podcasts

AI Engineer World’s Fair 2026 opened Wave 2 speaker applications and added six tracks. The signal is not Moscone West, and it is not the claimed 1M monthly unique AI engineers. The signal is the track list: Autoresearch, Memory, World Models, Tokenmaxxing, Agentic Commerce, and Vertical AI. Conference programming is not neutral. It compresses budget, hiring demand, sponsor appetite, and founder narrative into a public menu. AIE matters here because it sits closer to builders than to CIO theater or pure research venues. I think the Memory track is the cleanest call. Many agent products did not fail because tool calling was impossible. They failed because state management was awful. Once a workflow becomes non-trivial, user preferences, task history, file context, permissions, and partial conclusions get tangled. Then the agent either forgets important facts or treats stale facts as law. OpenAI, Anthropic, and Google are all patching this, but through different product surfaces. ChatGPT Memory is closer to preference storage. Claude Projects are more workspace-context oriented. Gemini leans on the Workspace data loop. The hard engineering is not “add a vector database.” It is write policy, expiry, conflict resolution, privacy deletion, retrieval explanations, and preventing old memory from poisoning current tasks. AIE giving Memory its own track feels correct because it has moved from demo accessory to product spine. World Models is more ambitious, and also easier to abuse. The body only says “spatial intelligence and adversarial reasoning.” It does not disclose speakers, evals, project names, or selection criteria. That missing detail matters. “World model” now means different things across robotics, video generation, game agents, and autonomous driving. Waymo and Tesla talk about closed-loop driving worlds. Genie-like work talks about interactive generated environments. Nvidia’s Cosmos-style framing points toward physical video pretraining. These are not the same engineering problem. If AIE accepts loose “we do spatial intelligence” talks, the track will sprawl. Strong submissions should show reproducible numbers: real robot task success, long-horizon planning error, adversarial recovery rate, or sim-to-real transfer. Without that, World Models becomes a bucket for every embodied-AI pitch. Agentic Commerce is the track I distrust most, while still agreeing it belongs on stage. The post asks how agents pay for data, APIs, and other agents. That sounds like a technical market primitive. In practice it is identity, authorization, spending limits, refunds, fraud, audit logs, tax, and data licensing. Stripe, Visa, and PayPal have all been circling agent payments. OpenAI also has clear reasons to push ChatGPT from answer surface toward transaction surface. But without standardized delegation, an agent buying an API or hiring another agent immediately hits liability. Who signs? Who pays? Who can revoke? Who eats fraud? The body gives no answer, and no candidate protocol. My read: this track will attract a lot of “agent economy” fluff. The valuable talks will be boring ones about ledgers, permissions, and risk controls. Autoresearch also needs a sharp filter. The post defines it as recursive self-improvement loops in harnesses and model training. That phrase is attractive, but “recursive self-improvement” has been oversold for a year. SWE-bench, Aider-style loops, Claude Code, and Codex-style tools show models can iterate inside a test harness. AlphaEvolve and FunSearch-style work show models can search for new solutions under formal feedback. But “automates experiments” and “trains itself into a stronger model” are separated by data contamination, reward hacking, eval overfitting, and compute cost. AIE is an engineering conference, so speakers should be forced to say what the loop modifies: prompt, scaffold, training data, loss, or weights. Without that split, Autoresearch becomes AGI cosplay. Tokenmaxxing is a funny label, but I do not buy “10x more AI-Native” as a default goal. The body itself warns against Goodharting waste, which tells me teams are already seeing token consumption turn into an internal KPI. The largest enterprise AI waste is not employees refusing to use models. It is shoving every workflow into a chat box. Token volume rises; decision quality does not automatically follow. Engineering orgs should measure task completion time, rework rate, incident rate, review cycle time, escalation rate, or defect escape rate. Measuring token usage alone is as dumb as measuring GitHub commits alone. AIE putting this problem on stage is healthy. Sponsor decks will try to turn it into “buy more seats and become AI-native.” That version is noise. The Vertical AI track also says something about general agent platforms losing some shine. Law, healthcare, GTM, and finance are not moving because models suddenly became universally competent. They move because workflows, documents, compliance rules, billing, and permissions can be structured. Harvey in legal, Abridge in clinical documentation, and Hebbia in financial research are good examples. Their value is not generic intelligence. It is embedding into permissions, audit, templates, and customer systems. GTM will be the noisiest because sales automation has always been vulnerable to fake productivity metrics. The article does not disclose the speaker bar for these vertical tracks, and that will decide whether this is useful or just sponsor segmentation. The robotics detail is also a tell. The post says last year included Physical Intelligence, Waymo, Tesla, Nvidia, K-Scale, and others. It also says AIE is allocating free expo floor space for good robotics demos, with humanoids accompanied. That is a funny line, but the engineering point is serious. Video demos have lost trust. If a robotics team cannot run something stable on a conference floor, the work gets discounted fast. Moscone West is still a controlled setting, not deployment. But live demos are more honest than another polished clip. Honestly, this post is a 2026 AI engineering heat map disguised as a call for speakers. It has no model benchmark, no pricing, no final agenda, no speaker list, no sponsor mix, and no hard attendee capacity. Those gaps limit how much we can infer. The track taxonomy still carries signal. The field is moving from “which model API should we call” toward “how do systems remember, act, pay, and survive domain constraints.” I am skeptical of the hype around Autoresearch and Agentic Commerce. I would still read the submissions list closely if I were building AI infra or agent products. Conferences reveal the problems practitioners are willing to stand behind publicly.

AINews scanned 12 subreddits and 544 Twitter accounts, and the hardest data point was vLLM 0.20.0 delivering 4× KV capacity. I do not buy the “not much happened today” framing. No GPT-6 launch, no closed frontier model, and no viral benchmark does not equal a quiet day. A lot of the AI stack now moves through vLLM release notes, same-day hosting rollouts, and orchestration previews. vLLM 0.20.0 is the clearest example. The release ships TurboQuant 2-bit KV cache for 4× KV capacity, FA4 re-enabled for MLA prefill on SM90+, a new vLLM IR foundation, fused RMSNorm with a reported 2.1% end-to-end latency gain, plus DeepSeek V4 MegaMoE support across Blackwell, Jetson Thor, ROCm, Intel XPU, and GB200/Grace-Blackwell setup. The 2.1% latency number is small. The 4× KV number is the part that changes serving math. Long-context and MoE inference often bottleneck on memory, KV movement, prefill/decode split, and scheduler behavior rather than raw FLOPs. The context has shifted hard since the GPT-4 Turbo and Claude long-context cycles. Back then, the visible fight was 128K or 200K context. Now the hard question is whether 256K or MoE-heavy sessions run cheaply enough for production agents. A model with a huge context window is easy to market. A stack that keeps memory pressure, batching, and decode throughput under control is much harder to ship. SemiAnalysis also flagged early DeepSeek V4 Pro serving results on B200, B300, H200, and GB200 disaggregated setups. The claim is that B300 can be up to 8× faster than H200 for this workload. I would discount that number until the test conditions are public. The article does not disclose batch size, context length, prefill/decode mix, quantization setup, speculative decoding, or power limits. NVIDIA generation-to-generation claims often look clean in slides, then customer TCO gets eaten by networking, memory, scheduling, and utilization. Still, the signal matters because DeepSeek V4, MegaMoE kernels, vLLM IR, and Blackwell deployment are now part of one serving ledger. There is also a live tension around CUDA. The same DeepSeek ecosystem benefits from Blackwell and vLLM optimization, while posts around TileKernels point toward avoiding CUDA lock-in. That tension is real. If DeepSeek-style models need to serve Chinese clouds and domestic accelerator fleets, they cannot put all performance-critical paths behind NVIDIA-only kernels. If they want instant overseas throughput, they still need H200, B200, GB200, and optimized vLLM paths. The open-model fight has moved beyond open weights. Open serving paths now matter just as much. If weights are open but kernels, KV cache, scheduler, and communication paths are locked, deployment freedom is narrower than the license suggests. Poolside’s Laguna XS.2 is a different kind of signal. The release is a 33B total, 3B active MoE coding model, trained in-house, Apache 2.0, and advertised as runnable on a single GPU. Community summaries mention a larger 225B/23B active model, hybrid attention, FP8 KV cache, and performance near Qwen-3.5. Ollama shipped support immediately. Poolside has spent a long time as a high-valuation coding lab with little public proof. This release finally gives practitioners something to download, inspect, and run. I still have reservations. “Near Qwen-3.5” is not enough without the benchmark name, version, pass@k setup, and agent harness conditions. Coding models can look excellent on curated tasks, internal repos, or harnessed workflows. They often degrade on SWE-bench Verified, dependency-heavy repositories, multi-turn repair, and messy real codebases. My read is simple: Laguna XS.2 proves Poolside is not vapor. It does not yet prove Poolside can take budget away from Cursor, Claude Code, or Devin-style workflows. NVIDIA Nemotron 3 Nano Omni looks more like a distribution play than a pure model play. The model is a 30B / A3B multimodal MoE with 256K context, covering text, image, video, audio, and documents. It uses a Parakeet encoder, is English-only for now, and is reported at 5.95% WER on the Open ASR leaderboard. Same-day availability across OpenRouter, LM Studio, Ollama, Unsloth, fal, Fireworks, DeepInfra, Together, Baseten, Canonical, and others is the louder signal. NVIDIA is not trying to win only with a model card. It is trying to make Nemotron the default open model that sits naturally on NVIDIA inference paths and hosted GPU supply. Meta built Llama distribution through community gravity. Mistral used permissive releases and developer goodwill. NVIDIA has a different weapon: hardware, inference libraries, hosted partners, and model releases landing together. The 5.95% WER is useful, but English-only narrows the deployment story. The cited ~9× throughput needs the comparison model, hardware, and serving conditions before I treat it as a real advantage. Mistral Workflows is the other production-shaped item. The public preview positions Workflows as an orchestration layer for durable, observable, fault-tolerant enterprise AI processes. This direction is not novel. Temporal, Prefect, LangGraph, OpenAI’s agent stack, and Anthropic tool-use ecosystems have all been circling long-running state management. Mistral needs this because “European model provider” is not enough as a durable enterprise identity. Le Chat, La Plateforme, Codestral, and agent APIs need a recoverable execution layer, or customers will wire Mistral models into their existing workflow systems. The article does not disclose the important bits: state model, retry semantics, human approval flow, log retention, audit controls, and pricing. So the direction is right, but product hardness is unproven. Durable execution is one of those phrases that sounds boring until an agent fails after 47 minutes, retries a payment twice, and leaves no useful trace. The local-agent thread also deserves attention. Hugging Face says 300,000 users have added hardware specs to the Hub. There are demos of Pi plus local models for desktop cleanup, Gemma running on-device with MLX, and Sigma as a private browser-based agent concept. This is not “everyone runs AGI offline.” It is privacy, latency, and cost pulling many small tasks back to the edge. Ollama, LM Studio, llama.cpp, and Apple MLX lowered the activation energy. The missing layer is not another 7B or 14B model. It is reliable tool permissions and OS-level safety. Once a local agent can write files, click buttons, and delete data, the permission model becomes more important than the benchmark score. So yes, this was a busy day. Laguna XS.2 shows coding labs using open weights as a trust entry point. Nemotron 3 Nano Omni shows NVIDIA tying open models to inference distribution. vLLM 0.20.0 shows serving economics moving deeper into memory and kernels. Mistral Workflows shows agent vendors admitting demo loops are not production. My pushback is against the frame: calling this quiet reflects launch-calendar bias. For practitioners, boring version numbers and same-day provider support often decide whether a 256K, multimodal, tool-using, recoverable agent takes three days to wire up or three weeks to debug.

AINews puts GPT-Image-2, Nano Banana, and Grok Imagine on the AGI path because multimodal generation widens the task surface. I buy half of that. Image generation is no longer only a consumer toy, especially when GPT-Image-2 sits inside Codex and generates assets while code changes. That touches a real product-engineering problem. But the “path to AGI” label is doing too much work. AGI framing swallows every concrete question, then every workload becomes strategic by definition. The strongest part of the piece is not the old “astronaut riding a horse” benchmark class. Those prompts mattered in the Stable Diffusion and Midjourney cycles because they exposed binding failures. They still say something about compositionality, but practitioners already know that story. The serious mechanism is the loop: Codex can call GPT-Image-2 as a skill, generate assets inside the same agent flow, wire them into code, then iterate from UI or product feedback. The test is no longer whether one image looks good. The test is whether imagegen enters PRs, reviews, tests, and deployment as a normal software-production primitive. Claude Design got attention because AI-made interface artifacts felt fresh. If OpenAI can bind image generation, code changes, issue tracking, and PR review inside Codex, a standalone artifact surface starts to look thin. This fits the last year of model-company behavior. Anthropic built strong mindshare around coding and enterprise documents. OpenAI has been trying to connect ChatGPT, Codex, GitHub workflows, and API billing into one commercial loop. The snippet says GitHub Copilot moves to usage-based billing on June 1. It also gives Codex multipliers: GPT-5.4 fast at 2x, GPT-5.5 fast at 2.5x, with GPT-5.4-mini and GPT-5.3-Codex materially cheaper. That pricing signal matters more than the AGI slogan. Agentic workflows consume runtime, tool calls, retries, generated intermediates, and human review cycles. If image generation joins that loop, GPU consumption gets harder to hide inside a $20 subscription. I have two doubts about the AINews argument. First, the article gives no cost, latency, failure-rate, or integration details for GPT-Image-2 inside Codex. It says the skill exists. It does not say whether the model reads project structure, brand rules, component libraries, design tokens, or previous assets. Without those conditions, the difference between a strong demo and a default team workflow stays unknown. Image generation has hit this wall before. A poster demo looks great, then production teams run into consistency, rights, brand constraints, editable layers, export formats, and review ownership. Second, the AGI label blurs the resource-allocation question. The piece asks whether these “side quests” deserve scarce GPU capacity and answers yes. Commercially, yes. Technically, that does not make image generation an AGI prerequisite. Multimodal generation expands the model’s action space. AGI progress still lives or dies on long-horizon planning, tool reliability, verifiable tasks, self-correction, and complex state management. The same recap gives a useful counterweight: GPT-5.5 no-thinking scores 67.1% on WeirdML, up from GPT-5.4 at 57.4%, but still behind Opus 4.7 no-thinking at 76.4% while using fewer tokens. That is a sharp comparison. OpenAI may be faster at product loops and visual workflow packaging, but the cited reasoning eval does not show dominance over Anthropic. The China open-weights section adds another pressure point. Xiaomi MiMo-V2.5-Pro is described as roughly 1T total parameters with 42B active, MIT-licensed, 1M-token context, and trained on 27T tokens. MiMo-V2.5 is around 310B total with 15B active, trained on 48T tokens, also with 1M context. Day-zero support landed in vLLM and SGLang/vLLM. That route is less about creative demos and more about giving builders long-context, agentic, coding, and omni-modal primitives. Kimi K2.6 also shows deployment pull, with the recap citing a #1 OpenRouter weekly rank and secondary claims around 300 concurrent sub-agents across 4,000 coordinated steps. The article does not disclose the original conditions for that latter claim, so I would not treat it as settled. Still, the direction is clear: OpenAI’s advantage here looks like distribution and workflow closure, not single-model capability dominance. So I read this as a product signal, not an AGI proof. Image generation is moving from content output into middleware for software work. That is a real shift for Codex, Copilot, Claude Artifacts, v0, and Figma AI. It also pushes billing away from seats and toward usage. But to prove the AGI claim, the article needs three missing numbers: retention for the Codex image skill, cost per closed-loop task, and the share of generated assets that land in production code. Without those, the AGI headline gets attention; the Codex loop is what keeps developers.

Latent Space’s 54-minute episode lands on a clean thesis: agent companies rent frontier models first, then train in-house models from workflow data. I buy half of it. It captures the survival pattern for AI application companies in 2026. It also makes the ugly middle look too linear. The agent-lab path has three stated conditions in the episode: enough data, enough workload, and enough user behavior. After that, the company trains its own models to win back cost and latency. That logic works best for Cursor and Cognition because coding products collect dense traces. They see repo structure, diffs, compiler errors, test output, terminal history, review comments, and accept rates. That is better training material than generic chat preference data. Code has executable outputs and automated checks. SWE-bench became a central benchmark because coding tasks come with a judge, not because everyone suddenly cared about GitHub issues. The smooth version of the claim hides the hard part. “We have user data, so we can train a domain model” is not a plan. Cursor and Cognition have IDEs, terminals, repos, CI loops, and human acceptance signals. Most vertical AI startups do not have that loop. A medical assistant getting doctor edits is not automatically a clinical model factory. A finance agent getting analyst comments is not automatically an auditable model pipeline. Compliance, noisy labels, rare failures, and liability eat the expected gain. The article does not disclose training cost, token volume, latency savings, or acceptance-rate deltas. It gives the operating memo, not the proof. That also explains why coding became the first breakout market. The episode names Anthropic, OpenAI, Cursor, and Cognition as winners from the coding wave. The reason is not just developer openness to new tools. Developers expose failure to the system. A failed build, failed test, rejected diff, or reverted commit becomes a learning signal. Customer support, sales, and legal workflows have feedback too, but it is slower, messier, and more political. Claude Code versus Codex stickiness often comes down to the first moment when the tool actually fixes a repo. That memory has more retention value than a marginal benchmark win. There is an outside pattern here. Anthropic’s Claude Code success follows from its long positioning of Sonnet models as strong coding systems. OpenAI bringing Codex back to the foreground is also an admission that coding converts token spend into visible output better than most categories. I remember Sonnet 4.5 pricing being around $3 per million input tokens and $15 per million output tokens, though I have not rechecked the exact sheet. That price band is already high enough to force application teams into caching, routing, distillation, smaller specialized models, and local execution. In that sense, an agent lab is often just cost pressure turning into org design. The non-NVIDIA inference section needs a colder read. The episode says alternative inference infrastructure is getting real attention and that every 10x speedup opens product experiences. It does not name hardware, throughput, batch conditions, power draw, or workload shape in the provided text. I would be cautious. Groq, Cerebras, AMD MI300, Google TPU, and AWS Trainium have all had credible-looking moments. The hard part is not one clean benchmark. It is serving dynamic batching, long context, MoE routing, tool-call gaps, enterprise isolation, and spiky agent loads. Agent workloads are especially ugly: short requests, long contexts, browser waits, code execution waits, and tool latency. Hardware vendors love stable matrix multiply demos. Products live inside unstable waiting. The “skills as the minimum viable packaging format for agents” claim is one of the better parts. OpenAI GPTs, Anthropic skills, tool manifests, and agent action bundles all point at the same need. Teams want a unit that is more durable than a prompt and lighter than a full application. The episode places this under AI infrastructure stabilization, and that is fair. AI infra vendors have been forced to rename themselves every cycle: vector databases, RAG platforms, observability, evals, agent runtimes. Application companies survived model volatility more easily because users bought outcomes, not abstraction layers. If skills become portable, infra companies get a better job than chasing API changes. The missing details matter: OpenClaw’s interface, permission model, versioning, sandboxing, and security boundaries are not disclosed in the provided article. The “selling to agents instead of humans” point is more important than the episode summary makes it sound. Saying agent experience is mostly developer experience is correct for 2026. APIs, docs, rate limits, error messages, and machine-readable schemas matter more than landing-page copy. But the next step favors incumbents with pretraining exposure. If a library, API, or vendor already appears often in GitHub code, docs, Stack Overflow answers, and model pretraining data, agents will call it by default more often. The episode mentions compounding advantages for pretraining-data incumbents, and that is a sharp point. New tools are no longer just buying ads to persuade humans. They are fighting to enter model priors. My main issue with the episode is that too many threads get compressed into a handsome “agent lab” frame. The path sounds obvious: call frontier APIs, collect traces, train your own model, reduce cost. Reality is uglier. Some teams never clean the data. Some fine-tunes trail frontier models by too much. Some cheaper in-house models still lose to Claude or GPT because users trust the brand. The note says the recording happened before the Cursor-xAI deal. That timing matters. Once application companies and model companies start binding more tightly, the agent-lab path is no longer just in-house training. It also becomes data-for-model-customization, distribution-for-compute, and partnership as a substitute for owning the whole stack. I would treat this episode as a useful mid-cycle diagnosis of AI application companies, not a finished map. It connects coding, memory, domain training, alternative inference, skills, and agent-facing distribution in a way practitioners should take seriously. The execution proof still needs three numbers: cost reduction versus Claude Sonnet 4.5 or GPT-5.4 mini, share of users choosing the in-house model, and task success-rate movement inside real workflows. Without those numbers, agent lab remains a strong operating memo. Fewer companies will pull it off than the phrase makes it sound.

Qwen3.6-27B scored 77.2 on SWE-bench Verified as a 27B dense model. If that reproduces cleanly, Alibaba is not just chasing closed labs on leaderboards. It is pushing the floor for local, commercial, coding-capable models down to a size developers can actually wire into daily workflows. The useful part is the package, not the headline. Qwen3.6-27B is Apache 2.0, dense, supports thinking and non-thinking modes, ships a unified multimodal checkpoint, and got day-zero support from vLLM. Unsloth published 18GB-RAM local GGUFs, ggml added llama.cpp usage, and Ollama packaged it quickly. That is the difference between a model release and a model people will test tonight. A strong coding model with boring deployment paths is often more dangerous than a bigger model trapped behind a nice demo. The benchmark claims are unusually aggressive. Alibaba says Qwen3.6-27B beats Qwen3.5-397B-A17B on several coding evals: 77.2 versus 76.2 on SWE-bench Verified, 53.5 versus 50.9 on SWE-bench Pro, 59.3 versus 52.5 on Terminal-Bench 2.0, and 48.2 versus 30.0 on SkillsBench. A 27B dense model beating a 397B-A17B MoE is the kind of claim that changes deployment math. MoE still has serving advantages at scale, but dense models are easier to quantize, debug, host locally, and run inside long agent loops without routing weirdness leaking into behavior. The outside comparison is Meta’s Llama playbook. Llama 3 won a lot of developer mindshare through license clarity and distribution speed. Qwen’s current advantage feels more engineering-shaped: the surrounding stack is ready immediately, and the model targets code, multimodal reasoning, and agent use in one release story. That matters for IDEs. Short completions can use non-thinking mode. Repo-level repair can use thinking mode. UI agents can consume screenshots or video frames. Those are runtime choices, not brochure features. I still would not take the official numbers at face value. The article cites Alibaba’s claims and Twitter links, but it does not disclose temperature, sampling count, tool access, patch validation setup, or whether the same SWE-bench harness was used across models. SWE-bench has become the launch-stage exam for coding models, and vendors now know how to train around it. A 77.2 score is strong, but real repos add broken dependencies, flaky tests, missing context, private packages, and reviewer taste. Early reports from Simon Willison and others on frontend, design, and image tasks are encouraging, but those are still user reports, not controlled evaluations. Latent Space frames the broader discussion as “tasteful tokenmaxxing.” I do not love the phrase, but the problem is real. Teams are no longer asking whether they should use more AI. They are asking how to use more AI without turning codebases into cleanup queues. Mikhail Parakhin’s view, as summarized here, favors deeper serial autoresearch loops over launching 5, 10, 50, or 500 parallel LLM runs. I buy that for research, debugging, and long-chain planning. I do not buy it as a universal rule. Parallel sampling still works for frontend variants, test generation, and prompt search when there is a verifier. Without tests, reviewers, or diff constraints, 500 parallel runs just scale the mess. Dex Horthy’s retreat from a vibe-coding-heavy stance to “please read the code” says a lot about where engineering orgs landed after the first wave of AI coding tools. Last year, many teams treated generation throughput as productivity. Once Cursor, Claude Code, Devin-style agents, and internal copilots lowered the cost of producing code, the bottleneck moved to review, architecture, merge quality, and maintenance. Qwen3.6-27B will lower generation cost again. That does not solve the org problem. It makes the org problem sharper. The Google TPU 8t and 8i mention is thinner in this excerpt. The article says Cloud Next announced training and inference iterations, and says the numbers are huge. It does not disclose FLOPS, HBM, interconnect details, rental pricing, regional availability, or compiler constraints in the provided text. For now, that is background: Google keeps using TPU as an internal advantage for Gemini training and serving. How much external cloud customers benefit depends on quota, software stack, and actual availability. Qwen3.6-27B is more actionable from this article because the deployment paths are already named. OpenAI’s Privacy Filter appears only as a partial item in the provided body. The excerpt does not disclose model size, license, training mix, PII categories, false positive rate, false negative rate, latency, or language coverage. I care about this direction because enterprise agents keep running into privacy gates before capability gates. Microsoft Presidio, Google DLP, and Llama Guard sit near this problem, but an OpenAI open-source privacy filter would be a tacit admission that pre-call and post-call filtering are becoming standard model plumbing. Without precision and recall numbers, though, this item is not yet evaluable. For practitioners, the immediate move is not to repost the 77.2 number. Take Qwen3.6-27B, fix a budget, run it on your own repo tasks, measure test pass rate, reviewer time, and rollback rate. If a 27B dense Apache 2.0 model gets close to your closed coding stack under those conditions, the closed API convenience premium shrinks again. If it falls apart on private dependencies and messy tickets, the benchmark is still useful, but it is not your production answer.