⚙ How GLM-5.3-Flash Served 70 Trillion Free Tokens on Chinese Chips GLM-5.3-Flash's anonymous "Ox Alpha" trial burned through roughly 70 trillion tokens in a week — and Zhipu says all of it ran on Chinese AI chip clusters. For many observers, that is a bigger story than the model itself. Zhihu contributor 恋猫 breaks down the systems engineering that made it work. His framing: the model architecture first reduces how much data needs to move, then the inference system compresses what remains. The reported scale is around 100,000 Chinese chips — from Huawei, Moore Threads and Hygon, according to @latepostnews — though @Zai_org itself has only said "tens of thousands." 1️⃣ The architecture cuts data movement first Versus GLM-5.3, Flash lowers attention compute by 3.01x and KV cache by 4.44x. Less data shuttling between memory and compute is the foundation everything else builds on. 2️⃣ Compute-for-bandwidth: trading FLOPs for HBM relief The bottleneck on these chips is HBM: the compute units still have headroom while memory bandwidth is nearly saturated. The fix, loosely speaking, turns "every step: read the full state, write the full state" into "read the state, recompute a little, write back periodically." A bit of extra matrix math buys a large reduction in HBM write traffic. 3️⃣ Communication-for-bandwidth: shard the cache across cards The cluster also uses high-speed inter-chip links and aggregated bandwidth to cut how much data each card must keep resident. The author's example: rank 0 holds the KV and indexer cache for some layers, rank 1 holds the rest. No card stores every layer's cache long-term; data is prefetched over the interconnect as each layer executes. The trade-off is real — more communication, far less resident cache per card. 4️⃣ EPD separation: three pools, scaled independently Finally, Encode, Prefill and Decode are split into three independently scalable resource pools, which matters most for multimodal traffic. Together, these pieces are what let the cluster run at high utilization. 5️⃣ The result: 3x end-to-end, cost "comparable to NVIDIA" Zhipu's own claim: end-to-end serving performance improved 3x on the same hardware, bringing per-token cost close to mainstream NVIDIA GPUs. The author's reading: individual Chinese cards may still be weaker, but model-system co-design lets the cluster as a whole reach international-mainstream throughput and cost. He adds a wry footnote: Zhipu's API used to be notorious among Chinese developers for 429 rate-limit errors. That it could absorb this launch's traffic at all — entirely on Chinese silicon — is, in his words, proof that optimization never ends. 🔗 Full Reading: #GLM# #Zhipu# #AIChips# #AIInfra# #LLM# #Inference# #OpenWeights#

💥 GLM-5.3-Flash: Reported Opus 4.8-Level Scores at One-Tenth the Price Zhipu has confirmed that Ox Alpha — the anonymous model that dominated OpenRouter and OpenCode last week — is GLM-5.3-Flash, a 320B-parameter MoE with 18B active, released with open weights and an API at roughly one-tenth of GLM-5.3's price. Zhihu contributor 小小将 opens with a confession: he had hoped Ox Alpha was an external team fine-tuning the open GLM weights. He was wrong — it was Zhipu's own model all along. His broader take: at this performance and this price, Flash becomes the new "kill line" for large models — the bar below which rivals simply get priced out. 1️⃣ The numbers behind the "kill line" Flash is somewhat larger than DeepSeek-V4 Flash, at 320B total and 18B active. On Zhipu's own benchmark comparisons, overall capability is roughly level with Claude Opus 4.8. More reference points he cites: 🔹 57 on the Artificial Analysis Intelligence Index — around GPT-5.6 Terra's level, slightly below GLM-5.3. 🔹 No.5 on Code Arena, one spot above GLM-5.3. 🔹 API at 1/10 of GLM-5.3's price, with a limited-time half-off promo bringing it to 1/20. 2️⃣ Same DeepSWE score, a fraction of the cost The author's sharpest comparison is on DeepSWE, where Flash scores 63% at a single-task cost of $0.24. DeepSeek-V4-Pro hits the same score at $1.67 per task — roughly seven times the cost for equivalent results. This is the author's cost arithmetic on reported figures, not an independent measurement. 3️⃣ The architecture that cuts the bill Part of the price drop is structural. Flash uses a hybrid of linear and sparse attention, sharply cutting attention compute, and a new IndexPool that compresses the indexer's cache from four copies to one — reducing latency and memory at million-token context. Versus GLM-5.3, the company reports attention compute down to about 1/3 and KV cache down to about 1/4.4. 4️⃣ The compute mystery, answered by domestic chips One reason few believed Ox Alpha was Zhipu's: the company was not thought to have enough spare compute for a massive free public test. The answer, per Zhipu: Flash is served from domestic Chinese chip clusters. The team built a custom inference engine on SGLang and worked around limited VRAM and bandwidth with quantization, layered deployment, and trading compute for communication — lifting end-to-end serving performance about 3x on the same hardware, with per-token cost now close to mainstream NVIDIA GPUs. 5️⃣ Chinese silicon just passed its biggest stress test The author's conclusion: Chinese chip clusters have shown they can carry large-scale inference for a frontier-level model — including a free, record-breaking public trial. If that holds, he argues, the outlook for Chinese models just got a lot brighter. 🔗 Full Reading: #GLM# #Zhipu# #GLM5Flash# #AIChips# #LLM# #AIInfra# #OpenWeights#

🧩 Qwen3.8-Flash-Next: A 6B-Active Preview of the Qwen4 Architecture On the same night Zhipu's GLM-5.3-Flash took over the timeline, @Alibaba_Qwen open-sourced Qwen3.8-Flash-Next — explicitly positioned as a preview of the Qwen4 architecture. Zhihu contributor Kitt在进化 argues that for people who actually run models locally, this is the more practical release of the night. The local-deployment groups he is in are, in his words, on fire. His take: the model is called 3.8, but the architecture is really Qwen4 in preview — and it borrows the best ideas from across the field. 1️⃣ Smaller, cheaper, and realistic for local deployment Flash-Next has 125B total parameters with only 6B active — far smaller than the 300B-class GLM-5.3-Flash. The API is priced at ¥1 input, ¥3 output, and ¥0.1 per cached million tokens, roughly matching DeepSeek-V4-Flash's off-peak rates. His comparison, per 1M tokens in RMB — input / output / cache hit: 🔹 Qwen3.8-Flash-Next: 1.0 / 3.0 / 0.1 🔹 GLM-5.3-Flash: 0.4 / 1.4 / 0.115 (limited-time promo rate) 🔹 DeepSeek-V4-Flash (off-peak): 1.5 / 4.5 / 0.05 2️⃣ A Qwen4 preview wearing a Qwen3.8 name The author points out this is the same play Zhipu made: GLM-5.3-Flash's architecture is also completely different from GLM-5.3. Shipping the new architecture as open weights early is deliberate pathfinding. Inference frameworks like vLLM and SGLang, plus the quantization toolchain, all need lead time to adapt before Qwen4 proper arrives. 3️⃣ An architecture that borrows from everyone The author reads Flash-Next as a synthesis of the field's best recent ideas. 🔹 Attention: Qwen's in-house QSA, built to balance throughput and speed on long context. 🔹 Knowledge: it absorbs DeepSeek's Engram line of work, packing prior knowledge into 51B of N-gram side parameters — high knowledge density at minimal compute cost. 🔹 Training: the Muon optimizer, popularized by Kimi, scheduled together with AdamW. 4️⃣ Half the active parameters, still overtaking At 6B active, Flash-Next is nearly half the size of DeepSeek-V4-Flash — yet Qwen's reported benchmarks show it overtaking that model on multiple coding and agent leaderboards. The author says he is not worried about real-world experience: the recent Qwen3.8-27B already proved itself in daily use, and this sits on the same foundation. 🔗 Full Reading: #Qwen# #Alibaba# #Qwen4# #OpenWeights# #LLM# #AIInference# #MoE#

⚔ GLM-5.3-Flash Aims Straight at the Post-Hike DeepSeek Zhipu's GLM-5.3-Flash — revealed this week as the anonymous "Ox Alpha" — has been open-weighted and priced at roughly one-tenth of GLM-5.3. Much of the early discussion compares it to DeepSeek's V4 Flash, which recently raised prices. Zhihu contributor 起步十档, who ran Ox Alpha inside real workflows before the reveal, gives a practitioner's verdict in one line: it is built to kill the post-hike DeepSeek. His case rests on three legs — performance, token efficiency, and an architecture change that makes the price possible. 1️⃣ It clears the bar for long-horizon work Official scores put Flash between Grok 4.6 and GLM-5.3, and clearly ahead of DeepSeek V4 Flash. In the author's own testing, its frontend ability roughly matches an early gray-test build of DeepSeek V4 Pro, while its backend is noticeably weaker than GLM-5.3 — but still usable on long-horizon tasks as long as the connection holds. His rule of thumb: any model past the Claude Opus 4.6 line is workflow-ready for long tasks. Beyond that, differences come down to reasoning style and accuracy, not viability. One caveat he flags: during the anonymous test the deployment was unstable, and some believe it served a mid-training checkpoint rather than the final model. 2️⃣ The real weapon: token efficiency Comparing peak API prices against the post-hike DeepSeek V4 Flash, the author notes cached input is actually 2x more expensive, while regular input and output sit at roughly 26% of DeepSeek's price. Since cached input is often the bulk of the bill, he wants real-world tests before calling a winner on price alone. But his own usage points the same direction. In one to two hours of real work — reading and writing files, running tests — Ox Alpha burned barely over 100K tokens. He estimates DeepSeek would need 250-300K for the same workload. His prediction: same tasks, run on both APIs, will come out cheaper on Flash — with clearly better performance. 3️⃣ His unexpected advice: skip the Coding Plan The plan only triples your quota for Flash. Using Zhipu's own best-case math — maximum usage, off-peak hours, the official 0.8x API-equivalent rate — the plan works out to about 53% of pay-as-you-go API cost. Since that scenario is already extreme, he concludes the API is the better deal for almost everyone. 4️⃣ The architecture change behind the price From GLM-5 through 5.3, Zhipu used DSA — essentially an optimized full attention — which costs more than DeepSeek's CSA/HCA, Kimi's KDA, or Qwen's linear-global hybrid. That is why GLM used to be pricier than larger DeepSeek models. Flash is the first GLM to switch to linear attention plus an HCA-like compressed attention, trained with the HCA recipe as well. That brings it in line with mainstream domestic practice — and the price fell accordingly. The author expects a future GLM-5.5 can scale up without costing much more than 5.3. On top of that, Zhipu added native multimodality and leaned on domestic compute, which he reads as the reason for the generous free quotas during the anonymous test. 5️⃣ Zhipu is still the team to beat The author's closing line is unambiguous: Zhipu remains, in his words, the number-one Chinese model company. That is his judgment, not a benchmark result — but the cost argument underneath it is now easy to check yourself. 🔗 Key links: Official announcement: Open weights (MIT): 🔗 Full Reading: #GLM# #Zhipu# #DeepSeek# #LLM# #AIInference# #TokenEfficiency# #OpenWeights#

🔌 GLM-5.3-Flash Served Its Viral Debut Entirely on Domestic Chinese Chips Zhipu's GLM-5.3-Flash — the 320B-A18B model revealed this week as the anonymous "Ox Alpha" — set usage records on OpenRouter and OpenCode during its undercover test. The company says all of that traffic was served by domestic Chinese chip clusters. Zhihu contributor 刘延 reconstructs how Zhipu lined up this infrastructure, and reads the official engineering details for hints about which chips are actually doing the work. The core judgment: the bigger story here is not the model itself, but that a frontier-level model handled global-scale, real-world inference on domestic silicon. 1️⃣ The timeline behind the launch The author pieces together a sequence from public reporting. 🔹 Zhipu was reported to have acquired an infrastructure company. 🔹 LatePost reported Zhipu had brought 50,000 domestic cards online; around the same time, its CodePlan subscription got cheaper with generous bonus quotas. 🔹 Ox Alpha went live anonymously and, in the author's words, blew up worldwide. 🔹 Zhipu then confirmed every request in that test ran on domestic chips. 🔹 The latest LatePost report puts the deployment at 100,000 domestic cards. Note the card counts come from media reports, not Zhipu itself. 2️⃣ The engineering: surviving 1M context on constrained hardware Zhipu's own statement is unusually specific about the constraints. The main bottleneck on these chips is memory capacity and bandwidth, and supporting a 1M-token context is the hardest part. The company's listed optimizations include trading compute for bandwidth and communication for memory, intra-node tensor parallelism for the linear attention and the LM head, ReplaySSM, W8A8 quantization, INT8/FP8/BF16 mixed cache quantization, and Layer Split. 3️⃣ Which chips? Reading the precision hints Here the author speculates, and it should be read as inference, not confirmation. 🔹 FP8 support suggests Moore Threads could be handling prefill, or possibly Hygon's DCU-3. 🔹 INT8 points toward Ascend 910B/C as the likely backbone. 🔹 No mention of FP4 suggests the newer Ascend 950 is probably not in the mix. 4️⃣ Why this matters If the reporting holds, this is the first time domestic Chinese chip clusters have carried a frontier model's global production traffic at this scale — including a free, record-breaking stress test from developers worldwide. The author treats it as a proof point: China's domestic chips are no longer just for training experiments or internal pilots, but can serve a top-tier model to the open internet. 🔗 Key links: Official announcement: Open weights (MIT): 🔗 Full Reading: #GLM# #Zhipu# #AIChips# #AIInfra# #Ascend# #Semiconductors# #OpenWeights#

⚡ GLM-5.3-Flash: Near-Flagship Logic at One-Tenth the Price Zhipu has confirmed that Ox Alpha — the anonymous model that just topped usage charts on OpenRouter and OpenCode — is GLM-5.3-Flash, a 320B-parameter MoE with 18B active, released with open weights. The company prices its API at roughly one-tenth of GLM-5.3, and reports an Artificial Analysis Intelligence Index score of 57, on par with Claude Opus 4.8. Zhihu contributor toyama nao, known for a long-running monthly logic benchmark built on self-designed problem sets, argues the launch fills a gap the market has had since DeepSeek raised prices: a model good enough to use daily and cheap enough to ignore. The core judgment: GLM-5.3-Flash is not a capability breakthrough. It is a cost breakthrough — same lineage as GLM-5.3, nearly identical results on many tasks, but a new path on inference efficiency. 1️⃣ Why the market needed this model After DeepSeek's price hike, its tier lost any clear price-performance leader. GLM-5.3 then improved quality without raising prices and quietly took that spot. Given compute costs in China, shrinking the model is the pragmatic route to genuinely low prices — and smaller models have already proven they can carry real workloads. What the market lacked was a model cheap enough that cost stops being a decision. Flash is that option. 2️⃣ Where Flash matches GLM-5.3 — and where the floor drops In the author's monthly logic evaluation, Flash matches the standard GLM-5.3 on most low- and mid-difficulty tasks, including coding. On hard tasks it can still reach the same ceiling — just not reliably. In practice that means more retries to get the best answer. Retries are cheap at Flash's pricing, but output speed hasn't improved, so the experience still degrades. 3️⃣ Hallucination: wider variance in both directions Flash's hallucination behavior shows a wider spread than the standard model. 🔹 At its best, it hallucinates less than GLM-5.3, catching very fine details buried in the context. 🔹 At its worst, it is worse — misreading long prompts and making basic mistakes. Longer inputs trigger the bad case more often. 4️⃣ The real story: token efficiency without longer reasoning Unlike many small models that stretch their reasoning chains to buy intelligence, Flash uses fewer tokens than GLM-5.3 on nearly every task — as little as 40% of the standard model's consumption in the best cases. On problems where most models brute-force the search space, Flash often narrows it down almost by intuition. The author notes GPT-5.6 still holds the token-efficiency crown on some difficulty levels. One catch: on tasks that genuinely require exhaustive search, Flash's consumption matches the standard model — and it can hit the API's 128K max output cap, leaving answers truncated. 5️⃣ Why this matters Four years into the LLM era, most people have still never used one for daily work. Products built on top of these models urgently need a model cheap enough to flip their ROI positive. As flagships cross the "good enough" threshold and their capability starts to overflow, that surplus intelligence should be inherited by a more civilian model. The market needed a low-price model — someone had to ship it. 🔗 Full Reading: 📊Author's monthly logic benchmark: #GLM# #Zhipu# #GLM5Flash# #LLM# #AIInference# #OpenWeights# #AIAgents#

Open Weights No single Open Weights model dominates across languages, with four different labs leading depending on the language: ➤ Boson AI's Higgs Audio V3 TTS is the top Open Weights model across 4 of the 9 new languages: Hindi, German, Portuguese and Vietnamese. ➤ @FishAudio OpenAudio S1 Mini leads Open Weights in Japanese, Spanish and Arabic. ➤ @MistralAI Voxtral TTS leads Open Weights in French. ➤ BreezeBlue Breeze TTS 2 leads Open Weights in Mandarin. Open Weights coverage remains much thinner than proprietary models, with just 2 to 6 Open Weights models ranked on each language leaderboard, compared with 15 to 24 public models overall.

Open weights. Shared progress. MiniMax H3 is moving fast. We built H3 for video generation with native stereo audio and multimodal reference control. The open-source community is making that capability faster, more accessible, and easier to build on. Recent highlights: • FastH3 — FastVideo, Nuva Lab and NVIDIA: 4-step distillation, now running on DGX Spark and Apple Silicon. • Sol-H3 — NVIDIA’s SANA team: 15 seconds of 768p video + audio in 6.6 seconds on 8×B300, in the team’s warm-inference benchmark.* • VDN — Haocheng Xi and the OpenVDN team: rethinking attention for faster H3 inference, with weights, training and inference code released. • PDD — NVIDIA’s distillation method, brought to H3 by Alibaba PAI as 8-step Acc-LoRAs, now supported in ComfyUI. • LightX2V — 4- and 8-step Turbo LoRAs, with workflows for text, image and reference-conditioned video + audio. Behind every release are people training, optimizing, quantizing, testing and sharing. Special thanks to: @haoailab @nuvalab @NVIDIAAI @xieenze_jr @HaochengXiUCB @ArashVahdat @julberner @LightX2V @ComfyUI And to the individual contributors pushing the work forward: @haozhangml @cxlcl1 @lawrence_cjs @yitongli165665 @haopengl33 @songhan_mit @shanasaimoe Thank you for building with H3 and helping make it faster, more accessible, and more useful for the community. Powerful models go further when we build together. Keep pushing H3. Excited to see what comes next. 🚀 Explore the ecosystem:

Open weights models are further from the frontier than we have seen in some time. Mythos was launched in March, and, as good as the open model releases have been since then (K3 & GLM-5.3 are great) they aren’t that close to Mythos or Astra in practice. I assume that will change