Alibaba Qwen 3.5 small models efficiency benchmark
Alibaba Qwen 3.5 small models beat GPT-4o-mini on benchmarks at $0.01/M tokens. Six MIT-licensed sizes, 119 languages, runs on iPhone.
Whether you're looking for an angel investor, a growth advisor, or just want to connect — I'm always open to great ideas.
Get in TouchAI, startups & growth insights. No spam.
TL;DR: On March 10, 2026, Alibaba released the Qwen 3.5 family on HuggingFace: six MIT-licensed models from 0.5B to 32B parameters. The 7B model scores 74.2% on MMLU, beating GPT-4o-mini's 72.9%, at $0.01 per million tokens (vs. $0.15/M for GPT-4o-mini). The 32B scores 85.8% on MMLU, edging past GPT-5.3's 85.4%. The 0.5B runs on an iPhone 15 Pro at 40 tokens per second via CoreML, with no network connection needed.
Alibaba's Qwen 3.5 family is the most direct empirical challenge to the "scale is all you need" argument in the AI industry's recent history.
The release arrived on March 10, 2026, with all six model weights available immediately on HuggingFace. No waitlist, no API-only preview, no commercial licensing negotiation. MIT license on every size, including the 32B.
The six sizes are: 0.5B, 1.5B, 3B, 7B, 14B, and 32B parameters. Each targets a distinct deployment tier, and the capability jumps between tiers are not uniform. The 7B is where the economic argument gets sharp. The 32B is where the benchmark argument gets uncomfortable for OpenAI.
Here is the full family at a glance:
| Model | Context window | Target deployment | License |
|---|---|---|---|
| Qwen 3.5-0.5B | 32K tokens | Smartphones, mobile apps | MIT |
| Qwen 3.5-1.5B | 32K tokens | IoT, embedded microcontrollers | MIT |
| Qwen 3.5-3B | 32K tokens | In-vehicle systems, smart home hubs | MIT |
| Qwen 3.5-7B | 128K tokens | Cloud API services, production workloads | MIT |
| Qwen 3.5-14B | 128K tokens | Mid-range enterprise reasoning | MIT |
| Qwen 3.5-32B | 128K tokens | High-stakes reasoning, research pipelines | MIT |
The context window split is deliberate. Mobile and embedded variants get 32K, which covers most single-session use cases on constrained hardware. The 7B and above get 128K, which is enough to process entire legal contracts, long codebases, or extended research documents in a single pass.
Three training choices separate Qwen 3.5 from earlier small models technically. First: DPO (Direct Preference Optimization) combined with GRPO (Group Relative Policy Optimization) for reasoning alignment. This combination trains reasoning capability more efficiently than classic RLHF, because DPO removes the expensive explicit reward model, while GRPO stabilizes gradient signals by comparing outputs across a group rather than against a single reference. The result is a model that reasons better per parameter than naive scale-up approaches would achieve. Second: flash attention v3, which cuts the memory and compute cost of the attention mechanism at inference time. Third: a context window architecture calibrated to each model's deployment hardware, rather than maximizing context uniformly and burning compute budget on mobile sizes that do not need it.
Key finding: Qwen 3.5-7B achieves 74.2% on MMLU and 85.1% on HumanEval using 7 billion parameters trained with DPO+GRPO, a combination that costs roughly one-fifteenth of running GPT-4o-mini per token on Together AI.
The benchmark claim that requires the most scrutiny is this: a 7-billion parameter model from Alibaba outperforms GPT-4o-mini on MMLU, the standard test for broad knowledge and reasoning across 57 academic subjects.
| Model | MMLU | HumanEval | Parameters | MIT licensed |
|---|---|---|---|---|
| Qwen 3.5-7B | 74.2% | 85.1% | 7B | ✓ |
| GPT-4o-mini | 72.9% | ~81% | Unknown (~8B) | ✗ |
| Llama 3.1-8B | 68.4% | 72.6% | 8B | ✓ |
| Mistral-7B v0.3 | 64.2% | 63.2% | 7B | ✓ |
The MMLU gap is 1.3 percentage points. That sounds narrow until you consider who Qwen 3.5-7B is beating. GPT-4o-mini is OpenAI's flagship small model, built with resources that dwarf what most research organizations can access. A fully open-weight model at a similar parameter count, available for free commercial use, beating it on the canonical knowledge and reasoning benchmark is not a marginal win.
HumanEval is the number that carries more signal for reasoning quality. An 85.1% score on code generation at 7B parameters indicates the DPO + GRPO training is genuinely improving reasoning, not just optimizing benchmark-specific patterns. Code tasks require the model to understand program logic, construct multi-step solutions, and avoid logical errors. You cannot memorize your way to 85%.
According to Ars Technica's benchmarking coverage, the efficiency gains at this parameter tier reflect a broader shift in how Chinese labs approach model training: systematic investment in training methodology research rather than scaling raw compute budgets. The evidence in the numbers suggests that methodology gap has closed, and in the 7B size class, reversed.
What should you actually do with this information? If you are running GPT-4o-mini for text classification, summarization, document analysis, or customer service routing, the benchmark data says Qwen 3.5-7B does that job better and costs 15x less per token. The switching cost is a few hours of integration work. The annual savings at moderate enterprise volume can reach seven figures.
Qwen 3.5-32B scores 85.8% on MMLU. GPT-5.3, OpenAI's recently released model that has become the enterprise standard for high-end reasoning, scores 85.4% on the same benchmark.
A 32-billion parameter open-source model from Alibaba outperforming a closed frontier model from OpenAI on the most widely used benchmark in enterprise AI evaluation is a claim that demands careful qualification. So here are those qualifications, stated plainly.
MMLU measures knowledge breadth and reasoning across 57 academic subject areas. It is the right benchmark for most enterprise text tasks: document review, knowledge base Q&A, classification, summarization, code review, contract analysis. It is not the right benchmark for multimodal reasoning (GPT-5.3 leads there), complex multi-step agentic workflows (GPT-5.3 leads there), or audio and image understanding (Qwen 3.5-32B is text-only in this release).
Given those qualifications, the Alibaba argument is precise: for the majority of production enterprise text tasks, MMLU-level capability is the ceiling. If Qwen 3.5-32B matches or exceeds GPT-5.3 on MMLU, and those tasks make up the bulk of your AI workload, you do not need GPT-5.3. You need Qwen 3.5-32B, self-hosted or on Alibaba Cloud, with MIT licensing and no per-token fees from OpenAI.
The MMLU comparison in full:
| Model | MMLU score | Open weights | Text-only |
|---|---|---|---|
| Qwen 3.5-32B | 85.8% | ✓ | ✓ |
| GPT-5.3 | 85.4% | ✗ | ✗ |
| Qwen 3.5-14B | ~80.1% | ✓ | ✓ |
| Claude 3.5 Sonnet | ~83.0% | ✗ | ✗ |
VentureBeat's AI inference cost analysis published March 2026 noted that Qwen 3.5-32B on Alibaba Cloud runs at approximately $0.02-0.04 per million tokens self-hosted, versus GPT-5.3's API pricing at rates that make 100M token daily workloads a budget line that most mid-market enterprises cannot absorb.
The "scale is all you need" argument that justified massive compute investments through 2024 is not disproven. Scale still produces capability gains at the frontier. But it is now demonstrable that well-trained smaller models can match frontier performance for specific capability classes, at a fraction of the infrastructure cost.
Qwen 3.5-0.5B runs natively on an iPhone 15 Pro at 40 tokens per second, generating one word roughly every 25 milliseconds, entirely on-device, with no network connection required.
Alibaba achieved this by exporting the 0.5B model to Apple's CoreML format, targeting the iPhone 15 Pro's Neural Engine directly. The model fits in a manageable memory footprint after INT8 quantization: well under 1GB, within the memory budget of current flagship smartphones.
Forty tokens per second is real-time by any conversational standard. A user asking a question gets a response that feels instant. No round-trip to a server, no API latency, no data leaving the device.
Android deployment runs through ExecuTorch, Meta's mobile inference runtime. Both paths are documented and available for developers now. The cross-platform story covers the full 6+ billion smartphone installed base.
What 40 tokens per second on a smartphone actually enables:
Offline assistants that work on planes, in basements, in developing-world environments with no reliable connectivity. The assistant degrades only when the hardware degrades, not when the network does.
Private AI for sensitive queries. Medical questions, personal financial decisions, and confidential business conversations stay on-device. No API logs, no server-side data retention, no terms-of-service analysis required about what the provider does with your query data.
Real-time translation across 119 languages. A 0.5B model running at 40 tokens per second can translate speech faster than cloud APIs that carry 200-400ms of round-trip latency. For live conversation, that latency difference is the gap between usable and unusable.
Edge inference for industrial devices. The same CoreML-optimized model runs on Apple Silicon iPads, MacBooks, and embedded systems. The 1.5B and 3B models extend this to industrial edge computing scenarios where cloud connectivity is intermittent or prohibited by security policy.
The implications for Apple and Google are uncomfortable. Both companies have invested heavily in proprietary on-device AI as a platform differentiator: Apple Intelligence, Google Gemini Nano. An MIT-licensed model achieving 40-tokens-per-second inference on existing hardware removes the competitive moat around proprietary on-device model ecosystems. Any developer can now ship a fully local AI feature on iOS without depending on Apple Intelligence at all.
The economics of Qwen 3.5 are where the efficiency argument becomes a budget conversation.
| Model | Provider | Cost per 1M input tokens | Cost per 1M output tokens |
|---|---|---|---|
| Qwen 3.5-7B | Alibaba Cloud | $0.008 | $0.008 |
| Qwen 3.5-7B | Together AI | $0.010 | $0.010 |
| Llama 3.1-8B | Together AI | $0.018 | $0.018 |
| Gemma 3-9B | $0.030 | $0.060 | |
| Claude 3.5 Haiku | Anthropic | $0.080 | $0.400 |
| GPT-4o-mini | OpenAI | $0.150 | $0.600 |
Qwen 3.5-7B at $0.01/M tokens on Together AI outperforms GPT-4o-mini on MMLU and costs 15x less. On Alibaba Cloud, the price drops to $0.008/M input tokens, making it 18.75x cheaper than GPT-4o-mini.
To make this concrete: a production application processing 100 million tokens per day, roughly the scale of a mid-sized enterprise AI deployment, costs $15,000 per day on GPT-4o-mini. On Qwen 3.5-7B via Together AI, that same workload costs $1,000 per day. On Alibaba Cloud, $800 per day. The annual difference: $5.4 million versus $292,000.
That is not a rounding error. It is a budget-line transformation.
TechCrunch's Qwen 3.5 release coverage noted that for high-volume applications (customer service automation, document processing, content classification, search augmentation), the Qwen 3.5-7B cost profile changes what is economically viable. Workloads that were cost-prohibitive on GPT-4o-mini pricing become routine at $0.01 per million tokens.
The decision framework for model selection in 2026 is getting simpler, not more complex. Run Qwen 3.5-7B for any text task where MMLU-level capability is sufficient and cost efficiency matters. Run GPT-4o-mini only when OpenAI ecosystem integration is required or multimodal capability is the deciding factor. Run Qwen 3.5-32B for enterprise reasoning tasks where you want 85.8% MMLU performance and control over your deployment environment. Run GPT-5.3 for frontier multimodal or agentic tasks where Qwen 3.5-32B's text-only architecture is a genuine constraint.
The structural pressure on GPT-4o-mini pricing is not theoretical. As AT&T's move to small language models demonstrated, enterprises that benchmark model selection cut AI costs by up to 90%. Qwen 3.5 gives those enterprises an open-source option that outperforms the incumbent on benchmarks and costs 15x less. The argument practically makes itself.
Qwen 3.5 supports 119 languages, including strong coverage of Arabic, Hindi, Bengali, Swahili, Malay, and Vietnamese, all languages that Western AI models have historically handled poorly.
GPT-4o-mini supports approximately 50 languages with meaningful quality. Gemma 3 has solid European language coverage but drops off significantly for Southeast Asian, African, and South Asian languages. Qwen 3.5-7B handles 119 languages at a quality level that reflects genuine training investment, not token-level support.
Why does this matter? About 4 billion people on Earth do not primarily use English. The AI models accessible to those people are determined by which models their languages are well-represented in. A model that handles Hindi poorly is, for practical purposes, not useful to Hindi speakers, regardless of its MMLU score on English benchmarks.
Alibaba's multilingual investment reflects its commercial position. Alibaba Cloud operates data centers and serves customers across Southeast Asia, the Middle East, and South Asia. These are markets where Alibaba has revenue incentives that Western AI companies largely lack. The 119-language coverage is not an academic achievement. It is a sales advantage in markets worth hundreds of billions annually.
Western AI labs have systematic weaknesses in non-Latin-script languages, morphologically complex languages, and languages with limited digital text in pre-training corpora. Chinese labs, starting from a non-English-first perspective, have consistently produced stronger multilingual models. Qwen 3.5 continues that pattern at a capability level that, for the first time, is also competitive with Western frontier models on English tasks.
The release creates pressure across the entire efficient model segment. The companies most directly affected are Microsoft, Meta, and Google.
Microsoft Phi-4 at 14B parameters has been positioned as the top small model for reasoning since late 2025. Qwen 3.5-14B enters the same size class with comparable reasoning benchmarks and MIT licensing, versus Phi-4's more restrictive terms. The fine-tuning ecosystem advantage Phi-4 built narrows when the competitor is also fully open.
Google Gemma 3 is competitive on multilingual benchmarks but priced at $0.03/M on Google Cloud, 3x the cost of Qwen 3.5-7B on Together AI. For workloads where language coverage overlaps, the cost argument for Gemma 3 is hard to sustain. Google's edge remains Vertex AI integration and multimodal Gemma variants that Qwen 3.5 does not offer in this release.
Meta Llama 3.1-8B has been the open-source community's default 7B-8B base model. The performance gap from the Qwen 3.5-7B's 74.2% MMLU versus Llama 3.1-8B's 68.4% is 5.8 percentage points. Developers choosing a base model for fine-tuning projects will need to reconsider that default. Llama retains real advantages: tooling ecosystem maturity, more community fine-tunes, and Meta's deployment infrastructure. But raw capability at the same parameter tier now favors Qwen 3.5.
GPT-4o-mini at $0.15/M tokens is increasingly hard to justify for text-only workloads. OpenAI faces two options: cut pricing aggressively, or differentiate more clearly on capabilities small models cannot match (multimodality, complex function calling, extended agentic reasoning). Based on what Qwen 3.5 delivers, that differentiation window is narrowing.
This mirrors what happened with DeepSeek's open-source releases earlier in 2026. Chinese open-source labs are not just closing the gap with Western models. They are setting efficiency benchmarks that Western labs then have to respond to. The Qwen 3.5 release is the clearest version of that dynamic in the small model space.
Summary of competitive impact:
| Model | Biggest threat from Qwen 3.5 | Can it respond? |
|---|---|---|
| GPT-4o-mini | 15x cost gap, MMLU deficit | Cut price or add multimodal differentiation |
| Llama 3.1-8B | 5.8% MMLU gap, same license | Llama 4 release needed |
| Gemma 3-9B | 3x cost gap, license restrictions | Gemma 4 or price cut |
| Phi-4 (14B) | MIT license parity, broader multilingual | New reasoning benchmark push |
Understanding why Qwen 3.5 beats GPT-4o-mini at 7B parameters requires understanding what DPO + GRPO actually does differently from standard RLHF training.
Traditional RLHF (Reinforcement Learning from Human Feedback) training runs in three stages: supervised fine-tuning, reward model training, and then reinforcement learning against that reward model. The reward model training phase is computationally expensive and introduces instability because the reward model itself can be wrong, and the reinforcement learning phase can overfit to the reward model rather than the underlying human preference. Scaling this reliably requires enormous compute.
DPO (Direct Preference Optimization) eliminates the explicit reward model entirely. Instead of training a separate model to score outputs, DPO directly optimizes the base model against preference pairs (preferred output vs. rejected output) using a mathematical equivalence that produces the same result as reward-model-based RLHF. Less compute, fewer training stages, more stable gradients.
GRPO (Group Relative Policy Optimization) adds another layer of efficiency. Standard policy gradient methods compute gradient updates by comparing each output to a fixed baseline. GRPO compares each output against a group of alternative outputs sampled from the current policy. This produces gradient signals that reflect actual relative quality, not just distance from an arbitrary reference, which stabilizes training with fewer samples.
The combination: DPO removes the reward model cost, GRPO stabilizes the training signal. Alibaba can train a 7B model to stronger reasoning capability than approaches that spend more compute on brute-force RLHF at larger scale.
Flash attention v3 is the inference-side counterpart. Standard attention is quadratic in sequence length: double the context length and you quadruple the compute. Flash attention reconfigures the attention computation to use GPU SRAM more efficiently, reducing memory bandwidth requirements dramatically. For 128K-token contexts on the 7B, this difference between flash attention v3 and standard attention is the difference between practical inference costs and prohibitive ones.
None of these techniques are proprietary. DPO, GRPO, and flash attention are published research, reproducible by any lab. The question is which labs are prioritizing this methodology investment versus raw scale investment. Based on results, Alibaba is.
The efficiency case for Qwen 3.5 is strong. That does not mean every team should migrate tomorrow. Here is the honest breakdown.
Switch to Qwen 3.5-7B now if:
Your workload is text-only, high-volume, and the task category falls within MMLU-testable capability (document analysis, summarization, classification, code review, knowledge base Q&A). The cost savings are immediate and the benchmark data supports the capability claim. Integration takes hours, not weeks.
Switch to Qwen 3.5-32B now if:
You need 85.8% MMLU performance for enterprise reasoning tasks, you are comfortable with self-hosted or Alibaba Cloud deployment, and you want MIT licensing for fine-tuning on proprietary data. The benchmark parity with GPT-5.3 on MMLU is real.
Wait or evaluate carefully if:
Your workload requires multimodal inputs (images, audio, video). Qwen 3.5 is text-only in this release. GPT-4o-mini and GPT-5.3 maintain clear advantages in multimodal reasoning. If multimodality is core to your application, the switch is premature.
Also wait if you depend heavily on OpenAI's function calling infrastructure, tool use reliability, or the broader OpenAI ecosystem (Assistants API, fine-tuning platform, Batch API). Qwen 3.5 is a strong model. Its ecosystem maturity around structured outputs and tool use does not yet match OpenAI's multi-year head start on those specific features.
For on-device deployment:
If you are building mobile apps that require local inference, Qwen 3.5-0.5B is the strongest fully open option available today. 40 tokens per second on iPhone 15 Pro via CoreML, MIT-licensed, 119 languages. The only caveat: this is a 0.5B model. Its reasoning ceiling is lower than the 7B. Match the size to the task.
For context on how the small model market has been evolving, the Meta Llama 4 open-source roadmap provides useful comparison on where Meta's response to efficiency-first Chinese models is heading.
Qwen 3.5 is a family of six open-source language models released by Alibaba on March 10, 2026. The models span six parameter sizes: 0.5B, 1.5B, 3B, 7B, 14B, and 32B. All six are released under the MIT license, meaning they are freely available for commercial use, fine-tuning, and redistribution. The models were trained using DPO and GRPO techniques for reasoning efficiency, plus flash attention v3 for inference speed.
Qwen 3.5-7B scores 74.2% on MMLU versus GPT-4o-mini's 72.9%, a 1.3 percentage point advantage. On HumanEval (code generation), Qwen 3.5-7B scores 85.1% versus GPT-4o-mini's approximately 81%. Both scores come from a model that costs 15x less per token on Together AI ($0.01/M vs. $0.15/M).
On Together AI, Qwen 3.5-7B is priced at $0.01 per million tokens (input and output). On Alibaba Cloud's API, the input token price is $0.008 per million. This compares to $0.15/M for GPT-4o-mini, $0.08/M for Claude 3.5 Haiku, and $0.03/M for Gemma 3-9B on Google Cloud.
Yes. Alibaba exported the 0.5B model to Apple's CoreML format, enabling native on-device inference on iPhone 15 Pro at approximately 40 tokens per second via the Neural Engine. Android deployment runs through ExecuTorch. The model runs without a network connection and processes data entirely on-device, with no API calls required.
Qwen 3.5-32B scores 85.8% on MMLU versus GPT-5.3's 85.4%, a 0.4 percentage point advantage. GPT-5.3 maintains clear advantages in multimodal reasoning and complex agentic tasks. For text-only tasks at MMLU capability level, Qwen 3.5-32B is slightly ahead of GPT-5.3 on this benchmark.
All six Qwen 3.5 models use the MIT license. MIT is one of the most permissive open-source licenses available: no restrictions on commercial use, modification, or redistribution. Organizations can fine-tune, deploy, and build commercial products on any Qwen 3.5 model without paying royalties or accepting usage caps.
Qwen 3.5 supports 119 languages, with strong coverage of Arabic, Hindi, Bengali, Swahili, Malay, and Vietnamese, languages that Western AI models have historically underserved. GPT-4o-mini supports approximately 50 languages with meaningful quality.
DPO (Direct Preference Optimization) replaces the explicit reward model in standard RLHF training, reducing compute cost while maintaining alignment quality. GRPO (Group Relative Policy Optimization) compares model outputs against a group of alternatives rather than a fixed reference, producing more stable gradient signals with less training data. Together, these methods let Alibaba train a 7B model to higher benchmark scores than traditional scale-up approaches achieve at equivalent compute cost.
The 7B, 14B, and 32B models support 128K token context windows. The 0.5B, 1.5B, and 3B models support 32K token context windows, sized for their mobile and embedded deployment scenarios where processing a 128K-token document in one pass is not a typical requirement.
Qwen 3.5 and DeepSeek target different size classes. DeepSeek V3 and R1 focus on large frontier models (60B+ parameters). Qwen 3.5 covers the full range from 0.5B to 32B, making it the more practical option for mobile, edge, and mid-scale enterprise deployments. For tasks requiring frontier-class reasoning with self-hosted deployment at large scale, DeepSeek remains the benchmark.
Training methodology matters more than parameter count in the current state of model development. Qwen 3.5-7B uses DPO + GRPO training, which extracts more capability per parameter than RLHF-based approaches at similar scale. GPT-4o-mini's parameter count is not publicly disclosed but is estimated at approximately 8B. The 1.3% MMLU gap reflects a training quality difference, not a size difference.
For text-only workloads (document analysis, classification, summarization, code review, Q&A), the benchmark data supports replacing GPT-4o-mini with Qwen 3.5-7B today. For multimodal workloads, GPT-4o-mini retains a clear advantage since Qwen 3.5 is text-only in this release. For workloads deeply integrated with OpenAI's Assistants API or structured output tooling, evaluate carefully before switching.
Qwen 3.5-7B in FP16 format requires approximately 14GB of VRAM. This fits on a single NVIDIA RTX 4080 (16GB) or RTX 3090 (24GB). With INT4 quantization via llama.cpp or GGUF format, the 7B model drops to approximately 4-5GB, runnable on an RTX 3060 (12GB) or M2 MacBook Pro with 16GB unified memory.
Standard attention computes quadratically with sequence length: doubling context length quadruples compute cost. Flash attention v3 reorganizes the attention computation to maximize GPU SRAM utilization, processing attention in tiles that fit in fast memory rather than slow HBM. For 128K-token contexts on the Qwen 3.5-7B, flash attention v3 is the difference between practical and prohibitive inference costs at production scale.
At 100 million tokens per day: GPT-4o-mini on OpenAI costs approximately $5.4 million per year ($0.15/M x 100M x 365). Qwen 3.5-7B on Together AI costs approximately $365,000 per year ($0.01/M x 100M x 365). On Alibaba Cloud at $0.008/M, the annual cost is approximately $292,000. The gap is $5.1 million annually at that volume.
For narrow, well-defined tasks, yes. Text classification, intent detection, short summarization, simple translation, and keyword extraction are all within the 0.5B model's capability ceiling at 40 tokens per second on iPhone. For tasks requiring multi-step reasoning, long-form generation, or complex code assistance, move up to the 3B or 7B.
Publicly reported MMLU scores at release: 7B at 74.2%, 32B at 85.8%. The 14B scores approximately 80%, consistent with the parameter scaling curve. The 0.5B through 3B models do not report MMLU scores in the official benchmarks, reflecting that these sizes are designed for narrow on-device tasks rather than broad knowledge benchmarking.
The 7B and larger models support function calling and JSON mode output. Documentation and community testing as of March 2026 indicate reliability is strong but not at the same maturity level as OpenAI's function calling infrastructure, which has had three years of production refinement. For tool-heavy agent applications, run your own reliability benchmarks before committing to a full migration.
Qwen 3.5-7B scores 85.1% on HumanEval, which tests Python code generation across 164 programming problems. This is higher than GPT-4o-mini's approximately 81% and significantly higher than Llama 3.1-8B's 72.6%. For code generation, review, and debugging tasks in Python, the 7B is the strongest open model at this parameter tier as of March 2026.
OpenAI has reduced pricing multiple times in response to open-source competition (GPT-3.5-turbo dropped from $0.002/1K to $0.0005/1K between 2023 and 2025). A 15x cost gap between equivalent-capability models creates strong market pressure. Based on the pattern, expect a GPT-4o-mini price reduction within 6-12 months, though OpenAI has not announced anything. The more sustainable response is capability differentiation: multimodality, function calling reliability, and ecosystem integration that smaller open models cannot easily replicate.
All six Qwen 3.5 model weights are available on HuggingFace under the Qwen organization. The models are available in standard Transformers format, GGUF format for llama.cpp, and with CoreML export for iOS/macOS deployment. No account required, no download restrictions, MIT license applies to all uses.
If you are running text-only AI workloads at any scale, benchmark Qwen 3.5-7B against your current model before your next contract renewal. The cost math has changed, and the benchmark data supports the switch. Read more about what the broader open-source AI shift from China means for enterprise AI procurement decisions.
Benchmark data sourced from official Qwen 3.5 model pages on HuggingFace. Pricing data verified via Together AI and Alibaba Cloud public model catalogs as of March 2026. Release coverage via TechCrunch, Ars Technica, and VentureBeat.
DeepSeek V4 trillion parameter open source multimodal model beats GPT-5.3 on MMLU. MIT license. $0.14/M tokens. Full benchmark breakdown.
Open source video AI LTX Helios march models hit 83.1 VBench. Run 10 hours of video for $0.19 vs $1,440 on Sora. Here's what changed.
Alibaba's Qwen 3.5 9B model outperforms OpenAI's 120B open-source model on benchmarks and runs on consumer laptops — a direct challenge to the 'scale is everything' AI narrative.