over 80% now. On Artificial Analysis's Intelligence Index, GLM-5.3 and Kimi K3 lead the top US open models by a wide margin; Lambert argues this reflects real capability gains, not distillation. Nathan Lambert
-
Jack Clark's Import AI 473 also found 55% of new uncensored Hugging Face models are now Chinese-origin. Jack Clark
-
OpenRouter: GLM 5.3 Flash and DeepSeek V4.1 Flash each routed about 18 trillion tokens the week of Sept 17-22. OpenRouter
-
OpenRouter's Sept 26 rankings: Opus 5.5 and Fable 5.1 lead by tokens; DeepSeek V4.1 Flash leads coding routes. OpenRouter
-
Chinese labs now clearly lead open-weight models, not just compete in them
Nathan Lambert's Sept 21 Congressional testimony argues Chinese firms have led open-weight models since roughly April 2025, distinguishing true open-source (weights, code, data) from merely open-weight releases. Sebastian Raschka's Sept 22 benchmark read of Xiaomi's new MiMo-V2.6-Pro found it ranks first among 114 tracked open-weight models on Artificial Analysis's Intelligence Index. Two independent practitioners, one week. Lambert · Raschka, via X
-
A stealth model called Space Bunny Alpha briefly led OpenRouter's daily token usage this week; its maker is still unconfirmed. OpenRouter
-
OpenRouter's usage ranking this week put Claude models in six of the top ten spots; Qwen 3.8 Max was the only non-US entrant. OpenRouter
-
A practitioner is funding an institutional alternative to frontier labs' declining transparency about post-training research
Nathan Lambert co-founded Trillium Labs, a nonprofit publishing open, fully transparent post-training research - data, code, evals and checkpoints - positioned against labs releasing less about how they build models. Announced Oct 1, built on arguments Lambert has made for months that closed labs aren't safer, just less examined. Lambert
-
Claude Opus 5.5 and Sonnet 5.5 topped OpenRouter's usage rankings this week; Qwen 3.8 Max was the top non-US model. OpenRouter
-
OpenRouter's usage rankings held Claude Opus 5.5, Sonnet 5.5 and Qwen3.8 Max on top; GPT-6 Astra closed within a point. OpenRouter
-
WATCHllama.cpp ships formal GLM-5.3-Flash support; three tracked bugs stay open
llama.cpp's Oct 5 v0.6.0 release added GLM-5.3-Flash support and new Metal kernels, but didn't cite or close the VRAM regression, the Metal decode stall, or vLLM's speculative-decoding bug tracked since early October - all three remain open as filed. llama.cpp releases · llama.cpp issue 29950 · vLLM issue 59724
-
OpenRouter's rankings still put Claude Opus 5.5 and Sonnet 5.5 on top; Qwen3.8 Max remains the leading non-US model. OpenRouter
-
A credible voice argues restricting open-weight models over cyber risk would be policy malpractice, not caution
Nathan Lambert's Oct 6 post, responding to the week's push to gate models like GLM-5.3 over exploit risk, argues closed models have caused more documented real-world attacks than open ones, and that banning open weights without also restricting closed-model APIs is an inconsistent, lose-lose position. His own argument, direct dissent. Lambert
-
SHIPReflection AI ships Beam, a 501-billion-parameter open-weight model
Beam, Reflection's first open release, matches GLM 5.2 on reasoning using 3 to 4 times less inference compute, pitched as a Western open-weight alternative to Chinese models. Apache 2.0 weights follow later in October; only the pitch is public now. TechCrunch
-
SHIPMistral previews Large 4, a 1-trillion-parameter open-weight model
Large 4 ('Le Chonk') runs 49 billion active parameters via MoE, available now only as a hosted API at $1.36/$4.18 per million tokens; open weights follow around October 27. Let's Data Science · officechai
-
StepFun's Step 5 Preview (600B MoE, 1M context) went live as a hosted preview Oct 8; open weights still due Oct 15. LLM Reference