Matt Turck shared his conversation with Mistral research lead Pierre Stock about Le Chonk. According to the conversation, Le Chonk is described as the largest open-source model in the West, trained by the team on 3,800 GPUs in their own data center, and is still in training. Stock also discussed what Mistral borrowed from DeepSeek, and the trade-off between open-source weights and cybersecurity refusals. The program also mentioned that Mistral Large 5 will be released later.
On October 21, 2026, LMSYS published a blog post introducing Elastic EP Scaling (elastic expert parallelism scaling) in SGLang. This mechanism allows MoE deployments to add GPUs in fine-grained increments without requiring a full replica or restart; in the example, it scales from 20 GPUs to 24 GPUs. At startup, max_ep_size reserves metadata and communication buffer capacity, and new ranks start on demand. Tested with DeepSeek-V3-Lite FP8 on 8×H100, the first completed request was 5.3×–8.2× faster than with a restart. LMSYS also noted that after a rank joins the Mooncake WORLD, automatic rollback is not currently supported, and a failure during finalization may require restarting the deployment. There are currently no subsequent developments.
restart; in the example, it scales from 20 GPUs to 24 GPUs. At startup, max ep size reserves metadata and communication buffer capacity, and new ranks start on demand. Tested with DeepSeek-V3-Lite FP8 on 8×H100, the first completed request was 5.3×–8.2× faster than with a restart. LMSYS also noted that after a rank joins the Mooncake WORLD, automatic rollback is not currently supported, and a failure during finalization may require restarting the deployment. There are currently no subsequent dev
On October 21, 2026, Modal released Modal Dojo, an LLM training library. According to Modal's official engineering blog, the library provides tuned default recipes, modifiable parameters, and an out-of-the-box metrics observability dashboard, and supports models ranging from 500 million to trillions of parameters; it is built on open-source frameworks such as Miles and Slime, can load datasets from Hugging Face, and lets agents query run status, logs, and trajectories via a CLI. The same source also states that Axiom Bio has used it to launch training of DeepSeek V4.1 Flash with a 64k context, running on 64 B300s. There are currently no subsequent independent developments.
eters, and an out-of-the-box metrics observability dashboard, and supports models ranging from 500 million to trillions of parameters; it is built on open-source frameworks such as Miles and Slime, can load datasets from Hugging Face, and lets agents query run status, logs, and trajectories via a CLI. The same source also states that Axiom Bio has used it to launch training of DeepSeek V4.1 Flash with a 64k context, running on 64 B300s. There are currently no subsequent independent developments.
As a unified API gateway, Flatkey supports calling multimodal models in Codex and Claude Code via CLI. According to a report dated October 21, 2026, the gateway can call text models such as GPT, Claude, DeepSeek, GLM, and Kimi, as well as image and video models such as GPT Image and Seedance; users can access 100+ official models and 1,000+ tool APIs with a single Key, with costs consolidated into one bill. The report also states that its everyday prices are at least 20% lower than official pricing; during the October 19–23 promotion, GPT, GLM 5.2, Kimi K3, and DeepSeek are 40% off, Claude is 20% off, and Seedance 2.0 is 15% off. There are currently no subsequent developments.
y, Flatkey supports calling multimodal models in Codex and Claude Code via CLI. According to a report dated October 21, 2026, the gateway can call text models such as GPT, Claude, DeepSeek, GLM, and Kimi, as well as image and video models such as GPT Image and Seedance; users can access 100+ official models and 1,000+ tool APIs with a single Key, with costs consolidated into one bill. The report also states that its everyday prices are at least 20% lower than official pricing; during the October
The RadixArk SGLang team has open-sourced SGLang-TorchTPU. According to information published on the LMSYS Blog on October 20, 2026, the team, together with Google and Meta, launched this PyTorch-based native backend, bringing SGLang's full serving stack to Google Cloud TPU, and it has already been used in Meta's large-scale production environment. Initial support covers models including Muse Spark Open, GLM-5.3, Kimi K3, DeepSeek V4, Qwen3.8, and Gemma 4, spanning features such as speculative decoding, KV cache offloading, PD disaggregation, and FP8 KV cache. The team also used TorchTitan and Miles to complete GRPO training of Qwen3-4B on a single 4-chip TPU v7x, with the reward curve consistent with a 6-card H200 baseline within the range of random fluctuation. There are currently no subsequent development facts.
ng stack to Google Cloud TPU, and it has already been used in Meta's large-scale production environment. Initial support covers models including Muse Spark Open, GLM-5.3, Kimi K3, DeepSeek V4, Qwen3.8, and Gemma 4, spanning features such as speculative decoding, KV cache offloading, PD disaggregation, and FP8 KV cache. The team also used TorchTitan and Miles to complete GRPO training of Qwen3-4B on a single 4-chip TPU v7x, with the reward curve consistent with a 6-card H200 baseline within the r
On October 20, 2026, Zyphra released the second generation of inference optimizations, ZOE, adding it to the ZOE inference optimization series. According to information published by Zyphra Research, on AMD Instinct MI355X, GLM-5.3's time to first token (TTFT) at 512K context dropped from 178.8 seconds to 64.6 seconds, about 2.8x; at 90% prefix cache hit rate, it dropped from 42.3 seconds to 11.8 seconds, about 3.6x. Its CSA top-K selection HIP kernel raised DeepSeek-V4-Pro's single-user decoding throughput at 1M context from 43.5 to 61.1 tokens/s, about 1.40x. GLM-5.3's AMD MXFP4 quantization compressed the model from 756 GB to 410 GB, with benchmark scores such as AIME25 within 3 points of FP8. Currently this release is the only development with evidence, and there are no subsequent developments yet.
d from 178.8 seconds to 64.6 seconds, about 2.8x; at 90% prefix cache hit rate, it dropped from 42.3 seconds to 11.8 seconds, about 3.6x. Its CSA top-K selection HIP kernel raised DeepSeek-V4-Pro's single-user decoding throughput at 1M context from 43.5 to 61.1 tokens/s, about 1.40x. GLM-5.3's AMD MXFP4 quantization compressed the model from 756 GB to 410 GB, with benchmark scores such as AIME25 within 3 points of FP8. Currently this release is the only development with evidence, and there are n
Michael Levitt, winner of the 2013 Nobel Prize in Chemistry, predicted at the Second International Conference on AI Scientists (ICAIS 2026) that by 2056, young scientists aided by AI agents may accomplish in a single year the work his generation did over a lifetime. He also said this prediction may not be aggressive enough, and that the efficiency gain could far exceed 30 times. Levitt believes AI will erase the age gap among scientists, and stressed that the coexistence of multiple models such as DeepSeek, Kimi, Claude, and ChatGPT is a safeguard for humanity's safe use of AI. Currently, this prediction is reported by a single source, with no further independent evidence or subsequent developments.
a single year the work his generation did over a lifetime. He also said this prediction may not be aggressive enough, and that the efficiency gain could far exceed 30 times. Levitt believes AI will erase the age gap among scientists, and stressed that the coexistence of multiple models such as DeepSeek, Kimi, Claude, and ChatGPT is a safeguard for humanity's safe use of AI. Currently, this prediction is reported by a single source, with no further independent evidence or subsequent developments.
Artificial Analysis released its evaluation on October 19, 2026; MiniMax-M3.1-Flash scored 42 on its Intelligence Index, 13 points higher than MiniMax-M3, at less than half the latter's cost per task. The same evaluation shows the model scored 48% on Humanity's Last Exam and 22% on CritPt, both higher than GLM-5.3-Flash and Gemini 3.8 Flash, and slightly lower than DeepSeek V4.1 Flash 1017. The model has a 1M-token context window, supports text and image inputs, is priced at $0.15/$0.60 per 1M input/output tokens, and its weights are currently closed-source; MiniMax announced that it will open the weights on October 23.
nts higher than MiniMax-M3, at less than half the latter's cost per task. The same evaluation shows the model scored 48% on Humanity's Last Exam and 22% on CritPt, both higher than GLM-5.3-Flash and Gemini 3.8 Flash, and slightly lower than DeepSeek V4.1 Flash 1017. The model has a 1M-token context window, supports text and image inputs, is priced at $0.15/$0.60 per 1M input/output tokens, and its weights are currently closed-source; MiniMax announced that it will open the weights on October 23.
Prime Intellect released prime-rl v0.10.0 and verifiers v0.4.0 on October 19, 2026. According to its release information, prime-rl adds the reinforcement learning algorithm IPO and sets it as the default, replacing the previous DPPO-based default loss. In tests comparing Qwen3-235B-A22B SFT against torchtitan, prime-rl achieved a 1.15× speedup (3732 vs 3253 tokens/s/GPU); the MFU of GLM-5.3 full SFT increased from 22.3% to 40.1%. The evaluation CLI has moved from verifiers into prime-rl, and support has been added for models including DeepSeek V4 Flash, Qwen3.8 Flash Next, and Nemotron-H. There are currently no subsequent independent development facts.
ng algorithm IPO and sets it as the default, replacing the previous DPPO-based default loss. In tests comparing Qwen3-235B-A22B SFT against torchtitan, prime-rl achieved a 1.15× speedup (3732 vs 3253 tokens/s/GPU); the MFU of GLM-5.3 full SFT increased from 22.3% to 40.1%. The evaluation CLI has moved from verifiers into prime-rl, and support has been added for models including DeepSeek V4 Flash, Qwen3.8 Flash Next, and Nemotron-H. There are currently no subsequent independent development facts.
Tenstorrent closes $1.5 billion funding round at a $6.2 billion valuation. According to a TechCrunch AI report on October 20, 2026, the round was led by Trace Capital and Innovation X, with participation from Cadence Ventures, Eclipse, Fidelity, SanDisk, and T. Rowe Price; the valuation is more than double the $2.6 billion at the end of 2024. The report said CEO Jim Keller stated that after demonstrating that the company's Galaxy servers can run DeepSeek models, the company decided to remain independent and continue raising funds, and that the new capital will be used to expand chips for inference on open-weight models and video generation. There are currently no subsequent developments.
ipation from Cadence Ventures, Eclipse, Fidelity, SanDisk, and T. Rowe Price; the valuation is more than double the $2.6 billion at the end of 2024. The report said CEO Jim Keller stated that after demonstrating that the company's Galaxy servers can run DeepSeek models, the company decided to remain independent and continue raising funds, and that the new capital will be used to expand chips for inference on open-weight models and video generation. There are currently no subsequent developments.
On October 20, 2026, Joseph Sifakis, winner of the 2007 Turing Award, called for the establishment of global standards and regulatory standards in the field of AI at the "2026 Sustainable Global Leaders Conference," saying this is a shared responsibility of humanity. Citing ChatGPT and DeepSeek outperforming humans on medical exams as an example, he pointed out that large models do not understand social norms and value systems and therefore cannot replace doctors. He also warned that large companies may use regulatory gaps to shirk responsibility and seek commercial gain. At present, this call has been reported by a single source, and there are no subsequent factual developments.
l standards and regulatory standards in the field of AI at the "2026 Sustainable Global Leaders Conference," saying this is a shared responsibility of humanity. Citing ChatGPT and DeepSeek outperforming humans on medical exams as an example, he pointed out that large models do not understand social norms and value systems and therefore cannot replace doctors. He also warned that large companies may use regulatory gaps to shirk responsibility and seek commercial gain. At present, this call has be
On October 20, 2026, Memory Tensor open-sourced and released the MemT 1.0 series of memory-native models. The series includes two models: Nano has 36B parameters with 3B activated, based on Qwen3.6-35B-A3B; Flash has 288B parameters with 17B activated, based on DeepSeek-V4-Flash. The models add a memory branch parallel to attention in selected Transformer modules, encoding experience as updatable matrix states that participate in subsequent generation. On 500 SWE-bench Verified tasks, Nano achieved an average success rate of 69.73% at key checkpoints, versus 66.07% for the no-experience baseline; under a 128K long history, end-to-end inference speed is approximately 1.86 times that of the full-context approach, and the persistent state footprint per trajectory is one sixty-sixth of the full-context KV Cache. There are currently no subsequent development facts.
f memory-native models. The series includes two models: Nano has 36B parameters with 3B activated, based on Qwen3.6-35B-A3B; Flash has 288B parameters with 17B activated, based on DeepSeek-V4-Flash. The models add a memory branch parallel to attention in selected Transformer modules, encoding experience as updatable matrix states that participate in subsequent generation. On 500 SWE-bench Verified tasks, Nano achieved an average success rate of 69.73% at key checkpoints, versus 66.07% for the no
Perplexity unveils post-training method JOLT, letting the same model serve as both teacher and student, with the teacher also participating in training so that feedback is always available. Perplexity's release states that on MATH, JOLT reaches accuracy comparable to GRPO with 6.5 times fewer training tokens. A subsequent MarkTechPost report states that the method was proposed by Perplexity and NVIDIA, named JOLT (Joint On-policy Learning and Teaching); JOLT+ achieves a 72.09% state-test pass rate on AppWorld, 9.9 percentage points higher than GRPO's 62.20%. The report states that testing covered Qwen3-0.6B, Qwen3-4B-Instruct, Qwen3-8B, Qwen3-30B-A3B-Instruct and DeepSeek-V2-Lite, and that the code has not yet been released.
LT reaches accuracy comparable to GRPO with 6.5 times fewer training tokens. A subsequent MarkTechPost report states that the method was proposed by Perplexity and NVIDIA, named JOLT (Joint On-policy Learning and Teaching); JOLT+ achieves a 72.09% state-test pass rate on AppWorld, 9.9 percentage points higher than GRPO's 62.20%. The report states that testing covered Qwen3-0.6B, Qwen3-4B-Instruct, Qwen3-8B, Qwen3-30B-A3B-Instruct and DeepSeek-V2-Lite, and that the code has not yet been released.
QuestMobile releases the 2026 Q3 AI application market report. The report shows that in September, Doubao ranked first among AI-native Apps with 419 million MAU, while Qwen and DeepSeek ranked second and third with 158 million and 123 million respectively. In the office Agent ranking, Tencent WorkBuddy topped the list with 32.353 million MAU, while Baidu Dazi and Qwen Office ranked second and third with 9.77 million and 7.905 million respectively. In September, the AI office track reached 142 million MAU, up 50.6% year-on-year, and the PC client user scale of office Agents grew 512.2% year-on-year. Other reports say that under Analysys's desktop measurement, WorkBuddy's monthly visits increased from 20.97 million in June to 85.34 million in September.
QuestMobile releases the 2026 Q3 AI application market report. The report shows that in September, Doubao ranked first among AI-native Apps with 419 million MAU, while Qwen and DeepSeek ranked second and third with 158 million and 123 million respectively. In the office Agent ranking, Tencent WorkBuddy topped the list with 32.353 million MAU, while Baidu Dazi and Qwen Office ranked second and third with 9.77 million and 7.905 million respectively. In September, the AI office track reached 142 mi
On October 20, 2026, Memory Tensor released and open-sourced the MemT 1.0 series of memory-native large models. According to a report by QbitAI, the series includes MemT-1.0-Nano, with 36B parameters and 3B activated (based on Qwen3.6-35B-A3B), and MemT-1.0-Flash, with 288B parameters and 17B activated (based on DeepSeek-V4-Flash); a memory computation pathway running in parallel with attention was added to selected Transformer modules. The team built a 34-billion-token memory training corpus, of which 20 billion tokens were used for continued pre-training and 14 billion for post-training. In a comprehensive evaluation across 11 memory benchmarks, Flash and Nano averaged scores of 50.6 and 47.0, respectively, ranking first and second among the parametric memory methods evaluated; on 500 SWE-bench Verified tasks, the proportion of tasks that Nano solved successfully in all three attempts rose from 54.00% on the no-experience baseline to 62.80%. Subsequent reports said that MemT-1.0 connects a native memory module in parallel within the Transformer, turning memory into a state in the forward computation stream, and released up to 288B weights and a 66-page technical report; the authors said that under a 256K ultra-long history, memory usage dropped from 5.06G to a constant 40M, and KV cache was reduced
by QbitAI, the series includes MemT-1.0-Nano, with 36B parameters and 3B activated (based on Qwen3.6-35B-A3B), and MemT-1.0-Flash, with 288B parameters and 17B activated (based on DeepSeek-V4-Flash); a memory computation pathway running in parallel with attention was added to selected Transformer modules. The team built a 34-billion-token memory training corpus, of which 20 billion tokens were used for continued pre-training and 14 billion for post-training. In a comprehensive evaluation across
Artificial Analysis has launched leaderboards for OpenAI's two health dialogue benchmarks, HealthBench and HealthBench Professional. HealthBench contains 5,000 conversations with general users and health workers, while HealthBench Professional consists of 525 conversations all aimed at clinicians. The leaderboard results show that on HealthBench Professional, GPT-6 Astra (max) ranks first with a length-adjusted score of 66.0%, followed by GPT-6.1 Sol (max) at 65.1% and Claude Sonnet 5.5 (max with fallback) at 63.0%; on HealthBench, Claude Sonnet 5.5 (max with fallback) leads with 62.1%, while DeepSeek V4.1 Flash (max) scores 60.9%. In addition, Claude Opus 5.5 and Claude Fable 5.1 give average answers of about 5,900 characters, losing more than 11 points after length adjustment. There are currently no further developments.
n HealthBench Professional, GPT-6 Astra (max) ranks first with a length-adjusted score of 66.0%, followed by GPT-6.1 Sol (max) at 65.1% and Claude Sonnet 5.5 (max with fallback) at 63.0%; on HealthBench, Claude Sonnet 5.5 (max with fallback) leads with 62.1%, while DeepSeek V4.1 Flash (max) scores 60.9%. In addition, Claude Opus 5.5 and Claude Fable 5.1 give average answers of about 5,900 characters, losing more than 11 points after length adjustment. There are currently no further developments.