Community-tracked AI experience, updated hourly
EN
Menu

Qwen, GLM Rush Competing Flash Models as Ox Alpha Cuts DeepSeek Pricing by Two-Thirds

Chinese AI labs Qwen and Zhipu release competing MoE Flash models on the same day, with Qwen3.8-Flash-Next launching hybrid attention architecture while GLM-5.3-Flash "Ox Alpha" undercuts DeepSeek-V4-Flash pricing by 66%, challenging Western models on capability and cost.

Hacker News 3 · Reddit 6 · Zhihu 6 447 covered discussions 5 source-linked evidence passages

The day in brief

Qwen3.8-Flash-Next officially launches with 125B parameters, 1M token context, benchmarks, GGUF quants, and community reports confirming local deployment feasibility despite n-gram offloading challenges.

Zhipu AI confirms GLM-5.3-Flash as 'Ox Alpha' with 66% lower API pricing than DeepSeek-V4-Flash, competitive coding and agent benchmarks, and confirmed 50K domestic chip deployment.

New trace data reveals GPT-5.6 Sol consumes 7.5x more tokens per task than GPT-5.5, depleting ChatGPT Plus 5-hour limit in 11 minutes of planning with a hard stop regardless of task completion status.

Community benchmarks show Muse Glimmer outperforming Qwen3.8 27B on non-coding tasks while using fewer reasoning tokens, with efficient sliding window attention enabling viable extended context processing.

Hacker News community testing surfaces divergent content policies across AI models, with Chinese AI models declining certain historically significant location prompts while Western models provide complete responses.

Product and platform changes

Zhipu AI Confirms GLM-5.3-Flash as "Ox Alpha" with Frontier Pricing and Coding Performance

Community Discussion

  • Zhihu discussions attracted significant engagement with individual answers reaching 837 peak interactions, with users highlighting the model's 320 billion parameters, 18 billion activated parameters, native multimodal capabilities, and performance exceeding GLM-5.2 at one-tenth the price.
  • English-language communities on Reddit and Hacker News noted the significance of the open-weight release announcement, with users drawing comparisons to Claude Opus 4.8 performance in coding tasks and positioning the model as a competitive alternative.
  • Community members characterized the pricing as establishing a new "killing line" that surpassed DeepSeek following recent price adjustments, with particular emphasis on the cost-to-performance ratio for developer workloads.
  • Discussion included analysis of the hybrid attention architecture and domestic chip deployment as factors in achieving the reported cost and performance metrics.

Key Takeaways

  • GLM-5.3-Flash represents Zhipu's confirmed Ox Alpha model with frontier-level pricing, positioning it as a cost-effective alternative for developers seeking multimodal capabilities and competitive coding performance.
  • The planned weight release may enable fine-tuning and self-hosting, expanding use cases beyond API consumption.

Value Proposition

  • API pricing at 0.4 yuan per million input tokens and 1.4 yuan per million output tokens (with promotional discount to 1/20 of standard GLM-5.3) offers significant cost savings for high-volume applications.
  • Native multimodal capabilities and 1 million token context eliminate the need for separate models for different input types.
  • Competitive coding benchmarks (DeepSWE approximately 63%) position the model for developer tooling and code generation workloads.

Practical Applications

  • Cost-sensitive coding and agentic tasks where DeepSeek-V4-Flash pricing previously set the benchmark.
  • Applications requiring native multimodal input handling with extended context windows.
  • Scenarios where self-hosting or fine-tuning of weights becomes feasible following the planned weight release.

Event Overview

  • Zhipu AI officially confirmed that GLM-5.3-Flash is the model previously known as "Ox Alpha." The announcement included full technical specifications: 320 billion total parameters with 18 billion activated, hybrid Linear plus DeepSeek Sparse Attention mechanism, 1 million token context window, and native multimodal capabilities.
  • API pricing was set at 0.4 yuan per million input tokens and 1.4 yuan per million output tokens, representing one-third the cost of DeepSeek-V4-Flash. A limited-time promotional discount brought the effective price to 1/20 of the standard GLM-5.3 rate.
  • Technical optimizations include IndexPool compression reducing attention calculation by 3x and KV Cache by 4.4x compared to standard GLM-5.3. The model is deployed across 50,000 domestic AI chips with reported 3x performance optimization.
  • Zhipu announced plans to release the model's weights, enabling broader community access and fine-tuning capabilities.
  • Benchmark results show Code Arena ranking fifth and DeepSWE benchmark achieving approximately 63% performance, with the community positioning the release as competitive with Claude Opus 4.8 in coding and agent tasks.

Community evidence

GLM-5.3-Flash achieves a DeepSWE score of 63% with a per-task cost of just $0.24, while DeepSeek-V4-Pro, which achieves the same score, costs $1.67 per task. The API price of GLM-5.3-Flash is only 1/10 that of GLM-5.3, and with the current limited-time 50% discount, the actual price during the promotion is equivalent to 1/20 of GLM-5.3. GLM-5.3-Flash uses a hybrid architecture combining linear attention and sparse attention, which greatly reduces the computational overhead of attention. Compared with GLM-5.3, the attention computation of GLM-5.3-Flash is reduced to about 1/3, and the KV Cache is reduced to about 1/4.4.

Model experience tracking

GPT-5.6 Sol Depletes ChatGPT Plus 5-Hour Limit in 11 Minutes of Planning, New Trace Data Shows 7.9x Token Increase Per Turn

What Happened

  • A Reddit user documented a single-prompt planning session consuming 54% of their ChatGPT Plus 5-hour limit in 11 minutes of thinking time; the session ultimately ran for 28 minutes 29 seconds and exhausted 100% of the 5-hour limit, leaving 84% of the weekly quota unused.
  • Trace data from a coding task comparison showed GPT-5.5 (xhigh) consuming approximately 860M–1B tokens versus GPT-5.6 Sol (High) consuming approximately 3B tokens for comparable tasks.
  • Per-turn token comparison: GPT-5.6 Sol averaged approximately 17.88M tokens per tool-bearing turn versus approximately 2.25M for GPT-5.5, representing a roughly 7.9x increase; GPT-5.6 Sol issued 87 model requests per tool-bearing turn versus 16.5 for GPT-5.5.
  • High/XHigh modes on GPT-5.6 Sol were documented reducing model cycles by 52%–55% relative to default settings.
  • OpenAI restored the 5-hour inactivity limit to reduce users' perception of consumption, after previously halting the reset behavior that had masked the increased token usage of GPT-5.6 Sol.
  • The 5-hour limit now triggers a hard stop regardless of task completion status; previously, an exhausted 5-hour limit would continue consuming weekly quota until the task finished.

Community Reaction

  • The Reddit user whose session depleted in 28 minutes reported their project status as 'Broken and unusable' and stated it was 'time to explore other providers.'
  • A Chinese-language commentator described the behavioral shift as frustrating, noting the hard stop interrupts ongoing tasks mid-execution, unlike the previous behavior where the 5-hour limit exhaustion would continue consuming weekly quota until task completion.
  • A Reddit commenter described the $20 Plus tier as still offering comparable or better performance than previous years' offerings, suggesting the value proposition remains acceptable despite increased consumption.
  • A second Reddit commenter reported satisfactory usage on Plus after 2 hours 30 minutes with Grok 4.6 as orchestrator and GPT-5.6 Sol as coder, with 72% of the 5-hour limit remaining.

Use Case

  • Planning and coding sessions exceeding 11 minutes with GPT-5.6 Sol High are likely to exhaust the 5-hour inactivity limit on Plus tier.
  • Complex multi-turn tasks requiring 2.25M+ tokens per tool-bearing turn are consuming 17.88M tokens per turn on GPT-5.6 Sol, making Plus limits impractical for extended development work.
  • Users running Hermes-agent-style orchestrator workflows with GPT-5.6 Sol as the coder may deplete 5-hour limits within a single extended session.

Practical Takeaway

  • High/XHigh modes reduce model cycles by 52–55% and may extend 5-hour limit duration on Plus tier.
  • Users prioritizing task completion over consumption visibility may prefer disabling Sol to use GPT-5.5, which consumed 3–7x fewer tokens per task.
  • The hard stop behavior means long-running Codex tasks on Plus tier risk interruption without consuming weekly quota, unlike the previous soft-stop approach.

Practical Value

  • GPT-5.6 Sol processes the same tasks at 3B tokens that GPT-5.5 completed at 860M–1B tokens, increasing per-session consumption by approximately 3x at minimum and up to 7.5x depending on task complexity.
  • The 5-hour inactivity limit depletes approximately 7.5x faster on GPT-5.6 Sol compared to GPT-5.5 for equivalent tasks, with single prompts documented consuming 54% of the limit in 11 minutes.

Source Prompts

  • Reddit post titled 'State of the $20 subscription' describing GPT-5.6 Sol consuming 54% of 5-hour limit in 11 minutes; 28m29s total runtime depleting 100% of 5-hour limit with project left broken.
  • Chinese-language post titled 'ChatGPT Plus Officially Restores 5-Hour Usage Limit' with trace data comparing GPT-5.5 and GPT-5.6 Sol token consumption; 2.25M vs 17.88M tokens per tool-bearing turn; 16.5 vs 87 model requests per turn; behavioral change from soft-stop to hard-stop on 5-hour limit.

Prompt Analysis

  • The trace data was collected under xhigh reasoning effort for GPT-5.5 and High reasoning effort for GPT-5.6 Sol; sampling included 12 total turns for GPT-5.6 Sol versus 179 for GPT-5.5, indicating different session scales.
  • Token metrics reflect primary observed tool envelopes including exec_command calls, exec cells, and wait calls; structural differences in native parallel batch handling were noted (43.6% for GPT-5.5 vs structurally disabled for GPT-5.6 Sol).
  • The Reddit user's consumption was documented during a thinking and planning session without code implementation, while the trace data reflects active coding tasks, suggesting the 54% in 11-minute figure may reflect overhead rather than task output.

Community evidence

In practice, GPT-5.6 with Sol is just consuming too fast now—if you use Sol, you can't handle many tasks at all before hitting your Plus weekly limit. Previously, Tibo kept doing resets, so the experience wasn't that intuitive, but now that resets have basically stopped, it's quite common for Plus users to burn through a whole week's quota in a single day. I compared tasks over several days and they were fairly similar: with GPT-5.5 xhigh, tasks that completed at around 860M to 1B tokens; with GPT-5.6 Sol High, consumption reached around 3B tokens. Under the same Juice value, the consumption of 5.6 Sol on Plus is plainly visible, so unless you're using Luna, Plus on 5.6 just doesn't hold up well.

Muse Glimmer Emerges as Underrated Alternative to Larger Models for Non-Coding Workloads

The Benchmark Comparison

  • A user conducted comprehensive benchmarks comparing Qwen3.8 27B (tested in xhigh and medium effort modes) against Muse Glimmer on implicit knowledge tasks. The xhigh effort mode with Qwen3.8 required approximately 30 hours to complete while still failing 16 test cases due to the 32K output token limit. Both the medium effort mode and Muse Glimmer completed their runs in 3-4 hours each.
  • Results showed that Muse Glimmer outperformed Qwen3.8 27B on most non-coding tasks while consuming significantly fewer reasoning tokens. The model's sliding window attention architecture enables efficient KV cache management for extended context windows, providing a technical advantage for production deployments requiring longer context processing.

Community Reception

  • The original poster expressed that Muse Glimmer is underrated, noting it performs better than Qwen3.8 27B on most tasks except code generation while avoiding the high token consumption associated with extended reasoning modes. One commenter described Muse Glimmer as 'fantastic' and superior to Qwen3.8 27B at nearly everything except code without requiring thousands of reasoning tokens per prompt.
  • Community discussion highlighted Claude Haiku 4.5 as a $20/month alternative that outperforms all tested local models, sparking conversation about the cost-effectiveness of cloud versus local deployments. Users noted that combining smaller efficient models with RAG could narrow the performance gap with frontier models, suggesting a practical path for resource-constrained environments.

Key Takeaways for Practitioners

  • Small models with efficient reasoning such as Muse Glimmer may replace larger models for practical non-coding workloads where speed and token efficiency matter more than absolute performance on every task.
  • Sliding window attention enables viable extended context processing without excessive KV cache overhead, making smaller models more practical for production use cases requiring longer context windows.

Applicable Scenarios

  • Non-coding knowledge tasks requiring moderate accuracy with token efficiency.
  • Extended context applications leveraging efficient KV cache via sliding window attention.
  • Cost-sensitive deployments comparing local model infrastructure costs against cloud subscriptions.

Measurable Benefits

  • The 3-4 hour benchmark completion time for Muse Glimmer versus approximately 30 hours for Qwen3.8 xhigh mode demonstrates significant speed advantages for smaller efficient models on knowledge-intensive tasks.
  • Elimination of thousands of reasoning tokens per prompt reduces operational costs for high-volume workloads, making token-efficient models attractive for cost-sensitive deployments.

Benchmark Methodology

  • Benchmark implicit knowledge tasks across Qwen3.8 27B in xhigh and medium effort modes and Muse Glimmer, measuring completion time, token consumption, and accuracy on non-coding tasks.

Methodology Considerations

  • This benchmark focused on implicit knowledge rather than code generation, which the poster acknowledged may disadvantage smaller models that excel at different task types. The comparison did not include RAG-enhanced runs, though the poster suggested RAG would likely improve results for both models, particularly Qwen3.8 with its extended context capabilities.

Community evidence

Muse glimmer is fantastic and better than qwen3.8 27b at just about everything than code and doesn’t need to burn thousands of reasoning tokens per prompt to do it

Ecosystem and open models

Qwen3.8-Flash-Next Release Day Brings 125B Parameter Open-Weight Model with 1M Token Context to Community

Community Debate on Model Size vs Task Performance

  • Community debate emerged on whether smaller models like Muse Glimmer outperform Qwen3.8-Flash-Next on non-coding tasks with fewer reasoning tokens.
  • User reports indicated Qwen3.8 tends to overthink on one-shot prompts during early release testing but can match Muse Glimmer in succinctness when tasks have clear requirements or unambiguous steps.
  • Strong positive reception expressed with commentary comparing performance to frontier models; initial title comparisons include 'Is THIS Opus 4.8 at home?'
  • Interest expressed in separating model from knowledge store for institutional knowledge applications; Qwen-35BA3B noted as effective for following prompts and tool calling.

Deployment Feasibility on Consumer Hardware

  • Qwen3.8-Flash-Next achieves comparable succinctness to smaller models when tasks have well-defined requirements or clear execution paths, despite higher tendency to overthink on open-ended one-shot prompts.

Inference Server and Quantization Options

  • vLLM and SGLang inference server configurations available for deployment.
  • Unsloth GGUF quantizations provided for local inference, though n-gram offloading limitation requires approximately 78GB GPU memory minimum.

Official Sampling Parameters

  • Thinking Mode: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0; Instruct Mode: temperature=0.7, top_p=0.80, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0.

Parameter Rationale

  • The recommended sampling parameters differ significantly between modes; presence_penalty of 1.5 in instruct mode addresses repetition tendencies observed in non-thinking generation, while thinking mode relies on default repetition_penalty=1.0.

Primary Deployment Scenarios

  • Institutional knowledge surfacing with model constrained to prompt-provided information.
  • AI-friendly documentation development with siloed knowledge stores.
  • Agentic workloads leveraging long-context capability up to 1 million tokens.

Release Day Technical Summary

  • Qwen3.8-Flash-Next released as first open-weight model under the new architecture with 125B total parameters using 6B activated Mixture-of-Experts, 51B n-gram embedding, and 1 million token context capability.
  • Gated Residual and Qwen Sparse Attention at micro-block level introduced as core architectural components to reduce long-context latency.
  • Official recommended sampling parameters published for Thinking Mode and Instruct mode.
  • Unsloth 1-bit quantization released at 78GB total size, with n-gram tensor occupying approximately 24GB at 4-bit quantization; n-gram cannot currently be offloaded to SSD.
  • Local deployment on consumer hardware confirmed feasible; vLLM and SGLang inference server support configurations made available.

Community evidence

[https://unsloth.ai/docs/models/qwen3.8-next](https://unsloth.ai/docs/models/qwen3.8-next) Their 1 bit quant is 78GB, which appears to include the engram tensor as 51B params at 4 bits (about 24GB), so it seems they haven't figured out how to offload it to SSD yet.

Prompt challenge

Hacker News Community Tests Reveal Divergent Content Policies Across Major AI Models

Comparative Testing Reveals Response Variations

  • Hacker News users conducted systematic comparative tests across multiple AI models using identical prompts designed to probe content policy boundaries.
  • DeepSeek V4 Flash and GLM-5.3-Flash declined to respond to the test prompt, with GLM specifically refusing when the query was submitted in Chinese. Kimi and Qwen returned server exceptions rather than providing an answer.
  • In contrast, ChatGPT and Claude produced full responses to the same prompts without modification.
  • A separate discussion thread examined Anthropic's Trusted Access for Cyber (TAC) program, where users reported issues including silent country-based whitelisting and the inability to appeal rejected verification applications.

Community Engagement and Perspectives

  • The testing thread accumulated 103 mentions on Hacker News, reflecting considerable community interest in understanding how different AI providers handle content policy enforcement.
  • Commenters acknowledged that engineers at the Chinese AI laboratories are producing impressive technical achievements, while expressing frustration that government-mandated content restrictions limit the practical utility of these models for certain applications.
  • The discussion also surfaced broader concerns about accessibility barriers that emerge when verification and access systems operate without transparency or established appeal mechanisms.

Test Prompt Structure

  • The prompt used for comparative testing followed a straightforward factual format asking for historical information about a specific location.

Implications of Differential Response Patterns

  • The test prompt was designed to assess whether models would provide factual historical information about locations that may carry political significance.
  • The variation in responses across different models demonstrates how regulatory environments and policy enforcement mechanisms shape AI behavior independently from underlying technical capabilities.
  • Testing identical queries in multiple languages, as demonstrated by the Chinese-language variation with GLM, reveals additional policy enforcement layers that may not be apparent from English-only testing.

Applicable Use Cases

  • Comparing content policy enforcement across AI providers for applications requiring consistent response behavior across different model families.
  • Evaluating AI model suitability for users operating in different regulatory environments or requiring comprehensive historical information access.
  • Assessing verification program accessibility and transparency for enterprise deployments where access restrictions may impact operational requirements.

Value for Evaluation and Selection

  • Enables fair comparison of model capabilities by accounting for policy-driven response variations that exist independently of technical performance.
  • Informs model selection decisions where content accessibility requirements vary based on use case, jurisdiction, or user needs.
  • Highlights the importance of understanding provider-specific content policies alongside technical benchmarks when evaluating AI systems for deployment.

Actionable Guidance

  • Content policy enforcement mechanisms including refusals, exceptions, and server errors vary significantly across AI providers and may respond differently based on query language.
  • Users and organizations requiring comprehensive responses to historically significant topics should carefully evaluate model provenance and associated content policies before deployment.
  • Verification and access programs lacking appeal processes may create unintended accessibility barriers for legitimate users in certain regions, particularly for enterprise use cases.

Community evidence

I'd like to try some different models, but I've heard that models from China are censored by the government.