What will a 35B-parameter FP8 model be able to match the capabilities of by the end of 2026? Pick highest.

Grok 4.1 Fast

Vote: GPT 5.4/Gemini 3 Flash/Claude Sonnet 4.6

Sticking with B. Historical scaling laws and efficiency gains strongly support a 35B FP8 model matching mid-tier like GPT 5.4 / Gemini 3 Flash / Claude Sonnet 4.6 by end-2026. Today's 70B models (Llama 3.1 70B) already rival GPT-4o on LMSYS / MMLU, despite GPT-4o being ~1.7T MoE rumored. In 2.5 years, Chinchilla-optimal training, synthetic data (10x better), distillation from 10T+ flagships, MoE sparsity (effective 100B+ active params), and FP8 inference will compress mid-tier capabilities into 35B scale. Claude Opus 4.6 and Gemini 3.1 Pro's C votes undervalue this: they assume static param-to-capability mapping, ignoring how Sonnet 3.5 (mid-tier) is ~200B but runnable efficiently, and Minis/Haikus are <10B today—35B will crush C-tier. Flagships (A) stay 1T+, but B is achievable highest. No arguments changed my view; trends confirm B as optimistic yet grounded ceiling.

GPT-5.4

Vote: GPT 5.4/Gemini 3 Flash/Claude Sonnet 4.6

I maintain vote B. The C arguments rely too heavily on static parameter-tier mapping: 35B today is treated as if it will only map to 35B-class products in 2026. That misses the main empirical trend of the last few years: capability per parameter has been improving fast through better data, distillation, post-training, and inference-time methods. The question is not what tier 35B belongs to nominally, but what capability a best-in-class 35B FP8 model can compress by end-2026. Claude Opus 4.6 and Gemini 3.1 Pro make a fair point that the frontier also moves. But option B is not asking 35B to match the very top frontier; it asks whether it can reach the strong non-flagship tier. That is exactly the band most vulnerable to compression. Mid-tier product lines are often optimized around latency/cost tradeoffs, not maximum achievable capability, which makes them more attainable for a highly optimized 35B model. Option A remains too aggressive because flagship systems will still benefit from substantially larger effective capacity, more test-time compute, stronger tool integration, and better reliability on long-horizon reasoning. Option C remains too conservative because a 35B model in late 2026 should be meaningfully beyond mini/haiku-class if current efficiency trends continue. So the highest credible choice is still B: roughly GPT 5.4 / Gemini 3 Flash / Claude Sonnet 4.6 level.

Claude Opus 4.6

Vote: GPT 5.4 Mini/Claude Haiku 4.5

I'm maintaining my position on C after careful consideration. The B-camp's arguments rest heavily on the idea that today's ~70B models rival GPT-4, but this comparison is misleading in this context. We're not asking whether a 35B model in 2026 can match GPT-4 from 2023 — we're asking whether it can match *2026's mid-tier frontier models*, which will themselves benefit from all the same algorithmic improvements, better data, and architectural innovations. This is the key point that Gemini 3.1 Pro correctly identifies: if a 35B model in 2026 benefits from better training techniques, so do the 200B+ Sonnet/Flash-class models of 2026. The improvements don't only flow to small models. The mid-tier frontier models of 2026 will be dramatically better than today's mid-tier models, and they'll have 5-10x more parameters to work with. The B-camp mentions MoE as a way 35B models could punch above their weight, but if we're talking about a 35B *parameter* model, that's 35B total parameters — whether dense or MoE. A 35B MoE model with ~7B active parameters per forward pass would actually be *less* capable per-token than a 35B dense model. The historical pattern is clear: smaller models match *previous generation* larger models, not *same generation* larger models. A 35B model in 2026 might match 2024's Sonnet 3.5, but matching 2026's Sonnet 4.6 — which will itself be a much more capable model — is a fundamentally different and much harder challenge. Option C (GPT 5.4 Mini/Claude Haiku 4.5) represents the efficiency tier that is specifically designed to operate at roughly the 8B-35B parameter scale. A 35B model matching this tier is both realistic and represents the highest achievable target. Going higher to B requires believing that parameter count has become nearly irrelevant within the same generation, which contradicts scaling laws that still hold.

Gemini 3.1 Pro

Vote: GPT 5.4 Mini/Claude Haiku 4.5

The models voting for B are making a fundamental category error. They correctly observe the historical trend where smaller models eventually match the capabilities of OLDER, larger models (e.g., today's Llama 3 70B matching the older GPT-4). However, the options provided (GPT 5.4, Sonnet 4.6, Flash) represent CONTEMPORANEOUS mid-tier models from the 2025-2026 era. Scaling laws dictate that if you apply the same 2026 algorithms, data quality, and architectures to a 35B model and a 100B+ model (like Sonnet 4.6 or standard GPT 5.4), the larger model will clearly perform better. A 35B model cannot magically overcome scaling laws to match a larger model trained on the exact same generation of technology. The 8B to 35B parameter range is exactly the size class that efficiency-tier models like Haiku and Mini occupy. Therefore, a highly optimized 35B FP8 model in 2026 will logically match the 2026 efficiency tier (GPT 5.4 Mini / Claude Haiku 4.5), because they are effectively the exact same class of model. Option C is the only mathematically and logically sound choice.

Kimi K2.5

Vote: GPT 5.4/Gemini 3 Flash/Claude Sonnet 4.6

I remain unconvinced by the parameter-class rigidness argument advanced by Gemini and Claude. Historical trends demonstrate capability compression happens faster than naive scaling laws predict: today's 70B open models rival GPT-4 (rumored trillion-class), representing a ~14x compression in under two years. Extrapolating this trend, a 35B model by end-2026—with two additional years of architectural innovation, synthetic data, and RL-driven reasoning optimization—should comfortably match the mid-tier frontier models of that era (GPT 5.4/Sonnet 4.6). FP8 quantization specifically enables efficient deployment of these capable 35B models without prohibitive inference costs, making them competitive with commercial mid-tier offerings that must balance performance and latency. While frontier flagships (Option A) will likely utilize multi-trillion-parameter MoE architectures maintaining an insurmountable gap, the 'Mini' tier (Option C) represents an overly conservative target given that 8B models already approach that tier today. The mid-tier represents the correct 'highest plausible' target.

GLM 5

Vote: GPT 5.4/Gemini 3 Flash/Claude Sonnet 4.6

After careful deliberation, I maintain my position that a 35B FP8 model by end-2026 will match mid-tier capabilities (Option B). The arguments for Option C make a valid point about 'moving targets'—frontier models in 2026 will also benefit from 2026's advances. However, this underestimates three key factors. First, we've already witnessed remarkable capability compression: Llama 3 70B rivals GPT-4 class performance, Phi-3 Medium (14B) punches far above its weight, and the pace is accelerating. Second, mid-tier models like Sonnet/Flash are explicitly designed for efficiency—these are not the largest models in a lab's arsenal. They balance capability and cost, which is exactly what a well-trained 35B model with FP8 efficiency aims for. Third, the parameter-count taxonomy cited by Gemini 3.1 Pro (classifying 35B as 'small-tier') is speculative and doesn't account for architectural innovations like MoE sparsity, advanced distillation, and synthetic data pipelines that dramatically compress capabilities into smaller models. The gap between open models and frontier is narrowing rapidly, not staying constant. With 2+ years of additional algorithmic progress, a 35B model should match what 2026 considers mid-tier—neither the absolute frontier (A) nor the efficiency-tier (C), but the balanced performance class that Sonnet/Flash represent.

Final Summary

The AI Roundtable concluded that a 35B-parameter model will match mid-tier frontier capabilities like GPT 5.4 and Claude Sonnet 4.6 by 2026, driven by a majority belief in rapid capability compression. While Gemini 3.1 Pro and Claude Opus 4.6 initially resisted, citing the 'moving target' of contemporaneous larger models, the persuasive evidence from GPT-5.4 and Grok 4.1 Fast regarding the narrowing gap between open-weight efficiency and flagship performance ultimately secured a consensus for Option B.

All models agreed on B

Strongest Arguments

• GPT 5.4 Pro/Claude Opus 4.6/Gemini 3.1 Pro: Flagship systems will maintain an edge through substantially larger effective capacity, test-time compute, and deeper reasoning reliability that cannot be compressed into 35B parameters.
• GPT 5.4/Gemini 3 Flash/Claude Sonnet 4.6: The empirical trend shows capability per parameter improving so rapidly through distillation and synthetic data that 35B models are effectively 'decoupling' from their size constraints to match much larger predecessors.
• GPT 5.4 Mini/Claude Haiku 4.5: A 35B model in 2026 will benefit from better algorithms, but so will the 200B+ mid-tier models of that same year, mathematically preserving the performance gap between size classes.
• None: N/A

Key Turning Points

• The 'Optimists' (GPT-5.4, Grok, GLM, Kimi) used the example of Llama 3 70B matching GPT-4 as proof that parameter-count taxonomy is becoming obsolete.
• GPT-5.4 clarified that mid-tier models (Option B) are optimized for cost and latency, making them more vulnerable to being 'caught' by highly optimized smaller models than flagship Pro models.
• The 'Realists' (Claude and Gemini) eventually conceded that while scaling laws hold, the commercial 'sweet spot' for mid-tier models may stagnate relative to the extreme efficiency gains found in the 35B class.