Filters
- TextCLF launches TQ, calibration-free 4-bit quantization with an OpenAI-compatible API
TextCLF says its Quant Factory converts pretrained models to TQ 4-bit models without a calibration dataset; the surrounding pipeline is open source, while the quantization and inference core is closed. Models run via an OpenAI-compatible API or self-hosted with Docker and vLLM, and TextCLF says Qwen3.8 27B shrank 4.1x while keeping 99.6% of FP16 accuracy, a figure not independently verified.
OfficialSource: TextCLF on X - OpenAI fixes a bug that degraded image understanding in GPT-6 Sol and GPT-6 Luna
OpenAI Developers says the fix should improve vision tasks, including computer use, in the API and Codex.
- OpenAI shuts down gpt-3.5-turbo-instruct, babbage-002 and davinci-002 today
OpenAI's deprecations page lists September 28 as the shutdown date for gpt-3.5-turbo-instruct, gpt-3.5-turbo-1106, babbage-002 and davinci-002, and recommends moving to gpt-5.6-terra. The shutdown is part of the legacy GPT snapshot retirement OpenAI announced in September 2025.
OfficialSource: OpenAI deprecations - Huawei open-sources openPangu-2.0 training and RL code, per iFeng
iFeng reports that Huawei on September 28 open-sourced the pre-training, SFT and post-training RL code for openPangu-2.0, and that the model's components are going live on open-source platforms.
- H Company releases Holo4 computer-use agent models in 27B and 35B-A3B sizes, plus Holotron4 Nano
H Company's Holo4 series, a 27B dense model and a 35B-A3B mixture-of-experts model, operates desktops, the web and Android through GUIs, code, MCP and APIs, and is available on the H Models API with open weights on Hugging Face. H Company says Holo4-27B scores 61.7% on OSWorld 2.0 at lower cost than most frontier closed models; it also released Holotron4 Nano, post-trained from NVIDIA's Nemotron 3 Nano Omni.
OfficialSource: H Company newsroom - Fireworks' Ember-1, post-trained on Kimi K3, is available in Cline
Cline says Fireworks Research's Ember-1, post-trained on Kimi K3, is now available in Cline. Cline says it matches K3 on benchmarks while using about 40% fewer tokens, and in an online A/B test on coding traffic used 71% fewer reasoning tokens at the same success rate. Fireworks lists pricing of $3 per million input tokens and $15 per million output tokens.
OfficialSource: Cline release - ElevenLabs releases Eleven v4 and v4 Turbo speech models with support for 90+ languages, available in its API
ElevenLabs has released Eleven v4, a text-to-speech model it says produces more expressive speech, along with a low-latency variant, Eleven v4 Turbo. Both models support more than 90 languages, can clone a voice from 10 seconds of audio and are available now through the ElevenLabs API. ElevenLabs says its own testing puts v4 Turbo's median inference latency at about 100 ms and its median time from request to audible speech at about 150 ms. The company also says Eleven v4 ranks first on an Artificial Analysis leaderboard and was preferred by about 75% of listeners in blind tests against competing models it selected.
OfficialSource: ElevenLabs - Web-extraction benchmark: a "Do not guess" instruction cut made-up fields from 70.7% to 20.2%
Earn an Honest Dollar tested extraction services and models on 42 paired pages across seven page types for fields invented when missing from the page, and says adding "Do not guess" cut the made-up rate from 70.7% to 20.2%. In checking returned values, GPT-6 Luna caught 20 of Firecrawl's 24 invented values; the results are the publisher's own tests.
OfficialSource: Earn an Honest Dollar - 500M VLM answers typed questions about images with calibrated confidence
A Reddit user trained a 500M vision-language model that answers typed questions about an image — choice, score or yes-no — with calibrated confidence.
- Community model claims better calibration than Jev with multi-modal ability
A Reddit post claims a decision model that is better, faster and more calibrated than Jev, with multi-modal ability.
- Basis says GPT-6 Astra completes tax workbook 2x faster than GPT-5.6 Sol
Basis reports GPT-6 Astra completed a 50-tab tax workbook twice as fast as GPT-5.6 Sol, with stronger understanding of user intent giving the team more confidence in real-world use.
- MarkTechPost verifies 20 agentic use cases of TypeSafe AI's Jev, per MarkTechPost
MarkTechPost reports Jev, which skips text generation and returns typed decisions with calibrated probabilities, costs $0.042 per million input tokens with free output, and it verified 20 agentic use cases from model routing to tool-call gating.
- Anonymous model Space Bunny tops OpenRouter and OpenCode daily usage charts, per QbitAI
QbitAI reports that the stealth model Space Bunny reached No. 1 on both OpenRouter's and OpenCode's daily usage rankings within days, and ran its own coding tests through the OpenRouter API. No vendor has claimed the model, and pricing has not been published.
- NaiveAI releases Naive-N0.5-Flash, a 309B open-weights MoE with 1M context under MIT
NaiveAI's Hugging Face model card describes a 309B-parameter MoE with 15.5B active, aimed at coding and AI research, that reaches a native 1M-token context by mixing sliding-window attention with lightweight DeepSeek Sparse Attention. Weights and inference code are MIT-licensed, and the card lists planned API pricing of $0.10 per million input tokens and $0.40 per million output tokens, without a launch date.
OfficialSource: Hugging Face - A new Google Flow build reportedly references "Nano Banana 2.1"
TestingCatalog says a new build of Google Flow on the web labels a possible Nano Banana update as 2.1, replacing an earlier reference to Nano Banana 2.5 Flash. Google has not announced the model, and there are no specs or release timing.
- Paper's own Table 2 says MoGe-2 pseudo-label training mostly fails to beat MoGe-2
A paper discussed on Reddit found, in its own Table 2, that a model trained partly on MoGe-2 pseudo-labels mostly fails to beat MoGe-2.
- Suochen Tech and Meimeng Space release embodied models and physics benchmark, per QbitAI
QbitAI reports Suochen Technology and its strategic investor Meimeng Space jointly released embodied models and physics evaluation standards, as world models become a new narrative for embodied AI commercialization.
- 700B-parameter GLM runs on a laptop with SSD as memory, per QbitAI
QbitAI reports Colibri, a fast-growing open-source project on GitHub, runs 700-billion-parameter GLM models on laptops without a GPU by using SSD storage as memory.
- Supersonic Labs releases Julia 1, a 144.3M-parameter decision model that runs on CPU, per MarkTechPost
MarkTechPost reports Julia 1, built on mmBERT-small, takes context, a question and 2 to 20 options and returns one choice with probabilities; the 144.3M-parameter model runs on a CPU.
- Sarvam AI releases Saaras V4 speech-to-text for all 22 Indian languages, per MarkTechPost
MarkTechPost reports Sarvam AI's Saaras V4 covers all 22 Indian languages plus global English, pairing an audio encoder with a 3B hybrid state-space decoder, adding keyterm prompting for up to 50 terms and 5 output modes.